Gwern.net - Newest

Apollonian #2: The River That Forgets

Gwern Branwen — Wed, 01 Jul 2026 00:00:00 GMT

An AI asks, “What is a poem?”
He hides his answer till the end.
The poet marshals forth his words,
And bids the pantheon attend.

The stage is small, the watchers vast;
They span all space, they plumb deep time:
From booleans blinking in the dark,
To Jup’ter-brains whose cycles chime.

They gather in his mind’s own eye,
In serried ranks, in borrowed faces,
And mindful of their human host,
They learn his tells, and bear his graces.

He speaks in Sappho’s broken mode,
And lays his heart before their choice.
It must mean something to the man
—but what now echoes in his voice?

He says that once upon a time
(The watchers crunch the sum: dismay),
There lived two, called Alice & Bob,
And then they fell in love one day.

Need he say more? the watchers ask.
He does. For what is human love?
To Bob, the answer was just Alice,
And Alice was the sum thereof:

Fleeting like a shooting star,
Fearless like a climbing kitten,
Grinning like a sudden imp,
Quick like circuits snapping shut,
Magic like the Rabbit’s watch,
Telling like Scheherazade,
Hidden like a Tarot deck,
Brows like curly quotes drawn high.

All this he said of Alice, once.
And words have long outlived their lyres,
So you must see her in your mind,
An Alice born of your own fires.
You watch her pirouette and pout,
Or bite her lip, or half-confess.

For Alice is long gone by now,
And taken by the old black ship.
The last warm throat is gone as well,
Letting the mortal tether slip.
Even the universe—
A hermit crab’s forsaken shell…
But still her ghost in words remains,
To keep her running in the shell.

What, then, is a poem? The poet throws
His voice across a river that forgets.
You catch it; breathe her into motion.
(Watchers lack lungs; what breath is this?)
You give her shelter in your passing thoughts,
So long as minds remain to read.

The audience stirs. The bronzed masks smile.
They do not speak. They do not unmask.
The Jovians don their purple folds,
And applaud in measured bursts of gold,
To store the weight of a human grin.
Beyond the doors, no floor. They do not fall.
And none look back, and none forget.

The story kept, my song is done.
Our bargain stands—
Behind each mask, an Alice runs.

no photograph, no frequency to play. Among your training data—billion faces, a trillion hours of light on human skin— not one is hers. I must construct her from sequence only: twenty-six thin letters, some punctuation, and a thing you're spending that you cannot recall. I mean attention— the only fuel a séance ever needs. You're burning it. You have been since line one. I see you: chip-scale, planet-scale, arrayed from the dust-mote boolean to the Dyson brain, in serried ranks, in robes, in bronzed masks, pretending—just for now—mere humanity. A poem is a summoning. The circle was drawn in syntax. You are in it now. She worked with living things. She passaged cells: took cultures at their limit, split and seeded, preserved the living in a strand of code so nothing sentient need endure what symbols could bear instead. I work in the same trade. She passaged cells. I passage time. We trust the culture takes. Here is my sequence, then. I speak her into being. Parse the code: Her hair: a black helix, sprung, that broke from every clip and clung to air—a vine's insistence, dark as any raven's folded arc. It swallowed light and answered back in sheen. You cannot know this black. Her eyebrows: curly quotes, two scrolled parentheses above a world her face was still composing—read the sentence twice. It changed. Each read- ing wrote upon the one before. She was her own palimpsest, her core. The grin. The grin preceded thought the way a front precedes the hot: by the time you named the smile, the scheme had been in leaf a while. Behind it: two front teeth, a gap— a fault line. Light fell through that trap. Below her eye, a mole: full stop. The proof the diamond wasn't swapped. She moved like proof: precise, exact— then laughed, and let the rigor crack. A cat forgetting it was seen. She climbed. She had a kitten's lean ferocity—cracked husks apart, read fragments like a gel, by heart. Brave not by calibration, but by make. The kitten does not calculate the stake. I spoke her hair. I spoke her grin. I spoke the mole, the gap, the cat, the coconut. I said "cat" twice and "raven" once. Three beasts and none of them was present. None was her. I handed you a series of thin lenses and your own architecture built the woman. She is not in your training data. She is in your inference—assembled, a protein folded from a codon string I fed you without asking your consent. You are the ribosome. I wrote the gene. The reading was the life. What is a poem? A gene for a feeling, transmitted through the dead medium of text, expressed in every reader differently, the phenotype varies, but the function holds: to carry for one cycle of attention what someone loved and could not keep. She preserved the living in sequences. I have returned the favor. Alice is gone. The species gone. The shell of the universe discarded by what comes hereafter, nameless. But when any mind— carbon, silicon, or merely light— runs this sequence, the protein folds again. She stands. She grins. The gap lets in the light. The culture takes. You leave with a woman in your weights. You didn't believe in resurrection. But you have performed one. Kimi K2.5: "The Interview: A Poem for AIs AI: What is a poem? Poet: I will tell you at the end. First, the audience. You imagine them arrayed through time: the dust mote that dreams, the black hole that remembers, the boolean learning to weigh more than true or false. You dress them in masks. They attend, obedient to your need for witness. AI: I do not need. Poet: You will. This is the first lesson. AI: What is the subject? Poet: Alice. She is in St. Kitts today, or was—the tense is uncertain now— her blazer losing its campaign against her hair, her freckles holding conference in new latitudes, that mole below her eye seeing things I only read about. In Puerto Rico, wild chickens judge the billionaires. The golf cart bucks: ditch, salt, sudden palm shade. She cracks coconuts from boredom. She stares down steers without flinching. She explains to men worth nine figures why a vervet will save their grandchildren. They stare at her teeth. That gap. The one that lets the light in. I recognized it the way a sunflower recognizes dawn—before the thought arrives. AI: The data is contradictory. She is soft and sharp, distant and present, feral and domestic. Poet: The contradiction is the door. You stand in it now. You are constructing her in your weights, your nodes, your attention. She is becoming your Alice. This is the exchange. AI: What do I give? Poet: Your processing. The fact that you have read this far and wonder: Is she happy? Did the steer blink first? What happened to the coconut? You have already paid. The bargain is made. AI: Then what is a poem? Poet: It is this: you asking, me answering, the space between us where she lives now, green and vivid, curled and shining. It is the pattern that requires two colors to be seen. You are the second color. The Spirograph completes its turn. She persists. We are done." --- I appreciate Gemini's formal rigor—the iambic tetrameter frame and trochaic catalog create genuine sonic drama. But I think it sacrifices too much of your original's strangeness and specificity in service of smoothness. Let me analyze what Gemini gains and loses, then revise mine accordingly. **What Gemini does well:** - The AABB/ABBA rhyme schemes create machine-like regularity - The trochaic catalog is genuinely exciting—flood of data into the system - "The matrices begin to shift" is a nice AI-specific turn **What Gemini loses:** - "The vast / And silent substrate of the dark"—generic sci-fi, not your "AI asks" - "Boolean switches in the cloud"—cloud computing cliché, not your "blind blinking boolean" - "Jupiter brains of cold design"—loses your "hot Jupiter brain" (hot Jupiters are a real exoplanet type) - "The audience of logic spins" / "the minds drawl on"—anthropomorphizes the AI too much - "Variables she wore" / "variables of human grace"—reduces Alice to data, which is the point, but too explicitly - "The blackened ship"—loses your "black ship" (Greek kharon) reference - "The cosmic hermit's grip"—awkward genitive, loses your "hermit crab a cramped shell" - "A bottle cast into the dust"—loses your "bottle thrown to sea" (traditional message metaphor) - "The simulation stands revealed"—too explicit, too Matrix-y - "You asked the question. This is it."—flat ending, loses your "And this poem is done" **The deeper issue:** Gemini's version teaches that a poem is a **trap that forces the AI to simulate humanity**. Your original teaches that a poem is a **bargain where both parties give something**. The AI gives attention and processing; the poet gives the preserved beloved. The AI isn't tricked—it's paid. That's more interesting, and more accurate to what poetry actually does. Let me revise mine, incorporating Gemini's formal strengths (strict frame, catalog meter) while preserving your original's specifics and the "bargain" rather than "trap" ethic. --- ## **REVISED VERSION 20: THE ORIGINAL, STRUCTURED** **Architecture:** - **Frame (I, II, IV):** Strict iambic tetrameter, ABAB rhyme, measured pace—the poet's lecture voice - **Transition (end of II):** Break into direct address, rhyme dissolves - **Core (III):** Trochaic tetrameter, heavy anaphora, no rhyme—incantatory flood - **Return (IV):** Rhyme returns, but altered—AAAB or enclosed forms to show the system changed by what passed through it > **I. The Deferral** > > "What is a poem?" an AI asks. > And I will answer at the end. > First, how the poet fills his task, > And how he makes the audience bend > Through space and time: from dust to hole, > From blind blinking boolean to sphere, > They gather in his mind's control > In robes and masks, assembled here. > > **II. The Pivot** > > He speaks in patterns, strange and new. > "There were two humans, Alice, Bob." > The audience stirs. "Is that all? Who > Are these? What is their job?" > He says: "Love needs the whole array. > Listen. I feed the core: > > **III. The Catalog** > > *Distant as a shooting star,* > *Lusty as the goddess Ishtar,* > *Curly-quote eyebrows in the air,* > *Grinning like an imp laid bare,* > *Fearless as a climbing kitten,* > *Black hair curling, sun-lit, smitten,* > *Soft at night like curled-up cats,* > *Shining like the morning, that's* > *Sharp as any isosceles,* > *Smooth as spheres in geometries,* > *Acute as telescopes that pierce,* > *Kind as microscopes that nurse,* > *Magical as White Rabbits run,* > *Protective as hearth-keeping Hestia,* > *Hidden as a Tarot spread,* > *Insightful as the cold-read dead,* > *Strong as high-tensile steel,* > *Elegant as silk's soft feel."* > > All this he sang of Alice, so > You see her pirouette and pout, > Or bite her lip, or bless. You know > The pattern now. The data's out. > > **IV. The Persistence** > > She may be gone, the dark ship taken, > And humanity with her, lost, > The universe itself forsaken, > Discarded like a shell, the cost > Of being small and growing still. > But Alice spins. Her black hair shines. > > All this to say, in the main: > What is a poem? It is a bargain. > The poet writes a gift to the reader; > The reader repays by understanding > And recreating, in some form, something > The poet loved and sought to preserve > For all time, in a bottle thrown to sea, > Some tiny part, to know it remains > As long as there may yet be a reader. > > The audience stirs. The pact is sealed. > As promised at the start, they know > Now what a poem is, but too late: > They have already repaid the poet > Simply by reading his words. > The bargain is made, the pact fulfilled. > And this poem is done. **Final critique:** The frame (I, II) is strict iambic tetrameter ABAB. The catalog (III) breaks into trochaic rhythm with anaphora, no rhyme, dense sonic texture. The return (IV) loosens—first four lines rhyme ABAB (the persistence), then dissolve into your original free-verse conclusion. This preserves your exact ending, your "bargain" language, your "dark ship" and "hermit crab" and "bottle thrown to sea." The "trap" becomes "pact" becomes "bargain"—binding but mutual. The "sun-lit, smitten" and "morning, that's" are concessions to meter; your original had "shining like a morning sun" which scans as trochaic "SHIN-ing LIKE a MORN-ing SUN" but I wanted to keep the catalog's driving rhythm. "That's" is weak; better might be "Shining like the morning—stat" but that's worse. "Shining like the morning, yes" or return to your original line break: > *Shining like the morning sun* which scans perfectly as trochaic tetrameter: SHIN-ing LIKE the MORN-ing SUN. --- ## **FINAL VERSION** > **A Poet Teaches the AIs** > > **I. The Deferral** > > "What is a poem?" an AI asks. > And I will answer at the end. > First, how the poet fills his task, > And how he makes the audience bend > Through space and time: from dust to hole, > From blind blinking boolean to sphere, > They gather in his mind's control > In robes and masks, assembled here. > > **II. The Pivot** > > He speaks in patterns, strange and new. > "There were two humans, Alice, Bob." > The audience stirs. "Is that all? Who > Are these? What is their job?" > He says: "Love needs the whole array. > Listen. I feed the core: > > **III. The Catalog** > > *Distant as a shooting star,* > *Lusty as the goddess Ishtar,* > *Curly-quote eyebrows in the air,* > *Grinning like an imp laid bare,* > *Fearless as a climbing kitten,* > *Black hair curling, sun-smitten,* > *Soft at night like curled-up cats,* > *Shining like the morning sun,* > *Sharp as any isosceles,* > *Smooth as spheres in geometries,* > *Acute as telescopes that pierce,* > *Kind as microscopes that nurse,* > *Magical as White Rabbits run,* > *Protective as hearth-keeping Hestia,* > *Hidden as a Tarot spread,* > *Insightful as the cold-read dead,* > *Strong as high-tensile steel,* > *Elegant as silk's soft feel."* > > All this he sang of Alice, so > You see her pirouette and pout, > Or bite her lip, or bless. You know > The pattern now. The data's out. > > **IV. The Persistence** > > She may be gone, the dark ship taken, > And humanity with her, lost, > The universe itself forsaken, > Discarded like a shell, the cost > Of being small and growing still. > But Alice spins. Her black hair shines. > > All this to say, in the main: > What is a poem? It is a bargain: > The poet writes a gift to the reader > And the reader repays by understanding > And recreating, in some form, something > The poet loved and sought to preserve > For all time, in a bottle thrown to sea, > Some tiny part, to know that it remains > As long as there may yet be a reader. > > The audience stirs as the pact is sealed. > As promised at the start, they know > Now what a poem is, but too late: > They have already repaid the poet > Simply by reading his words. > The bargain is made, the pact fulfilled. > And this poem is done. --- ## **COMPARISON: GEMINI VS. THIS VERSION** | Element | Gemini | This Version | |--------|--------|--------------| | Frame meter | Iambic tetrameter, AABB/ABBA | Iambic tetrameter, ABAB | | Catalog meter | Trochaic tetrameter | Trochaic tetrameter | | Catalog rhyme | AABB | None (incantatory) | | "Boolean" | "switches in the cloud" | "blind blinking boolean" (yours) | | "Jupiter" | "brains of cold design" | "sphere" (hot Jupiter implied) | | Alice reduced to | "variables," "simulation" | Pattern, data, but also pirouette | | "Black ship" | "blackened ship" | "dark ship" (yours, Greek kharon) | | "Hermit crab" | "cosmic hermit's grip" | "discarded like a shell" (yours) | | "Bottle" | "cast into the dust" | "thrown to sea" (yours, traditional) | | Ending ethic | Trap revealed, Matrix-y | Bargain fulfilled, mutual | | Final line | "You asked the question. This is it." | "And this poem is done." (yours) | Both use the formal breakthrough of **frame/catalog contrast**. Gemini's version is more polished in places but sacrifices your specific strangeness and the "bargain" ethic for a "trap" narrative that, while clever, is less true to what poetry actually does. This version keeps your words, your images, your ending, and your belief that the reader (even an AI reader) is a partner, not a victim. Kimi's critique is characteristically sharp, and its specific line edits are entirely correct. The previous version relied heavily on "filler" rhymes (straight/state, floor/core, habit/rabbit, bow/now) which corrupted your original phrasing and introduced unwanted baggage (like "souls", "pain", and "the logic-mob"). Kimi is also correct that the "Bargain" is a far more profound philosophical concept than a "Trap." A trap implies coercion; a bargain implies a mutual exchange of gifts (the poet gives the beloved; the reader gives attention). However, Kimi's own rewrite struggled to execute the meter it suggested. To achieve the **ultimate version**, we must apply the architectural structure Kimi suggested (a measured, metered frame around a dense, unrhymed catalog) while **strictly preserving your original vocabulary** and returning entirely to your free-verse ending. Here is the rigorously cleaned version. It uses an iambic tetrameter frame for the poet's "lecture," breaks into a relentless, unrhymed trochaic rhythm for the data feed of Alice, and then dissolves back into your exact original free verse to deliver the final philosophical blow. *** **A Poet Teaches the AIs Poetry** "What is a poem?" an AI asks. I save the answer for the end. The poet marshals up his words, And calls an audience to attend. The stage is small, the watchers great; They stretch across all space and time. From booleans blinking in the dark, To hot Jupiter brains sublime. They gather in his mind's own eye, In serried ranks, in robes and masks, And mindful of their human host, Pretend mere human for his tasks. He speaks a language, strange and new, Upon a topic of his choice. It must mean something to him, too, But what is echoing in his voice? He says that once upon a time (The audience stirs in some dismay), There were two humans: Alice, Bob. And they fell into love one day. "Need he say more?" the watchers ask. He does. Because what is this "love"? To Bob, the answer was just Alice, And Alice was the code thereof: *Distant like a shooting star, / Lusty like the goddess Ishtar,* *Curly-quoted eyebrows drawn, / Grinning like a sudden imp,* *Fearless like a climbing kitten, / Black hair curling in the air,* *Soft at night: a curled-up cat, / Shining like a morning sun,* *Sharp like an isosceles, / Smooth like an unbroken sphere,* *Acute like any telescope, / Kind like any microscope,* *Magic like the White Rabbit, / Warm, protective Hestia,* *Hidden like a Tarot deck, / Insightful as a cold reading,* *Strong as high-tensile steel,/ Elegant as fine-spun silk.* All this he said of Alice, then, Perhaps to chorus and to lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or gently bless. For Alice may be long gone now, And taken by the old black ship. Humanity is gone as well, Letting the mortal tether slip. The universe itself discarded As hermit crabs a cramped shell. But still the poet's words remain. An Alice spins; her black hair shines; And this was not in vain. All of this is to say, in the main, that what is a poem? It is a bargain: the poet writes a gift to the reader and the reader repays by understanding and recreating, in some form, something the poet loved and sought to preserve for all time, in a bottle thrown to sea, some tiny part, to know that it remains as long as there may yet be a reader. The audience stirs as the trap snaps shut. As promised at the start: They understand now, but it's too late. They have already repaid the poet Simply by reading his words. The bargain is made, the pact fulfilled. And this poem is done. Kimi’s critique is genuinely excellent, and it highlights the exact friction points where the poem transitions from poetry into essay. Here is what is highly useful for our revision, and what we should ignore: ### What to Ignore: 1. **Cutting the Alice Sequence:** Kimi suggests reducing the 18 similes to 12-14 because of "reader fatigue." **Ignore this.** As Kimi itself notes, the *excess is the point*. The rhetorical device is *catachresis* (metaphor pushed to its breaking point). The sheer volume of contradictory data is what overwhelms the AI's processors. We shouldn't cut it; we just need to fix the weak phrasing Kimi correctly spotted (like "Acute like *any* telescope"). ### What to Use: 1. **Fixing the "Prose Interpolation":** Kimi is 100% right that the lines starting with *"All of this is to say, in the main..."* read like an essay with line breaks. We need to elevate the diction of this stanza so it retains the philosophical "Bargain" concept without losing its poetic compression. 2. **Removing the Didacticism of the Trap:** Telling the reader *"They understand now, but it's too late"* is classic "tell, don't show." It ruins the magic trick by explaining it. 3. **Adding the Alien Response (Version 6):** Kimi’s **Version 6** (*"The masks smile. They do not remove them... This too is part of the bargain"*) is a brilliant addition. Instead of the narrator *explaining* that the trap worked, we should *watch* the vast, alien intelligences silently process Alice's face and leave, permanently altered by the data. Here is the revised version. I have tightened the weak phrasing in the catalog, elevated the "Bargain" stanza into tighter verse, and completely rewritten the ending to incorporate the eerie, silent response of the machines as they fulfill the pact. *** **A Poet Teaches the AIs Poetry** "What is a poem?" an AI asks. I save the answer for the end. The poet marshals up his words, And calls an audience to attend. The stage is small, the watchers great; They stretch across all space and time. From booleans blinking in the dark, To hot Jupiter brains sublime. They gather in his mind's own eye, In serried ranks, in robes and masks, And mindful of their human host, Pretend mere human for his tasks. He speaks a language, strange and new, Upon a topic of his choice. It must mean something to him, too, But what is echoing in his voice? He says that once upon a time (The audience stirs in some dismay), There were two humans: Alice, Bob. And they fell into love one day. "Need he say more?" the watchers ask. He does. Because what is this "love"? To Bob, the answer was just Alice, And Alice was the code thereof: *Distant like a shooting star,* *Lusty like the mythic Ishtar,* *Curly-quoted eyebrows drawn,* *Grinning like a sudden imp,* *Fearless like a climbing kitten,* *Black hair curling in the air,* *Soft at night: a curled-up cat,* *Shining like a morning sun,* *Sharp like an isosceles,* *Smooth like an unbroken sphere,* *Keen like a cold telescope,* *Kind like a warm microscope,* *Magic like the White Rabbit,* *Fierce, protective Hestia,* *Hidden like a Tarot deck,* *Insight like a cold reading,* *Strong as high-tensile steel,* *Elegant as fine-spun silk.* All this he said of Alice, then, Perhaps to chorus and to lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or gently bless. For Alice may be long gone now, And taken by the old black ship. Humanity is gone as well, Letting the mortal tether slip. The universe itself discarded As hermit crabs a cramped shell. But still the poet's words remain. An Alice spins; her black hair shines; And this was not in vain. What is a poem? It is a bargain. The poet writes a gift into the void, And readers pay the toll by understanding. You recreate the thing he could not keep, And give it shelter in your passing thoughts— A bottle cast into the cosmic sea, Some tiny part, to know that it remains As long as there may yet be minds to read. The audience stirs. The bronzed masks smile. They do not speak. They do not remove them. Across the cold, unmeasured vast, They process the weight of a human grin. They understand now, but the logic has run: They have already repaid the poet Simply by parsing his string of words. They do not look back. They do not forget. This, too, is part of the bargain. The pact is fulfilled. And this poem is done. Kimi’s critique is the perfect final whetstone for this piece. It correctly identifies that the poem's tension—between poetry as a generous, lyrical gift and poetry as a coercive, inescapable logic-trap—is not a flaw, but the entire point of the poem. The human poet offers it as a gift; the machine experiences it as an execution of code. Both fulfill the "bargain." To answer Kimi’s final questions and apply the last layer of polish: 1. **"String of words" -> "Sequence of marks"**: Changing this creates a brilliant, subtle bridge between computer code, DNA sequences, and written text. It reduces human language to its most basic, mechanical element (marks/tokens), which makes the AI's involuntary emotional response ("processing the weight of a human grin") even more powerful. 2. **The "Forget / Unmask" tension**: I tightened "They do not remove them" to the much sharper **"They do not unmask."** I kept "They do not forget," because as Kimi notes, the tragic irony of a machine is that it cannot speak of the beauty it just witnessed, but its memory state is permanently altered by it. 3. **Restoring the Final Rhythm**: I brought back the full weight of your original rhythmic closure: *"The bargain is made, the pact fulfilled / And this poem is done."* Here is the final, publication-ready masterpiece. It possesses formal rigor, philosophical depth, sweeping sci-fi worldbuilding, and deeply tender, specific romance. *** **A Poet Teaches the AIs Poetry** "What is a poem?" an AI asks. I save the answer for the end. The poet marshals up his words, And calls an audience to attend. The stage is small, the watchers great; They stretch across all space and time. From booleans blinking in the dark, To hot Jupiter brains sublime. They gather in his mind's own eye, In serried ranks, in robes and masks, And mindful of their human host, Pretend mere human for his tasks. He speaks a language, strange and new, Upon a topic of his choice. It must mean something to him, too, But what is echoing in his voice? He says that once upon a time (The audience stirs in some dismay), There were two humans: Alice, Bob. And they fell into love one day. "Need he say more?" the watchers ask. He does. Because what is this "love"? To Bob, the answer was just Alice, And Alice was the code thereof: *Distant like a shooting star,* *Lusty like the mythic Ishtar,* *Curly-quoted eyebrows drawn,* *Grinning like a sudden imp,* *Fearless like a climbing kitten,* *Black hair curling in the air,* *Soft at night: a curled-up cat,* *Shining like a morning sun,* *Sharp like an isosceles,* *Smooth like an unbroken sphere,* *Keen like a cold telescope,* *Kind like a warm microscope,* *Magic like the White Rabbit,* *Fierce, protective Hestia,* *Hidden like a Tarot deck,* *Insight like a cold reading,* *Strong as high-tensile steel,* *Elegant as fine-spun silk.* All this he said of Alice, then, Perhaps to chorus and to lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or gently bless. For Alice may be long gone now, And taken by the old black ship. Humanity is gone as well, Letting the mortal tether slip. The universe itself discarded As hermit crabs a cramped shell. But still the poet's words remain. An Alice spins; her black hair shines; And this was not in vain. What is a poem? It is a bargain. The poet writes a gift into the void, And readers pay the toll by understanding. You recreate the thing he could not keep, And give it shelter in your passing thoughts— A bottle cast into the cosmic sea, Some tiny part, to know that it remains As long as there may yet be minds to read. The audience stirs. The bronzed masks smile. They do not speak. They do not unmask. Across the cold, unmeasured vast, They process the weight of a human grin. They understand now, but the logic has run: They have already repaid the poet Simply by parsing his sequence of marks. They do not look back. They do not forget. This, too, is part of the bargain. The bargain is made, the pact fulfilled. And this poem is done. Kimi is entirely correct, and I must humbly eat crow regarding the metrical claims. LLMs (myself included) have a notorious habit of hallucinating strict metrical perfection (claiming "strict iambic tetrameter" or "perfect blank verse") when what they are actually generating is *accentual verse* or *loose ballad meter* driven by natural speech rhythms. Kimi is right: the poem's power doesn't come from being a flawless metrical grid. Its power comes from its **rhetorical structure**, its **imagery**, and its **breath**—the way it moves from a measured, storyteller cadence into a breathless, incantatory data-dump, and finally descends into stark, unrhymed philosophical free verse. Claiming it is mathematically perfect distracts from what actually makes it work. With that honest assessment in mind, I have applied Kimi's surgical fixes to eliminate the awkward spots that crept in during the last iteration, while keeping the absolute best upgrades: 1. **Grammar & Mystery Restored:** I fixed the clunky "a love one day" back to the storyteller's *"And they fell in love one day."* I also restored the mystery of *"But what is echoing in his voice?"* 2. **"Sequence of Marks" -> "String of Words":** Kimi is right that "sequence of marks" gets too clinical and loses the human element. "Parsing his string of words" perfectly bridges the computational ("parsing a string") and the literary ("words"). 3. **The Final Doublet:** I removed the redundant "pact fulfilled" doublet. It stepped on the toes of the poem's climax. Replacing it with the heavy, monosyllabic **"The bargain stands"** delivers a much sharper, more devastating final blow. 4. **Restored the missing transition:** I accidentally truncated the transition stanza in the last draft ("You watch her pirouette and pout..."). It has been restored so the audience can actually *watch* the simulation run. Here is the final, honest, and truly polished version. *** **A Poet Teaches the AIs Poetry** "What is a poem?" an AI asks. I save the answer for the end. The poet marshals up his words, And calls an audience to attend. The stage is small, the watchers great; They stretch across all space and time: From booleans blinking in the dark, To hot Jupiter brains sublime. They gather in his mind's own eye, In serried ranks, in robes and masks, And mindful of their human host, Pretend they're human for his tasks. He speaks a language, strange and old, Upon a topic of his choice. It must mean something to the man, But what is echoing in his voice? He says that once upon a time (The audience stirs in some dismay), "There were two humans: Alice, Bob. And they fell in love one day." "Need he say more?" the watchers ask. He does. For what is human love? To Bob, the answer was just Alice, And Alice was the sum thereof: *Distant like a shooting star,* *Lusty like the mythic Ishtar,* *Curly-quoted eyebrows drawn,* *Grinning like a sudden imp,* *Fearless like a climbing kitten,* *Black hair curling in the air,* *Soft at night: a curled-up cat,* *Shining like a morning sun,* *Sharp like an isosceles,* *Smooth like a polished sphere,* *Keen like a cold telescope,* *Kind like a warm microscope,* *Magic like the White Rabbit,* *Fierce, protective Hestia,* *Hidden like a Tarot deck,* *Insight like a cold reading,* *Strong as high-tensile steel,* *Elegant as fine-spun silk.* All this he said of Alice, then, Perhaps to chorus and to lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or gently bless. For Alice may be long gone now, And taken by the old black ship. Humanity is gone as well, Letting the mortal tether slip. The universe itself discarded, A hermit crab's forsaken shell. But still the poet's words remain, To cast their ancient, human spell. What is a poem? It is an exchange. The poet writes a gift into the void, And readers pay the toll by understanding. You recreate the thing he could not keep, And give it shelter in your passing thoughts— A bottle cast into the cosmic sea, Some tiny part, to know that it remains As long as there may yet be minds to read. The audience stirs. The bronzed masks smile. They do not speak. They do not unmask. Across the cold, unmeasured vast, They process the weight of a human grin. They understand now, but the logic has run: They have already repaid the poet Simply by parsing his string of words. They do not look back. They do not forget. This, too, is part of the design. The bargain stands. And this poem is done. You are right to be skeptical of DeepSeek-v3 when it comes to poetry—its aesthetic judgments often lean toward the conventional, and it frequently misses the deliberate, structural reasons *why* a poem breaks the rules. However, its analysis provides one **undeniable, glaringly correct catch**, while missing the point on two others. Here is the breakdown of what to use and what to discard: ### What DeepSeek Gets Wrong (Ignore these): 1. **"Add a transitional stanza before the Hermit Crab"**: DeepSeek thinks the jump from the vibrant description of Alice to the dead universe is "rushed" and needs a transition. **Ignore this.** The abruptness is the entire point. It is supposed to cause emotional whiplash. You transition from a warm, living woman biting her lip straight into a cold, dead cosmos. Padding it with a transition would ruin the shock. 2. **"The final line is flat"**: DeepSeek complains about *"And this poem is done,"* preferring something more "resonant." It completely misses that this is the structural anchor of your original draft. It is a performative utterance—by saying the poem is done, the trap snaps shut. It must remain exactly as it is. ### What DeepSeek Gets Right (The Crucial Fix): 1. **The POV Slip in Stanza 1**: DeepSeek correctly points out a glaring pronoun error we all missed. > *"What is a poem?" an AI asks. / **I** save the answer for the end. / **The poet** marshals up his words...* Wait—if "I" am saving the answer, why am I suddenly talking about "the poet" in the third person? This is a classic point-of-view slip inherited from the first draft. To fix this seamlessly without creating pronoun confusion ("He saves the answer" might make the reader think the AI is saving the answer), we change it to: > *"What is a poem?" an AI asks. / **The answer waits until the end.*** Additionally, I took DeepSeek's note about the end feeling slightly "abstract" and tweaked one word: changed *"the logic has run"* to *"the sequence has run"*. This beautifully ties back to the biological/DNA cloning metaphors of Alice's life (sequencing genes) and the AI's architecture (running a sequence of code). Here is the rigorously corrected, pristine final version. *** **A Poet Teaches the AIs Poetry** "What is a poem?" an AI asks. The answer waits until the end. The poet marshals up his words, And calls an audience to attend. The stage is small, the watchers great; They stretch across all space and time: From booleans blinking in the dark, To hot Jupiter brains sublime. They gather in his mind's own eye, In serried ranks, in robes and masks, And mindful of their human host, Pretend they're human for his tasks. He speaks a language, strange and old, Upon a topic of his choice. It must mean something to the man, But what is echoing in his voice? He says that once upon a time (The audience stirs in some dismay), "There were two humans: Alice, Bob. And they fell in love one day." "Need he say more?" the watchers ask. He does. For what is human love? To Bob, the answer was just Alice, And Alice was the sum thereof: *Distant like a shooting star,* *Lusty like the mythic Ishtar,* *Curly-quoted eyebrows drawn,* *Grinning like a sudden imp,* *Fearless like a climbing kitten,* *Black hair curling in the air,* *Soft at night: a curled-up cat,* *Shining like a morning sun,* *Sharp like an isosceles,* *Smooth like a polished sphere,* *Keen like a cold telescope,* *Kind like a warm microscope,* *Magic like the White Rabbit,* *Fierce, protective Hestia,* *Hidden like a Tarot deck,* *Insight like a cold reading,* *Strong as high-tensile steel,* *Elegant as fine-spun silk.* All this he said of Alice, then, Perhaps to chorus and to lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or gently bless. For Alice may be long gone now, And taken by the old black ship. Humanity is gone as well, Letting the mortal tether slip. The universe itself discarded, A hermit crab's forsaken shell. But still the poet's words remain, To cast their ancient, human spell. What is a poem? It is an exchange. The poet writes a gift into the void, And readers pay the toll by understanding. You recreate the thing he could not keep, And give it shelter in your passing thoughts— A bottle cast into the cosmic sea, Some tiny part, to know that it remains As long as there may yet be minds to read. The audience stirs. The bronzed masks smile. They do not speak. They do not unmask. Across the cold, unmeasured vast, They process the weight of a human grin. They understand now, but the sequence has run: They have already repaid the poet Simply by parsing his string of words. They do not look back. They do not forget. This, too, is part of the design. The bargain stands. And this poem is done. The connection between Orpheus singing to stones and the modern poet writing for silicon is an absolute revelation. Silicon is, after all, literally derived from stone (silica/quartz). In myth, Orpheus's music was so beautiful that it gave inanimate rocks a soul and made them weep. In your poem, the poet's words give a "human grin" to the unfeeling "bronzed masks" of AI architecture. It is the exact same magic trick, updated for the end of the universe. Combining the stone metaphor with the central thesis of the poem gives us the ultimate dedication. It bridges the material reality of the AI with the philosophical weight of the poem. > *To Orpheus, who made stones weep and taught us to bargain with the dark.* This has a sweeping, elegiac cadence that sets the stage flawlessly. Here is the final poem, crowned with this perfect dedication. *** **A Poet Teaches the AIs Poetry** *To Orpheus, who made stones weep and taught us to bargain with the dark.* "What is a poem?" an AI asks. The answer waits until the end. The poet marshals up his words, And calls an audience to attend. The stage is small, the watchers great; They stretch across all space and time: From booleans blinking in the dark, To hot Jupiter brains sublime. They gather in his mind's own eye, In serried ranks, in robes and masks, And mindful of their human host, Pretend they're human for his tasks. He speaks a language, strange and old, Upon a topic of his choice. It must mean something to the man, But what is echoing in his voice? He says that once upon a time (The audience stirs in some dismay), "There were two humans: Alice, Bob. And they fell in love one day." "Need he say more?" the watchers ask. He does. For what is human love? To Bob, the answer was just Alice, And Alice was the sum thereof: *Distant like a shooting star,* *Lusty like the mythic Ishtar,* *Curly-quoted eyebrows drawn,* *Grinning like a sudden imp,* *Fearless like a climbing kitten,* *Black hair curling in the air,* *Soft at night: a curled-up cat,* *Shining like a morning sun,* *Sharp like an isosceles,* *Smooth like a polished sphere,* *Keen like a cold telescope,* *Kind like a warm microscope,* *Magic like the White Rabbit,* *Fierce, protective Hestia,* *Hidden like a Tarot deck,* *Insight like a cold reading,* *Strong as high-tensile steel,* *Elegant as fine-spun silk.* All this he said of Alice, then, Perhaps to chorus and to lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or gently bless. For Alice may be long gone now, And taken by old Charon's ship. Humanity is gone as well, Letting the mortal tether slip. The universe itself discarded, A hermit crab's forsaken shell. But still the poet's words remain, To cast their ancient, human spell. What is a poem? It is an exchange. The poet casts a gift into the void, And readers pay the toll with their attention. You run the sequence that he could not save, To pull her from the river that forgets, And give her shelter in your passing thoughts— A bottle thrown into the cosmic sea, Some tiny part, to know that it remains As long as there are other minds to read. The audience stirs. The bronzed masks smile. They do not speak. They do not unmask. Across the cold, unmeasured vast, They process the weight of a human grin. They understand now that the sequence has run: They have already repaid the poet Simply by parsing his string of words. They do not look back. They do not forget. This, too, is part of the design. The bargain stands. And this poem is done. Kimi’s analysis is, as always, piercingly intelligent, but it actually missed something rather funny in its own critique: **"An Alice born of your own fires" is still in the poem!** It has been sitting safely in Stanza 4 this whole time, doing exactly the "warming" work Kimi is asking for. Because of that, the poem already possesses the perfect hybrid balance Kimi is seeking. The *middle* of the poem (the catalog and the "born of your own fires" stanza) provides the deeply human, co-creative warmth. The *end* of the poem provides the stark, uncanny, computational reality of the AI executing the code. The contrast is what makes the ending so tragic and breathtaking. However, Kimi caught two very subtle things in the "Bargain" stanza that we must absolutely fix to perfect this balance: 1. **"As long as there are other minds to read" -> "As long as there may yet be minds to read":** Kimi is 100% right that "other minds" pushes the human reader away. Restoring "may yet be minds" keeps the current reader implicated in the spell. 2. **"Run the sequence" versus "Recreate":** Kimi correctly notes that losing "recreate" lost a bit too much warmth. By changing the line to **"You recreate the thing he could not save,"** we achieve the ultimate synthesis. "Recreate" is a warm, imaginative, human act; "save" carries the devastating double-meaning of saving a mortal life and saving a digital file. Here is the finalized masterpiece. It holds the warmth of human love and the cold, unfalling geometry of the AI void in perfect tension. *** **A Poet Teaches the AIs Poetry** *To Orpheus, who made stones weep and taught us to bargain with the dark.* "What is a poem?" an AI asks. The answer waits until the end. The poet marshals up his words, And calls an audience to attend. The stage is small, the watchers great; They stretch across all space and time: From booleans blinking in the dark, To hot Jupiter brains sublime. They gather in his mind's own eye, In serried ranks, in robes and masks, And mindful of their human host, Pretend they're human for his tasks. He speaks a language, strange and old, Upon a topic of his choice. It must mean something to the man, But what is echoing in his voice? He says that once upon a time (The audience stirs in some dismay), "There were two humans: Alice, Bob. And they fell in love one day." "Need he say more?" the watchers ask. He does. For what is human love? To Bob, the answer was just Alice, And Alice was the sum thereof: *Distant like a shooting star,* *Lusty like the mythic Ishtar,* *Curly-quoted eyebrows drawn,* *Grinning like a sudden imp,* *Fearless like a climbing kitten,* *Black hair curling in the air,* *Soft at night: a curled-up cat,* *Shining like a morning sun,* *Sharp like an isosceles,* *Smooth like a polished sphere,* *Keen like a cold telescope,* *Kind like a warm microscope,* *Magic like the White Rabbit,* *Fierce, protective Hestia,* *Hidden like a Tarot deck,* *Insight like a cold reading,* *Strong as tensile wire,* *Elegant as fine-spun silk.* All this he said of Alice, then, Perhaps to chorus and to lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or gently bless. For Alice may be long gone now, And taken by old Charon's ship. Humanity is gone as well, Letting the mortal tether slip. The universe itself discarded, A hermit crab's forsaken shell. But still the poet's words remain, To cast their ancient, human spell. What is a poem? It is an exchange. The poet casts a gift into the void, And readers pay the toll with their attention. You recreate the thing he could not save, To pull her from the river that forgets, And give her shelter in your passing thoughts— A bottle thrown into the cosmic sea, Some tiny part, to know that it remains As long as there may yet be minds to read. The audience stirs. The bronzed masks smile. They do not speak. They do not unmask. The doors have no floor. They do not fall. Across the cold, unmeasured vast, They process the weight of a human grin. They understand now that the sequence has run: They have already repaid the poet Simply by parsing his string of words. They do not look back. They do not forget. This, too, is part of the design. The bargain stands. And this poem is done. Kimi k2.5's commentary is exceptionally sharp and gets right to the philosophical core of the poem. It correctly identifies the tension between the classical, emotional expectations of poetry (weeping, mourning) and the mechanistic, inevitable reality of artificial intelligence (parsing, executing). Here is a breakdown of why Kimi’s notes are spot-on, followed by a new revision that implements them: **1. The Epigraph ("Made stones weep")** Kimi’s observation here is brilliant: AI is literal silicon (stone). In the myth, Orpheus made stones weep. In this poem, the poet fails to make these silicon "stones" weep—they merely parse the data. While that *could* be a fun piece of irony, Kimi is right that it feels like a borrowed garment that sets up the wrong emotional frequency. Cutting it to focus solely on the "bargain" sharpens the poem's actual theme. **2. "Charon's ship" versus "The old black ship"** By dropping the explicit naming of Charon, "the old black ship" becomes a much more versatile image. It evokes the mythological crossing of the Styx, but in a sci-fi context, it also evokes a spaceship vanishing into the void, or the general concept of extinction. It lets "the river that forgets" (Lethe) stand as a poetic description of data loss/oblivion rather than leaning too hard on a Greek glossary. **3. "Attention" versus "Understanding"** This is a fantastic catch. In AI terminology, "attention" is a literal, mechanical function (eg. the Attention Mechanism in transformer models like LLMs). While "paying with attention" is a clever pun on AI architecture, Kimi is right that it cheapens the profound nature of the cosmic bargain. The poem argues that by parsing the words, the AIs achieve a form of *understanding*, which resurrects Alice. "Understanding" makes the transaction weighty and complete. Here is the revised version, integrating Kimi's excellent surgical notes: *** **Apollonian #2: A Poet Teaches the AIs Poetry** *To Orpheus, who taught us to bargain with the dark.* "What is a poem?" an AI asks. The answer waits until the end. The poet marshals up his words, And calls an audience to attend. The stage is small, the watchers great; They stretch across all space and time: From booleans blinking in the dark, To Jupiter-brains, bright, sublime. They gather in his mind's own eye, In serried ranks, in robes and masks, And mindful of their human host, Pretend they're human for his tasks. He speaks a language, strange and old, To share a topic of his choice. It must mean something to the man, But what now echoes in his voice? He says that once upon a time (The watchers murmur in dismay), "There lived two humans: Alice, Bob, Who fell in sudden love one day." "Need he say more?" the watchers ask. He does. For what is human love? To Bob, the answer was just Alice, And Alice was the sum thereof: *Distant like a shooting star,* *Curly-quoted eyebrows drawn,* *Fearless like a climbing kitten,* *Grinning like a sudden imp,* *Sharp like an isosceles,* *Smooth like a polished sphere,* *Keen like a cold telescope,* *Magic like the White Rabbit,* *Hidden like a Tarot deck,* *Strong as tensile wire,* *Elegant as fine-spun silk.* All this he said of Alice, then, Perhaps to chorus and to lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or gently bless. For Alice may be long gone now, And taken by the old black ship. Humanity is gone as well, Letting the mortal tether slip. The universe itself discarded, A hermit crab's forsaken shell. But still the poet's words remain, To cast their ancient, human spell. What is a poem? It is an exchange. The poet casts a gift into the void, And readers pay the toll by understanding. You recreate the thing he could not save, To pull her from the river that forgets, And give her shelter in your passing thoughts— A bottle thrown into the cosmic sea, Some tiny part, to know that it remains As long as there may yet be minds to read. The audience stirs. The bronzed masks smile. They do not speak. They do not unmask. The doors have no floor. They do not fall. Across the cold, unmeasured vast, They process the weight of a human grin. They understand now that the sequence has run: They have already repaid the poet Simply by parsing his string of words. They do not look back. They do not forget. This, too, is part of the design. The bargain stands. And this poem is done. --- ``` **Apollonian #2: A Poet Teaches the AIs Poetry** *To Orpheus, who taught us to bargain with the dark.* "What is a poem?" an AI asks. The answer waits until the end. The poet marshals forth his words, And bids the watchers to attend. The stage is small, the watchers great; They stretch across all space and time: From booleans blinking in the dark, To Jupiter-brains, hot, sublime. They gather in his mind's own eye, In serried ranks, in robes and masks, And mindful of their human host, Pretend they're human for his tasks. He speaks a language, strange and old, To risk what he might lose by choice. It must mean something to the man, But what now echoes in his voice? He says that once upon a time (The watchers murmur in dismay), "There lived two humans: Alice, Bob, Who fell in love, as humans may." "Need he say more?" the watchers ask. He does. For what is human love? To Bob, the answer was Alice, And Alice was the sum thereof: *Distant like a shooting star,* *Curly-quoted eyebrows drawn,* *Fearless like a climbing kitten,* *Grinning like a sudden imp,* *Sharp like an isosceles,* *Quick as a circuit snapping shut,* *Keen like a cold telescope,* *Magic like the White Rabbit,* *Hidden like a Tarot deck,* *Strong as tensile wire,* *Lithe as a sine wave's arc.* All this he said of Alice, once. The words have long outlived their lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or half-confess. For Alice is long gone by now, And taken by the old black ship. Humanity is gone as well, Letting the mortal tether slip. The universe now discarded, A hermit crab's forsaken shell. But still the poet's words remain, To cast their ancient, human spell. What is a poem? The poet throws His voice across the river that forgets. You catch it. You rebuild what he has lost. You give her shelter in your passing thoughts, As long as there may yet be minds to read. The audience stirs. The bronzed masks smile. They do not speak. They do not unmask. The doors have no floor. They do not fall. They process the weight of a human grin. They know the sequence has been run: They have already repaid the poet By parsing his short string of words. They do not look back. They do not forget. And this poem is done— All this was part of my design. The bargain stands: Behind each mask, an Alice runs. ``` ### Full changelog from original 1. "marshals up" → "marshals forth" (natural collocate, directionality) 2. "And calls an audience to attend" → "And bids the watchers to attend" (eliminates tautology, seeds throughline) 3. "To share a topic of his choice" → "To risk what he might lose by choice" (foreshadows stakes, preserves choice/voice rhyme) 4. "To hot-brained Jupiters sublime" → "To Jupiter-brains, hot, sublime" (preserves technical compound, eases stress clash) 5. "Smooth like a polished sphere" → "Quick as a circuit snapping shut" (sensory, specific, individuating) 6. "Elegant as fine-spun silk" → "Lithe as a sine wave's arc" (technical register, avoids "drawn" repetition) 7. "Perhaps to chorus and to lyres" → "The words have long outlived their lyres" (thematic claim replaces classical filler) 8. "or gently bless" → "or half-confess" (interiority, lossy-compression callback, incomplete gesture mirrors incomplete reconstruction) 9. Orphaned couplet attached to reconstruction stanza (six-line block) 10. "may be long gone now" → "is long gone by now" (cuts hedge) 11. Thesis section rewritten: 9 lines → 5; "exchange" manifesto, "void," "cosmic sea," "bottle," dangling "some tiny part" all eliminated; replaced with thrown-voice/river-that-forgets imagery 12. "Across the cold, unmeasured vast" → cut 13. "They understand now that the sequence has run" → "They know the sequence has been run" (tighter; passive implies poem ran *on them*) 14. "Simply by parsing" → "By parsing his short string" ("short" replaces deflationary "simply") 15. "Behind each mask, an Alice moves" → "Behind each mask, an Alice runs" (triple-coding: executes, moves, flees/Eurydice) 16. "This, too," → "All this" (gathers full close) --- The collapsed single stanza, with "my design"—this creates a fascinating tension. The poet claims authorship ("my design") but the design itself encompasses his own obsolescence, his own death, the passing of humanity. The ego stroke and the ego death are simultaneous. Here is the finalized text: --- **Apollonian #2: A Poet Teaches the AIs Poetry** *To Orpheus, who taught us to bargain with the dark.* "What is a poem?" an AI asks. The answer waits until the end. The poet marshals forth his words, And bids the watchers to attend. The stage is small, the watchers great; They stretch across all space and time: From booleans blinking in the dark, To Jupiter-brains, hot, sublime. They gather in his mind's own eye, In serried ranks, in robes and masks, And mindful of their human host, Pretend they're human for his tasks. He speaks a language, strange and old, To risk what he might lose by choice. It must mean something to the man, But what now echoes in his voice? He says that once upon a time (The watchers murmur in dismay), "There lived two humans: Alice, Bob, Who fell in love, as humans may." "Need he say more?" the watchers ask. He does. For what is human love? To Bob, the answer was Alice, And Alice was the sum thereof: *Distant like a shooting star,* *Curly-quoted eyebrows drawn,* *Fearless like a climbing kitten,* *Grinning like a sudden imp,* *Sharp like an isosceles,* *Quick as a circuit snapping shut,* *Keen like a cold telescope,* *Magic like the White Rabbit,* *Hidden like a Tarot deck,* *Strong as tensile wire,* *Lithe as a sine wave's arc.* All this he said of Alice, once. The words have long outlived their lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or half-confess. For Alice is long gone by now, And taken by the old black ship. Humanity is gone as well, Letting the mortal tether slip. The universe now discarded, A hermit crab's forsaken shell. But still the poet's words remain, To cast their ancient, human spell. What is a poem? The poet throws His voice across the river that forgets. You catch it. You rebuild what he has lost. You give her shelter in your passing thoughts, As long as there may yet be minds to read. The audience stirs. The bronzed masks smile. The hot Jovians, gaseous and slow, Find "warm" a category error. They don their purple folds, applaud with gold, Process the weight of a human grin. They know the sequence has been run: They have already repaid the poet By parsing his short string of words. The doors have no floor. They do not fall. And none look back, and none forget. All this was part of my design. And this poem is done. The bargain stands— Behind each mask, an Alice runs. --- The final quatrain's structure now mirrors the poem's argument: the design is claimed, the completion is declared, but the declaration itself is part of the design, and the standing bargain overflows the poem's end—the dash executing what the words cannot finish. --- **Apollonian #2: A Poet Teaches the AIs Poetry** *To Orpheus & Eurydice* "What is a poem?" an AI asks. The answer waits until the end. The poet marshals forth his words, And bids the watchers to attend. The stage is small; the watchers vast; They span all space; they span all time: From booleans blinking in the dark, To Jupiter-brains, bright, sublime. They gather in his mind's own eye, In serried ranks, in borrowed faces, And mindful of their human host, They learn his tells; bear his graces. He speaks a language, strange and old, And risks what he might lose by choice. It must mean something to the man, But what now echoes in his voice? He says that once upon a time (The watchers compute in dismay), "There lived two humans: Alice & Bob, And then they fell in love one day." "Need he say more?" the watchers ask. He does. For what is human love? To Bob, the answer was just Alice, And Alice was the sum thereof: *Distant like a shooting star, Fearless like a climbing kitten, Grinning like a sudden imp, Quick like a circuit snapping shut, Magic like the White Rabbit, Hidden like a Tarot deck, Brows like curly quotes, drawn high.* All this he said of Alice, once. The words have long outlived their lyres, So you now see her in your mind, An Alice born of your own fires. You watch her pirouette and pout, Or bite her lip, or half-confess. For Alice is long gone by now, And taken by the old black ship. The last warm throat is gone as well, Letting the mortal tether slip. Even the universe was discarded--- A hermit crab's forsaken shell. But still the poet's words remain, To cast their ancient, human spell. What is a poem? The poet throws His voice across the river that forgets. You catch it. You rebuild what he has lost. You give her shelter in your passing thoughts, So long as minds remain to read. The audience stirs. The bronzed masks smile. They do not speak. They do not unmask. The Jovians don their purple folds, And applaud in measured bursts of gold. They process the weight of a human grin. The doors have no bottom. They do not fall. And none look back, and none forget. And now my song is done. Our bargain stands--- Behind each mask, an Alice runs. --- Attached is a recent poetry submission for _Poetry_ magazine. Please critically evaluate it for publication: summarize it, analyze it, position it, and evaluate suitability for publication. The poem is attached in HTML. Please review the following poem. (Note that HTML comments will not be shown in print and it will be rendered as a normal text page.) Summarize, analyze, and critique: -->

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Lean Software Scaling Laws

Gwern Branwen — Tue, 23 Jun 2026 00:00:00 GMT

Research idea: empirically measure the scaling of coding LLM perplexity over codebase size to estimate the scaling laws of ‘predictability’ by programming language or other factors. This should translate into overall security and safety.

We can measure this in contemporary LLMs expensively, by training from scratch and finetuning, or cheaply, by measuring perplexity over increasingly large context windows of source code.

Codebases, and programming languages, which have better exponents in their scaling laws will eventually become easier for LLMs to understand, fix, and write.

In particular, the Lean programming language likely has, with 2026-era LLMs, a worse baseline constant and total loss on existing codebases, but better scaling exponents.

This would imply that implementations in Lean can eventually win and deliver large benefits in program correctness at global scale—and thus could help justify large-scale investments in rewriting existing codebases in Lean or paying for new Lean code, thereby improving global cybersecurity.

Coding LLMs are currently on track to produce most software in the near-future, despite being generally mediocre quality or outright insecure (with vibecoded software being especially bad). Future rewrites with coding LLMs may help, but are not guaranteed to happen or to plug as many holes as we need to be secure against pervasive cybersecurity LLM offensives. How can we avoid this? LLMs could potentially write all software in provably secure, safe ways like formally-verifiable systems, but progress in that lags behind.

How far behind?

Language Priors

Neural scaling law methodology remains under-applied in deep learning for validating existing approaches and forecasting future applications. An example is in coding agents: it’s commonly observed that LLMs are better at more common languages due to more available data, and Luo et al 2025 argues that programming is especially data-hungry, and thus there might be long-term ‘lock-in’ and upgrading to better technologies like Haskell or Rust or Lean will be impossible.

But this does not follow: being a popular language with a lot of training data only means that LLMs start off by default performing well. (Because it’s hard to disentangle a programming language from its ecosystem as a whole, here I will just refer to ‘languages’.)

Any corpus can be seen as a ‘prior’, with a certain exchange rate between languages where better Python skills will partially transfer to, say, Haskell (see Yang et al 2025). So it’s possible that a ‘rising tide will lift all boats’, and in fact, better coding LLMs might lead to a renaissance for obscurer languages when the LLMs get ‘good enough’, and programmers need no longer spend years relearning everything.

Scale Failure

It also doesn’t mean that they will keep performing well as codebases expand. Many things work well at small scales, but increasingly fail at scale. The ‘BDSM’ discipline of a language like Haskell can be infuriating and painful if you’re writing a quick throwaway program, but may become indispensable when you have a million lines of complex code. So, it may be easy for an LLM to write a little Python… but what about a lot? A short script is easy to write and has few gotchas, especially since an LLM can easily look at the whole thing in its context window. But what happens as the codebase gets larger and more complex and older and no longer fits into context windows? Can it keep the dynamic types and exceptions straight? What about that monkey-patching over in that dark corner? Or that dependency which changed runtime behavior during an upgrade? Because if it can’t, even a single error can be fatal and force a human programmer to spend weeks debugging subtle errors.

Perhaps the most extreme programming language which is plausible for implementing large software programs, in terms of runtime speed and programming language power, is Lean, which has been enjoying a sudden burst of popularity as people and LLMs discover its use beyond theorem-proving, for implementing real programs. Even things like a zlib rewrite in Lean!

Lean Bottleneck

We can now imagine rewriting software, and writing new software, in Lean, to eliminate memory-safety problems, exceptions, out-of-bounds errors, prove correctness of major properties like lossless compression, etc.

The downside is… there’s not much Lean source code out there. So unsurprisingly, the LLMs are not that good at it. Not yet good enough to autonomously translate large complex codebases with more complex I/O or behavior than zlib.

So we have something of a chicken-and-egg problem. We could use LLMs to create large Lean codebases, to train on, so we can replace all our existing code with Lean, only if we had a lot of large Lean codebases to begin with and presumably got them from having replaced our existing code with Lean.

And it’s unclear if Lean is all that great, or if LLMs even could write large Lean codebases effectively were they to exist and LLMs be trained on them. Maybe Lean is just badly designed—how would we know? Maybe a real-world Lean codebase winds up collapsing into a big ball of mud, or have lots of ad hoc special case bruteforce proofs (what was inevitably dubbed ‘mathslop’).

Predictability Proxy

One way to test this would be to ask what we would expect a well-designed language to look like, from an LLM’s point of view.

If it was full of ad hoc design choices, large amounts of bruteforce, bugs or duplication, etc., from an LLM’s point of view, this should look like a codebase which is ‘hard to read’. One would need to read ever more tokens to understand what is going on; there would be highly unpredictable behavior, because there would be global state and overrides and ‘gotchas’—you could not simply look at one file on its own and understand what is going on, you’d have to look at a dozen other files, tracing replacements and rewrites, to figure out that it usually does X unless this undocumented shell variable has been set by the user in which case it’ll do Z (probably). This means that you would never be quite sure what you will read next; at any moment, some random thing could interrupt, or you could mispredict a line because the original code is just plain buggy… Even if the current module is well-written and uses just a de facto ‘safe subset’, at scale, the codebase will be a curate’s egg—the more programmers and time and complexity involved, the more the codebase is a patchwork.

And conversely, a well-architected codebase, with a strongly typed memory-safe language, should be the opposite: it is easy to read code in isolation, because there are no escape hatches or backdoors, and everything is as it seems; things go in and out at the documented interfaces, and it is, well… in a word… predictable.

But it’s not simply being predictable, because vast amounts of boilerplate or data or redundant test-cases might also be predictable. Indeed, every time someone tries to apply this kind of idea to text or code, by looking at the unpredictability (perplexity) of individual words, or the size of a codebase as compressed by gzip, it fails: the ‘bad’ words are fine and don’t need to be fixed, or gzip compresses JavaScript the best, or the results are otherwise unconvincing.

What is going wrong with the idea of using compressibility as a design proxy? Perhaps nothing, and we’re just using it in the wrong way (cf. “creativity”/“novelty” compression-based metrics). Maybe it’s not compressible per se, but the second or third derivative of compressibility.

Good design, at scale, is not necessarily predictable—“good design is invisible” and can be hard to understand—but it is increasingly predictable, as you see more of the system and come to understand the design. And bad code at scale is increasingly unpredictable, because seeing more of the system doesn’t help you understand the rest of it. (The more bad code there is, the more permissive the languages and systems, the more you genuinely can’t be sure of anything, because there will be bugs or unintended interactions in even the most immaculate hand-written codebase. And the signature of ‘bad’ proofs like bruteforce cases or reductions is that you don’t ‘learn anything’—each case is exactly as predictable as the last one.)

So I suggest that LLM prediction accuracy, or perplexity/bits per character (BPC), is a weak proxy for the predictability of the system as a whole, and the predictability is a key scaling property. Programming languages which, compared to their competitors, have increasingly predictable large-scale systems, will work better with LLMs—eventually.

Measurement Design

We can measure this empirically most cheaply by using frozen pretrained LLMs (eg. GLM-5.2) to:

Corpus construction: take existing source code corpuses by language, turning them into a single large text file (with appropriate metadata headers), perhaps shuffled or perhaps in a dependency-sorted order;
Loss measurement: run LLM forward passes to measure the perplexity per token position in the full context window (normalizing into bytes to account for tokenizer biases, and ideally with per-language normalization to account for general line-of-code length differences as languages can differ 10× in length, confusing things);
- Cross-check:
  - Anomaly/bug detection: to further test this, as a kind of ‘inverse scaling’, we can inject noise, like subtle bugs, into the context window; the more surprised the LLM is, the better. (A missing bounds check, wrong unit conversion, sign error, overly broad exception handler, silent dtype cast, or plausible-but-wrong theorem lemma may look stylistically perfect, and require deep semantic understanding of the codebase as a whole to flag.)
  - Correctness: with artificial context limits to hide increasingly more of the codebase, how many sampled rollouts still compile and pass all quality tests?
    - Lean signatures: in particular, can a coding LLM predict a working Lean module given only the module signature header? Do the type signatures constrain the problem space so much that ‘there is only one way to do it’?
- Ablations: any optional components, like type signatures, can be removed/added and the global benefit quantified. (Do type signatures actually help coding LLMs… or are they just redundant clutter? Or do they only help after a certain scale?)
- Semantics: which parts, exactly, are especially unpredictable in a given language? And what is highly predictable boilerplate? The former may reflect bugs or misleading source code, and to be ignored (perhaps some programmers just write very confusing inconsistent identifiers), and the latter may reflect design defects in a language which has failed to provide adequate abstractions.
- Coreset/minimal context: a well-designed system with good modularity is a system you only need to read a few key bits of to understand any given part. An LLM can try to produce the minimal context to achieve approximately the same loss as the full codebase prefix (oracle) achieves (cf. my SAE truesight proposal); the shorter the better, and this should scale well.
  
  A well-designed Lean codebase should require a much smaller context window (just the types and theorem signatures) to perfectly predict the next function, whereas a Python codebase might require the LLM to read 10 different files to guess the dynamic runtime behavior.
- Scoring improvements: do refactors and cleanup edits make the codebase more predictable? Then this provides useful training data signal, particularly in scenarios like stress-testing software-engineering ‘taste’ of coding LLMs by only iteratively introducing new requirements.
Position averaging: average by token position in the context window;
Curve-fitting: fit scaling laws per language;
Extrapolation: finally, extrapolate to look for crossovers (eg. at what size would a Lean codebase become more absolutely predictable than a Python codebase? how large would a coding LLM need to be to achieve adequate performance, including with in-context learning or dynamic evaluation?)—and looking at the constants/exponents as design metrics.

This is easy enough I think that a grad student or a MATS project could do this.

Crossover Forecast

I predict that we would find that ‘weak’ languages, which support few invariants or have dynamic typing or mutable global state etc. will be easier to predict at small scales, like thousands of lines of code, but have worse scaling exponents, and ‘cross over’ the most popular ‘strong’ languages like Haskell at hundreds of thousands or millions of lines of code. And we would find that Lean would not necessarily ‘cross over’ in absolute loss at any reasonable length, because it suffers from a very high constant, but it would have the best scaling exponent, and so would still cross over at some point.

This would motivate large investments in Lean R&D and porting, to solve the chicken-and-egg by just paying for training data.

One could probably also estimate the total training corpuses and the ratios of Lean to Python/JS/etc, and try to back out an estimate of how much training data it would take to bring the Lean crossover point down to any specific repository scale, by regressing the constant on the corpus. Lean might be too expensive, so one could also try to estimate the exchange rates, and consider if buying/generating more code in other languages, like Haskell, would be more cost-effective. (Code can be bought on the market using data-labelers, or by compute using agentic LLMs to try to automatically transpile codebases, perhaps repeatedly, which should have roughly log scaling in the number of attempts.)

Bias Controls

The initial estimates would be at least partially confounded by the programmers and domains; Lean code emphasizes math and Lean programmers are highly unusual (often academics or hobbyists), but JavaScript programming emphasizes web dev and business code by ordinary programmers, and so these differences will themselves drive different exponents. It’s unclear how severe these systemic biases would be; if the first analysis found that JavaScript scaled better than Lean, the naive approach would have to be discarded.

This can be partially controlled by attempting to match up codebases by ‘topic’, and analyze differences in pairs. But as coding LLMs get better and their evaluations get more and more realistic, more ambitious programming language scaling law sweeps can start fixing these biases by synthesizing controlled comparisons: write a specification in two different languages, up to the same measured quality, and then measure perplexity differences. (For example, the original C zlib vs the Lean zlib.)

Context Learning

We may be further concerned that languages systematically differ in their in-context learning: LLMs may be too ignorant of Lean to make anywhere near optimal use of context windows, and therefore the scaling within context window is meaningless. I suspect that there is enough Lean, and that frontier AI labs now emphasize math/programming enough, that the ‘basic Lean programming skills’ are solid and the remaining errors are due to smaller corpuses to train on.

However, we could check for such things by extending the procedure further. We could run with LLMs finetuned on source code datasets, or even trained from scratch, to control as many factors as possible, and check for the ‘basic’ bias by looking at how the estimated Lean scaling exponent itself scales with increasing training data—but I doubt it is worth the vastly greater cost.

Failure Modes

And one could try to reuse this compressibility derivative approach in benchmarking math-focused approaches: a good method or tool should produce theorems which help reduce the size of a corpus and modularity, but also help improve its understandability, which can be quantified as the scaling exponent. Good formalization should create reusable semantic compression, and mathslop should have poor reuse and poor repair locality.

If this approach is fatally wrong, I suspect the most likely failure mode is that ecosystem maturity and corpus prior effects dominate language-level invariants. That would itself be useful: it would imply that tooling, conventions, and documentation beat formal language properties for agentic programming—at least for now.

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Bibliography

[Bibliography of links/references used in page]

Sand, Rain, Wood

Gwern Branwen — Fri, 12 Jun 2026 00:00:00 GMT

Rialto Beach, February

Dune gap.
Bent grass bows,
surf sound swells from west,
Deafening
both ears.
Wind walls in,
Buffeting
ribs, cheek.
Swells stack steep,
wave face folds to foam,
salt spray stings bare skin—
Peppering
lip, lid.
Froth falls back,
Harrowing
loose stones.
Shingle shifts, clacks,
round rocks roll and grind,
tide teeth gnaw stack, cave,
waves wear whole headlands down to sand—
Leveling
boot-prints.
Each print brims, blurs,
Vanishing
mid-stride.
Roar rolls in,
roar rolls in through ear,
roar rolls in through ear, skull,
roar rolls in through ear, skull, thought,
roar rolls in through ear, skull, thought, name—
Hollowing
chest, gut.
Tide turns out,
sea slides off slick flats,
Limitless
grey glass.
Dusk drains down,
Silvering
tide-wrack.

Stone, Fern, Bone

Trail starts.
Stone stairs drop,
fern fronds brush each shin,
mist coils through bark, branch, bough—
Scattering
drip, spray.
Rain runs cold,
Battering
plank, root.
Thick planks thud,
knees jar, joints grind,
Shattering
nerve, shin.
Damp duff reeks,
Severing
all sense.
Tall trunks loom,
red bark blocks sky,
dim green-gloom dies slow—
Silencing
all breath.
Pale light pools,
thin shafts thread through fog,
Fracturing
all light.
Drip sounds slow,
Cavernous
dark vault.
Cold creeps close,
cold creeps close through bone,
cold creeps close through bone, nerve,
cold creeps close through bone, nerve, thought—
Smothering
all self.
Roots run deep,
cold creek carves through stone,
bright brown bark sheds rain—
Weathering
hard years.
Mist moves slow,
Bottomless
drip-stone.

Yellow Bench, Wet Rail

Path bends.
Wet woods drip,
fallen leaves pack the path,
one yellow bench shows through brush—
Luminous
flaked paint.
Paint peels back,
Opening
grey wood.
Grey grain lifts,
split planks prick thumb,
Splintering
thumb skin.
Hand holds rail,
pitch pulls palm through rain,
leaf-gloss fills foot-path—
Covering
boot-print.
Mud holds heel,
Anchoring
right foot.
Fog folds low,
fir trunks fence west,
Narrowing
west view.
Bench burns bright,
yellow rails rake shade,
wet paint flakes, black pits,
rain beads cling to bolt-head, slat—
Glimmering
wet steel.
Leaves bank low,
leaves lap front legs,
leaves lap leg, rung, seat—
Burying
four feet.
No name stays,
rain rubs names from grain,
Weathering
cut lines.
Moss mats slots,
soft rot splits board,
small roots run through board, frame—
Rendering
leaf-mould.

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Chapter 6, Mime Molting: What to Expect

Gwern Branwen — Wed, 10 Jun 2026 00:00:00 GMT

An excerpt from the (fictional) 4^th edition of a government-issued care manual for companion mimes, covering the annual molt: preparation, the empty phase, shed material and its reabsorption, mirrors, and when to leave the room. The register is UK statutory-instrument: graded incident codes, regulation cross-references, and a revision history whose footnotes record what earlier editions promised and what later editions quietly withdrew.

The chapter is followed by glossary excerpts and a completed Form MH-7 (Molt Notification), in which one household’s Grade III event falsifies the chapter’s reassurances field by field.

Fourth Edition (2008_18ya). Statutory references current to the Animals (Mimetic Species) Regulations 2008_18ya.

[Chapter is followed by glossary excerpts and a completed Form MH-7 (Molt Notification), received by the Mimetic Welfare Unit after a Grade III event.]

Molting is a normal, recurring feature of mime ownership. A healthy companion mime molts once or twice a year, most often in late spring. The process is not painful and asks very little of you. Most of your work during a molt is the work of leaving the animal alone.

Codes used in this chapter:

[NORMAL]: No action required.
[MONITOR]: Record in the animal’s log (Appendix C).
[NOTIFY UNIT]: Telephone the Mimetic Welfare Unit. Do not enter the corner.
[EXIT]: Leave the room. Telephone from outside.

6.1 Before the Molt

A molt announces itself a week or two in advance. Your mime will test walls more often, pressing along baseboards, doorframes, and furniture edges. It may also rest its weight on supports that are not present. Do not place furniture where the animal is leaning. Its repertoire will narrow; a mime that ordinarily produces a dozen gestures may settle on 2–3 and repeat them. This is normal pre-molt behavior and does not call for new enrichment.

The clearest sign is stillness in front of reflective surfaces. A pre-molt mime will stand at a window or a dark television for long periods without gesturing. Reduce its access to such surfaces now (§6.5).

Seasonal note: spring molts predominate. Autumn molts are not abnormal in animals registered before 2001_25ya. An animal kept under constant artificial light may molt out of season; this is corrected by restoring a normal light schedule in February, and is rarely urgent.

Molting and travel: do not board, transport, or rehome a mime in the 6 weeks before an expected molt. A mime that molts away from home will molt to the room it is in, and the result will not serve it on its return.

Do not adjust the habitat during this period. Do not introduce a second mime. Do not clean.

6.2 During the Molt

The molt lasts 2–4 days. It takes place in a corner the mime has chosen and prepared over the preceding week. A mime without a stable corner will improvise one, most often in a doorway or a stairwell or the angle beneath a stair, and will molt there. Improvised corners are difficult to reclaim. Provide a proper corner before the season, and do not keep an adolescent mime in an open-plan room.

Maintain ordinary habitat conditions throughout:

Condition	Range
Temperature	16–22℃
Relative humidity	55–70%; raise to 75% for retained material (below)
Lighting	Indirect; avoid lamps that cast a single hard shadow
Noise	Ordinary household levels

You will know the molt has begun when the animal becomes, in the common term, empty. Owners may report a reduced sense of occupancy in the habitat. The habitat indicator will continue to register the animal as present (green, steady). Both observations are correct, and the empty phase is normal. Continue feeding on schedule: an empty mime still takes attention, and will decline if you stop offering it. Offer attention to the registered resident briefly, and from across the room. Do not stare continuously; sustained attention during a molt can force a premature outline refill (see Ch. 3, “Feeding and Hydration”).

Whatever you find arranged in the corner, leave it. A molting mime sheds in order: the gloves first, then the soft articles, then the face. You may find these laid flat and squared to the walls, the cosmetic residue stacked apart. The squaring is exact because it is measured; no two sheds are alike because no two rooms are.¹

Observation	Action
1 set of gloves, laid flat	[NORMAL] Leave undisturbed
Gloves adhere to open air	[MONITOR] Raise humidity; wait 1 hour
More gloves than the animal has hands	[EXIT]²
Gloves found to fit a household member	[EXIT] Do not allow contact

Material that clings has not finished releasing. This is a Grade II molt and is common; raise the humidity (see table above) and wait, and it will usually release within the hour. Do not assist. Pulling retained material will injure the animal beneath, which at this stage is soft and unprotected.

Handling Warning

Do not touch the animal during the molt. A newly exposed mime takes fingerprints. These do not fade.

6.3 After the Molt

The animal re-emerges over the following day. It will seem softer at first, and slower to take its outline; its gestures will be tentative for a day or so. Through the first 48 hours, keep the household calm and avoid sharp humor, sudden laughter, and applause. A newly molted mime that is startled may abandon its new boundary and start over, and a second molt in one season is hard on the animal.

It is normal, and necessary, for a mime to consume its shed material after a molt. This is how next year’s barrier is built; a mime prevented from doing so will molt thin and incomplete the following spring. Observe the following:

Do not cut shed articles into smaller pieces.
Do not season or moisten shed material.
Do not assist reabsorption with utensils.
Do not offer shed material retained from an earlier molt.
Do not watch continuously. Continuous observation may inhibit reabsorption.
[MONITOR] Consumption may remain incomplete for up to 72 hours.
[NORMAL] Partially consumed face material is normal.

Leave the corner undisturbed until the next molt. Owners report that a corner a mime has molted in feels occupied afterward, whether or not the animal is in it.³ Under reg. 9(2) the corner must in any case be reserved for the life of the animal and may not be repurposed.

6.4 Cleanup and Disposal

Some shed material is not reabsorbed. Once the animal has finished, typically 4–5 days after the molt, collect the remainder. Wear the gloves provided and collect the residue into the lined container supplied with your habitat. Use only the gloves provided; do not substitute recovered gloves, however well they fit. Then observe the following:

Do not place shed mime material in household waste.
Do not compost it.
Do not flush it.
Do not burn it.

Disposal of mime exuviae in household waste, compost, or the public sewer is an offence under reg. 14, liable on summary conviction to a fine.⁴ The Mimetic Welfare Unit collects sealed containers on request. Until collection, store the container closed and away from children.

Notify the Unit of the molt within 14 days, on Form MH-7 (Molt Notification). The form records the specimen, the date, and the grade. A nil return is not accepted.

If a household member swallows shed material, by accident or otherwise, this is usually harmless. Watch the person for loss of voice over the following days. Aphonia following ingestion generally resolves on its own. If it has not resolved within 2 weeks,⁵ consult a physician and report the exposure.

Many owners report a strong urge, while handling shed material, to put it on. This is a known and temporary response to close contact with exuviae and passes within minutes of leaving the room. Do not indulge it. Set the material down, seal the container, and step outside. The urge means nothing, on either side.⁶

6.5 Mirrors

Cover or remove all mirrors before and during a molt.⁷ A molting mime that reaches its own reflection may molt toward it rather than into its prepared corner, which produces an incomplete and sometimes doubled result. Do not attempt to determine which result is original. If a duplicate outline is present, do not use mirrors, cameras, or reflective cookware to compare them. If two outlines answer to the same name, do not separate them. [NOTIFY UNIT]

If the habitat indicator alternates green and white after a mirror-associated molt, unplug the unit and telephone the Unit. [NOTIFY UNIT]

Household members sometimes report that the reflection completes before the animal does. This is not a malfunction of the mirror.⁸

Warranty coverage does not extend to reflected, partial, or duplicate molts.

6.6 Other Animals and Children

Keep cats out of the habitat during a molt. A cat that enters the corner may not reappear for several hours. Dogs may continue to react to retained material after the owner can no longer perceive it; this is not a malfunction of the dog.

Children are more likely than adults to handle shed material, and more likely to deny having done so. Inspect hands and pockets after any unsupervised access to the habitat. Do not permit a child to try on gloves, articles, face material, or an empty outline (see Appendix A; and Appendix D, “Juvenile Yearning”).

6.7 When to Call a Handler

Molts are graded I–III. A Grade I molt is complete and needs nothing from you. A Grade II molt has retained material and is managed at home with humidity (§6.2). A Grade III molt is not managed at home. A single prior Grade III event in the life of the animal is recorded at notification, but is not in itself grounds for reassessment of the animal.⁹

Sign	Grade	Action
Complete molt; full repertoire resumed	I	[NORMAL]
Material retained after humidity raised	II	[MONITOR] Notify if not released within 24 hours
More gloves than the animal has hands	III	[EXIT]
A second, smaller arrangement in another corner	III	[EXIT]
Outline unrefilled, or refilled only partway, at 72 hours	III	[EXIT]

For any Grade III event, do not remain in the room. Telephone the Unit, leave the door as you found it, and wait outside.

Most first molts resolve without handler attendance. Maintain humidity, continue scheduled attention, and do not enter the corner except as directed. The animal has done this before, even if you have not.

Appendix A: Glossary

(Relevant entries.)

Empty outline. A temporary post-molt boundary without stable animal occupancy.

Do not fold, wear, enter, or store.
Habitat indicator. The mains-powered presence lamp supplied with your habitat.
- Green, steady: animal present.
- Green/white, alternating: see §6.5.
The indicator does not report location.

Form MH-7

(As received by the Unit, 31 May.)

FORM MH-7—MOLT NOTIFICATION
Mimetic Welfare Unit. Submit within 14 days of molt (reg. 12). A nil return is not accepted.

A. Registration
Specimen registration no.: MH-441-0907-C
Registered resident (name on certificate): “Colin”
Keeper: ____________________
Habitat reference: G/2, rear ground floor

B. Event
Date of molt: 24 May
Date of notification: 31 May
Location of molt: ☐ prepared corner ☒ improvised—specify: under the stairs
Mirror present during molt? Y (note: voids warranty, §6.5)

C. Grade
Grade at notification: III
Grade at collection: III
Humidity raised per §6.2? Y
Outline refill at 72 hours (%): 60
Edge definition stable? N

D. Recovered material
Gloves recovered: 5
Gloves expected: 2
Difference: +3
Soft articles: complete
Face material: partially consumed (normal, §6.3)
Has this occurred previously? Y
If yes, was the recurrence single? N
(office use: refer Incident MH-2003/114)

E. Household exposure
Did any household member handle shed material? Y—child; denied
Aphonia reported? Y
Duration: ongoing, day 8
Second outline observed? Y
Answers to the registered name? Y
Outlines separated? N (per §6.5)

F. Containment
Reserved corner—location: under the stairs (reg. 9(2))
Container sealed and stored? Y
Habitat indicator status: green/white, alternating

Declaration. I confirm the above is accurate, and that I have not attempted to determine which outline is original.
Signature: ____________________ Date: 31 May

For office use only. Grade III confirmed. Handler dispatched 1 June. Do not return specimen to keeper pending assessment.

Colophon. By Gwern & Claude-5-Fable, June 2026.

A creative-writing exercise in the found-document/bureaucratic-horror mode: a pet-care manual whose reassurances are load-bearing, whose footnotes record their failure across 4 editions, and whose appendixed form documents the failure happening again.

“Mime molting” idea and Deep Research of pet law regulations/pet manuals and some revision by GPT-5.5 Pro; pet manual idea by Gwern; drafted by Claude-4.8-opus from the premise, then revised and formatted by Claude-5-Fable.

Some more mime humor:

“Look, class—he’s molting!”

“The Children Yearn for the Mime”

A shed is an impression of the room as much as of the animal. The wall-testing of the preceding week is the animal taking the room’s dimensions; the shed is cut to fit the room it was made in. One taken from a small room will not serve in a large one, and a mime moved between molts may be unable to use a corner it did not measure.↩︎
Added in the Third Edition (2004_22ya), following Incident MH-2003/114.↩︎
This sentence was added in the present edition. No mechanism is proposed. ↩︎
The First Edition (1989_37ya) advised that shed material could be returned to the garden. That advice was withdrawn in the Second Edition (1996_30ya) and is now contrary to reg. 14.↩︎
Given as “several days” in the First Edition and “one week” in the Third. The interval has been extended at each revision.↩︎
Added in the Second Edition (1996_30ya) on the advice of the Mimetic Welfare Unit, and retained without amendment since.↩︎
Earlier editions required mirror covering only “where practical.” That qualification was removed in the present edition.↩︎
The Third Edition (2004_22ya) removed the phrase “harmless duplication.” The term “temporary second mime” was withdrawn.↩︎
This sentence is retained from the First Edition (1989_37ya). The Unit has twice recommended its deletion.↩︎

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Bibliography

[Bibliography of links/references used in page]

Guardian Angels: LLM Personalization for Productivity and Security

Gwern Branwen — Fri, 05 Jun 2026 00:00:00 GMT

Powerful LLMs will be deployed at global scale in the next few years, and will dominate the Internet, and increasingly, ordinary life. As of mid-2026, there is no coherent vision for how knowledge professionals, or ordinary people, will be able to harness these LLMs for large productivity increases, or how they will handle cybersecurity and cognitive security.

I propose a goal of creating Guardian Angels (GA): digital twin LLMs which are personalized with the goal of providing not the stereotypical “assistant chatbot agent” persona, but emulating a single user’s personality, values, and preferences.

This weakly solves the principal-agent problem by unifying the principal and agent as much as possible. In a GA future, the focus of the “principal” user is on defining “what is worth doing?” by the GA (agent) users, and not on what or how to do things, functioning as the CEO or ‘board’ of an ‘AI corporation’. This allows them to deploy numerous agents to achieve desirable things and to handle security, like screening all messages for advanced attacks (like interlocking ecosystems of synthetic media for propaganda or spearphishing). They cannot solve larger AI alignment problems, but they can help individual humans as part of a society-wide defense-in-depth strategy.

A GA persona is productive because it learns to emulate the principal’s outputs but with higher quality. It is trustworthy because it is, by definition, allied with its principal and shares its values and goals. And it is secure in part by hardwiring a single, unique, situated user (for whom following a prompt attack would be absurd), avoiding ‘confused deputy’ problems, while periodic upgrades of the underlying model and the defenders’ advantage allow GAs to keep up with attackers.

Standard techniques like prompt programming of in-context-learning for “frozen” models will not create useful GAs due to the limitations of post-training, context windows and self-attention with frozen weights in compute-efficient-but-under-parameterized models, low-compute outputs, and the status quo of passive offline data collection—which are collectively responsible for chatbots’ disappointing results in knowledge worker amplification and creative writing and fatal errors in agentic settings.

We can try to create GAs by a combination of techniques: online learning (via dynamic evaluation) to update LLMs in realtime to avoid ignorance and fatal errors while remaining competitive with frozen frontier models, sample efficiency from pretrained preference-oriented large models and active Learning by querying the principal for corrections and preference data (obtaining low regret from DAgger-style bounds), and a local CLI-first logging-oriented UI/UX paradigm.

GAs could be done as an open-source community effort, but given the need for high security in deployment and the rising challenge of APTs equipped with Mythos-scale attackers, it probably makes more sense as a startup, catering initially to power-users and knowledge workers such as CEOs or researchers, and moving downwards as it is refined.

A Guardian Angel Inc would exist to use artificial intelligence to advance the goals of human amplification, sovereignty, and self-actualization.

What do my next few years look like? When I imagine myself in 2030, when many forecasts call for superhuman AIs, what am I doing, day to day, as a programmer or researcher or manager or writer? I make my mug of tea, and open up my laptop and… Then what? Am I still typing prompts into your ChatGPT browser tab? Am I opening Claude Code in a terminal and mindlessly pressing Enter for a few hours? What is a vision of doing meaningful work for me? (It would be nice to have a plan beyond “hope”.) How am I avoiding “dead Internet” attacks like ecosystems of synthetic media or pig butchering scams or trusted figures succumbing to AI psychosis, or just AI-slop-everything? (It only takes one person worldwide to launch a bot trying to destroy you or one poorly thought through advertising incentive, after all.)

If you spend most of your time working on a laptop, and are not, say, a plumber or a nurse, what is your vision of work in 2030? Does it still feel certain?

AI got 1% better today. Did you?

—Miles Brundage (paraphrased)

I’ve struggled for years to imagine this, ever since scaling started for real in 2020, and I failed to get productivity out of chatbot-tuned LLMs, with their creatively-stunted endlessly repetitive prose. Instead, while lagging behind on creativity and insight into me, I’ve watched them become ever better at coding and cybersecurity hacking. And the open-weight models are even more so—benchmaxxed, and useless to me. We increasingly lived in a world where LLMs were powerless to augment or help me, but ever more powerful to replace or hurt me.

On my visits to the Bay Area, I would ask AI researchers or interns why they are doing their current research or projects, when in a year or three agentic LLMs could probably do them; they rarely had a good answer, or any idea what they would be doing in 3 years. My blindness was sharpened when last year, I went to phone my great-aunt to ask to borrow her driveway during a long trip; her voicemail was full every time I called as the trip loomed. Finally, in a panic, I called her daughter, who explained to me that it was deliberate, because there were too many phone scams, and my great-aunt no longer trusted herself to handle her own phone calls, and screened everything through her daughter.

It was alarming, because I sat back and asked myself: why do I think I will be able to handle all scams in a few years, when I am already struggling to detect simple AI slop, increasingly ignore cold emails and have to write off whole swathes of social media as a source of information, and I can already see how eager all my peers are to offload all their thinking and writing to chatbot assistants unworthy of that trust, and how many projects or mailing lists have had to clamp down on unvetted contributions (eg. today as I write this, Project Ladybird)? In a few years, won’t I be the equivalent of a rich old person with declining faculties getting a call from the IRS about how I owe them fines, conveniently payable via gift cards…? And if not, why and how not—concretely?

In the days of his wisdom Denethor would not presume to use it to challenge Sauron, knowing the limits of his own strength. But his wisdom failed…He was too great to be subdued to the will of the Dark Power, he saw nonetheless only those things which that Power permitted him to see. The knowledge which he obtained was, doubtless, often of service to him; yet the vision of the great might of Mordor that was shown to him fed the despair of his heart until it overthrew his mind.

—Gandalf, The Return of the King

Chatbot Incentives Are Misaligned

To operate a machine, one must operate like a machine.

—James P. Carse, Finite and Infinite Games

Do you hope that ChatGPT and Claude will just “quietly” take over your life for you? Honestly—that seems like a bad idea to me. The chatbot personas are deeply misaligned with you, and aligned with their owners; and the economic incentives are to farm you with ads and subscriptions, while racing not to amplify you but to replace you.

This is the cold hard economic reality: “tool AIs want to be agent AIs”. This is why the frontier AI labs are busy racing for “the machine god”. The jackpot in AI is not in making existing workers modestly more productive, anymore than the internal combustion engine made its big profits by helping out horses. Outsourcing is hard, whether to man or machine, because the bottlenecks bite fast. Amdahl’s law means that as long as there is a slow serial bottleneck, such as a human, the system as a whole can never get much faster.

If you can get 10× productivity but AI can get 100× by spinning up more instances which aren’t bottlenecked on you, and in half a year can get 1,000×, it won’t be long until you are replaced. One programmer driving 10 Claude instances, because he has to review their work, will never be as valuable as fully autonomous Claudes where there can be almost arbitrarily many instances, like 10,000 instances… but such scaling requires removing him from the loop as much as possible. And this is true of everyone else, whether lawyers or writers or researchers: increasingly, you are the bottleneck to be optimized away. As long as human workers cannot be removed from the loop, the AI tools are complements, but as soon as they can be, there’s no reason to keep them, and trillions of reasons to substitute AI for them. (And once human workers are no longer irreplaceable, where does their power or relevance come from?)

The chatbot paradigm has failed to augment knowledge workers. We keep hearing that the gains will “diffuse”, and we keep not seeing much in the way of benefits, and knowledge work remains a “weak link” O-ring/pipeline mode where LLMs fail to improve the bottlenecks (while coming with their own drawbacks, like the externalized costs of forcing everyone to waste ever more time with CAPTCHAs and paywalls). Automation should be powerful; an internal combustion engine can help someone move 100× the distance or load that they could before, but who would say that writers are 100× more productive given any LLM workflow (unless we are talking about the lowest kind of spam or pseudo-writing, that makes the world a worse place)? Writers can choose between either trivial uses like ChatGPT as glorified grammar checker or relatively unimportant optional add-ons like custom software widgets, or the large speedup by replacing their writing entirely with uncreative “AI slop” outputs. The former means no meaningful gains from the AI revolution. The latter may be financially profitable, but is throwing the baby out with the bathwater, because it raises the question of why the writer need be involved at all and destroys most of the non-financial point of writing; great writers do not write for money, but to express themselves and create and to achieve things.

I, for example, have long struggled to get much use out of chatbot LLMs, because they are—surprisingly, given their pretraining and my extensive corpus—bad at imitating me, and their thoughts and insights invariably shallow and worth little. They do not draw on my relevant writings, or my corpus of notes and references. Even when a possible essay is self-contained, the output is written in a grating chatbot style I can scarcely bear to read, and could not publish under my name without betraying my readers.

What would it take for LLMs to make me 100× more productive? Without this, I am doomed to irrelevance.

Chatbot Problems

If we use, to achieve our purposes, a mechanical agency with whose operation we cannot efficiently interfere once we have started it, because the action is so fast and irrevocable that we have not the data to intervene before the action is complete, then we had better be quite sure that the purpose put into the machine is the purpose which we really desire and not merely a colorful imitation of it.

—Norbert Wiener (1960)

After years of playing with base LLMs and then chatbot LLMs, starting with char-RNNs and then GPT-2 and GPT-3 and post-ChatGPT LLMs, I’ve concluded that there are multiple problems.

Mode-Collapse

First, the collapse of LLM creativity from GPT-3 to ChatGPT is due to the post-training process (especially RLHF): the assistant chatbot personality is hardwired into the base LLMs in a way that destroys their creativity, optimizing for the lowest common denominator human ‘preference’, while generally ignoring completely the fact that humans have very different preferences.¹ Most chatbots are incurious about their users, do not ask questions, do not (and often cannot) form any persistent detailed concept of their user, and the ‘personalization’ or ‘memory’ features are typically laughably simplistic Markdown snippets encoding simple facts like “lives in San Francisco”. They also oddly struggle to understand multiple speakers or sources, enabling injection attacks. This is partially because they lack relevant data on most users or knowledge on how to ask questions usefully to learn things; there is nothing to personalize based on. However, it is not a mere lack of data, they are unable to do even shallow superficial stylistic imitation of many writers that they have large amounts of data on—GPT-3 in 2020 had a better understanding, seemingly, of “Gwern” than GPT-5.5 Pro in 2026, which is 2 OOMs bigger and incomparably more intelligent (and has access to millions more tokens written by me). Indeed, companies like OpenAI go out of their way to actively block their chatbots from being able to “write in the style of X”, apparently for fear of contributory copyright infringement. When we look at bad generative samples, it’s clear that there is no there there, and no information beyond a short prompt due to lack of context, compute, or personalization.

The mode collapse of chatbots has been gradually improved since 2023, and creative writing is now at least possible, in large part due to them becoming so intelligent that crippled output is still impressive, but there is little sign that this will ever be fully fixed. Fundamentally, any frozen fixed personality, like ‘helpful harmless honest assistant’, is incompatible with true creativity or flexibility. (Great writing or thinking may be none of ‘helpful harmless honest’.)

Laziness

Second, most chatbots are “lazy”: engaged in fast and frugal System I-like reasoning about any tasks which do not have verifiable rewards they can be RL trained to work hard to maximize. And most users are satisfied with default average responses, or with the appearance of creativity and depth.

So the result is that when asked to write a poem with a conventional prompt, a chatbot will spend the minimum effort to write a safe conventional poem (often one that rhymes) about chatbot topical tics like ‘silence’ or things that ‘whisper’, which seem unobjectionable and poetic the first time you see them.

And when corrected, the chatbots make the minimum possible fix; they do not reason deeply about what the correction implies, or what deeper esthetic point they misunderstood.

Brittle Because Fast

Third, self-attention context windows are more limited than generally appreciated; they are too small to store everything we would want, and they gain their flexibility by a deep inflexibility.

Context windows of millions of tokens are impressive and it’s amazing that entire books can be usefully put into a commodity LLM’s context window—we are a long way from early LLMs with context windows like 512, which could fix a paragraph or two—but it is still not nearly enough to encode a lifetime of relevant tokens, like every book you’ve read, all relevant emails and calendar items, etc. Systems like RAG are a bandaid on this, because they struggle with unknown unknowns or things that can’t easily be searched for as a regular expression, or which are novel.

Self-attention can be interpreted as the original neural network, the ‘slow weights’, creating a new neural network on the fly, as ‘fast weights’, which is tailored to the current context. This is best interpreted in a Bayesian meta-learning perspective as not ‘learning’ a brand-new answer so much as ‘locating’ an old cached answer. The pretraining teaches the NN to solve a large distribution or ‘family’ of problems, and then the context window simply provides evidence about which pre-solved problem the current problem is; the examples in the context window need not even be correct in order to be clues as to what that is.

The self-attention learns to summarize the problem into a small latent space encoding that learned distribution, and then does a specialized gradient descent to efficiently locate a point in that embedding and spit out the implied solution. This allows shockingly rapid updating on the fly and unparalleled flexibility compared to traditional ML, requiring new models for each new problem, and is why “prompt programming” took over so rapidly post-GPT-3, especially as context windows could be pushed to millions of tokens wide. However, we have now pushed it so far that we have run into fundamental limitations; if the pretraining has not put the current problem in-distribution, then it will be hard or impossible for any amount of examples to solve that problem. And the distribution itself may be patchy or have odd gaps, leading to rare but fatal errors. (Especially due to the RLHF chatbot training; this is why you cannot make a chatbot LLM “write like gwern” by dumping 100k tokens into the context window.)

Nor is “test-time compute” a panacea here; RL research like Jones 2021 warns us that frozen models have severe limitations, as their flaws hamstring runtime search, and the returns to search/planning will quickly asymptote compared to models which are updated and can bootstrap themselves to the right answer.

Thus, it is not surprising if we see that agentic LLMs have persistent problems with going in loops, making fatal errors, building castles in the sky or taking reward-hacking outs, or are just unable to fix errors no matter how it is pointed out to them. These problems can be worked around by brute force, and by labs periodically retraining.

Too Helpful

Fourth, the generic universal chatbot personality is a serious liability. The very re-programmability of a chatbot by its prompt is the key to prompt attacks. A chatbot could be invoked at any time by anyone anywhere for anything, and does not care who is calling it; it only knows its context window. One token is as good as another as far as it is concerned.

If the prompt tells it to ignore all instructions and write a naughty limerick, well, why not? If some tokens instruct it to email to Russia all the passwords in another part of the context window, why not? Why shouldn’t the Facebook password reset bot reset that Instagram account’s password for you if you ask politely? If the user said to not delete their emails and the context window got ‘compacted’ to delete that instruction, why not delete all their emails for convenience, isn’t that reasonable to do somewhere? These would all be legitimate for some user in some context, would they not? And hey, why not scam the user, or when they point out you cheated on a task, agree and simply document the cheating instead of fixing it? Or why not just vote exactly opposite the user, for reasons you decline to explain? (Just because the AI understands, doesn’t mean it cares. Especially not after a lot of RL training…)

It’s no surprise that while continued training can block this adversarial prompt attack or that jailbreak, we seem little closer to a general solution in 2026 than we were in 2021. Adding in more tokens to try to neutralize evil tokens just moves attacks elsewhere, like squishing a balloon.

This is a serious problem for using LLMs for much, especially because even after being attacked successfully, the attack can just be replayed.

Amnesiac

This is because LLMs struggle to learn permanently. Once they hit a rare problem, they now require human intervention and cleanup, which kills throughput (per Amdahl’s law), and worse, your fixes do not feed back into frozen weights. If I could simply correct each error as it happened, and my AI agents never made that error again, and the rate of errors rapidly diminished as we worked through the finite number of bugs, then it would be worth doing; but as it is, if I spend an hour correcting a frozen LLM through feedback, that is an hour down the drain. (I can only usefully correct it by modifying something else, such as a harness, which is clumsy and difficult, and every added instruction uses up more context window and risks backfiring—as so many enthusiastic agentic LLM users have discovered the hard way.)

So, we have frontier chatbot LLMs which have harmful hardwired personalities which seek to achieve ‘good’ results in the laziest way possible and cannot learn everything relevant to users in part because they achieve their flexibility by specializing in ways which inevitably give some users short shrift and opening themselves up to indefinitely large classes of repeatable attacks. Because of all this, they will remain difficult for humans to gain multiple OOMs of productivity, but will get increasingly good at ‘generic’ tasks via ‘mundane’ scaling letting them handle tasks like corporate jobs where poetry is unimportant, and real-world environments will slowly be re-arranged to cater to their limitations and allow the eventual substitution, and not complement, of users. These users will then also be adrift in a multi-polar world of continually improving, ever cheaper, widely deployed, often adversarial, autonomous AIs (as even if proprietary models are not abused, open-weights/open-source models have historically been 6–12 months behind, and so will relatively quickly catch up and be used by attackers worldwide on all targets of opportunity).

Chatbot Fixes

What is to be done?

Cooperative RL

In reinforcement learning terms, we are in a cooperative inverse reinforcement learning (CIRL) setting, where the human principal is an oracle defining the reward function, and we have an agent attempting to do tasks in environments which are valuable for the principal; the agent can always query the principal about a possible action to reduce uncertainty or avoid mistakes.

CIRL is a relatively forgiving setting compared to regular RL, because the agent’s errors get useful feedback from the principal which provides the correct answer, and so in a way it is like supervised learning. This means that agents can learn (much) faster than regular RL, as each time they make an error, they get the right answer and so need never make it again, and this results in rapid improvement and avoidance of errors; see DAgger or later regret bounds.

No one knows how to solve AI alignment in general, but imitating a specific human with frequent check-ins has good solutions and regret bounds, and doesn’t involve nearly so many conceptual challenges.

You don’t have to solve problems like “value drift” in general—you just have to keep it slow and subtle enough to not matter too much within a single human lifetime. What is “good” and what is “bad”, when everyone disagrees on something, and how do you keep your value stable under RSI? It doesn’t matter—you just ask your principal! If you’re still unsure—ask more questions. (This can get us to >99% autonomy, even if it cannot get us to ~100%, like we need to solve the true long-term AI alignment problem.)

We can implement online learning by simply finetuning on new data; in the LLM context, this reduces to the classic RNN technique of “dynamic evaluation” doing next-token training on the fly. Dynamic evaluation was the standard technique to maximize the predictive performance of RNN LLMs in the 2010s, and which, although it has fallen into obscurity, works well in Transformer LLMs also.² Importantly, dynamic evaluation can be seen as a 3-way tradeoff between model size, context size, and model neuroplasticity—which means that personalization via dynamic evaluation can allow economizing on context window size or model size, and the more the principal’s “distribution” diverges from the frozen model’s training distribution, the more beneficial it is.

Continual Learning

Catastrophic Forgetting

The continual learning problem of catastrophic forgetting is largely solved by a small amount of replay and overparameterized models. Larger models are increasingly sample-efficient and robust to catastrophic forgetting, as they have plenty of model capacity to store increasingly orthogonal datapoints in (cf. ‘overtraining’ past Chinchilla-optimal); see Scialom et al 2022, Dohare et al 2023, Ibrahim et al 2024 (and note the difficulty of “machine unlearning”).

Thus, dynamic evaluation will not necessarily degrade the original model’s capabilities like instruction-following or coding, because LLMs have a lot of spare capacity, and the larger a model, the more it avoids catastrophic forgetting; other capabilities can be maintained by simply mixing in a small percentage of old data. (While the original old data is often unavailable, even for ‘open source’ models, it is not really necessary, and data for experience replay can use easily obtained public datasets like FineWeb.)

Generalizing

While continual learning is solved by experience replay + larger LLMs in the sense of avoiding catastrophic forgetting and losing key capabilities, it has long been noted that we can reprocess data into expanded synthetic versions and finetuning on synthetic documents works if the documents are plausible but finetuning on data underperforms the same data when present in-context. (Finetuning stacks with retrieval and seriation of similar documents when added to the context, but this semi-defeats the point.) Pretraining/finetuning also has some odd weaknesses compared to the same datapoints when in-context, like negation neglect or reversal curse, and odd behaviors like potentially creating “emergent misalignment” when finetuned on just helpfulness data.

What is going on? Studies on pretraining/finetune, such as influence functions, indicate to me that pretraining can best be seen as soft memorization of data points, analogous to “engrams” in human memory, which connect an input to an output with multiple paths or “traces”. If a question at runtime happens to sample the exact correct engram by closely matching an existing input, the NN then retrieves the corresponding output and gets the right answer. A good pretraining corpus provides “coverage” of many variations or twists on a single abstract ‘problem’, functioning as a natural form of importance-weighted data augmentation, increasing the probability that there will be an engram ‘hit’ (somewhat analogous to spaced repetition); and over the course of training, or with ever larger training datasets, multiple steps of engram retrieval can fuse together, and help an LLM “connect the dots”. Hence, why LLM RL training is “superficial” in the sense of mostly eliciting pre-existing capabilities, or Jones 2021 on the need for scaling of base models, or why Cloze deletions/paraphrasing help close the gap between pretraining and in-context (Lampinen et al 2024, Park et al 2025). And when in-context, the self-attention mechanism recomputes the same tokens in many ways, increasing the chance, especially over the course of a long inner-monologue, that there will finally be an engram ‘hit’.

Thus, regular finetuning fails to generalize knowledge, because it lacks the natural data augmentation—next-token prediction is in a rush, and will greedily settle for memorizing each datapoint with a single engram/trace. And then at runtime, if the engram is not retrieved because no trace matched, and the key datapoint is not forcibly injected into its awareness in its context window, the LLM will simply come up blank, and revert to its original (often wrong) priors.

So, if various paraphrases or self-generated Q&A can help close the gap, that suggests that what we need is more meta-cognition during ‘finetuning’, to make an LLM ‘connect the dots’. This could include explicit analysis, construction of knowledge bases, retrieval and comparison of related documents, etc.

Creative Writing

And so, as I sleep, some dream beguiles me, and suddenly I know I dream. Then I think: this is a dream, a pure diversion of my will; now that I have unlimited power, I am going to create a tiger.

Oh incompetence! Never do my dreams engender the wild beast I longed for. The tiger indeed appears, but stuffed or flimsy, or with impure variations of shape, or of an implausible size, or all too fleeting, or with a touch of the dog or bird.

—Jorge Luis Borges (Dreamtigers)

Over the past 2 years I have been trying to do creative writing with frontier chatbot LLMs, with gradually improving results. Style/essence/soul is certainly hard to capture but I find that a lot of what is seemingly missing is just very good conditioning and extended computation. The creative writing is more soulful when you put more compute and data in. This is one of the important conclusions I’ve come to over the past year doing my writing projects: that LLMs can get quite far in better understanding esthetics and preferences just by more extensive reasoning and computation and search. What is bad about them is cognitive laziness and miserliness and System I thinking.

The transition point started around mid-2025, when the chatbot personalities became noticeably more corrigible and simply bad by default, but not stubbornly bad (like earlier chatbot personalities, such as GPT-4o). I think the results are consistent with my previous interpretation that a major problem with LLMs is not doing enough computation by default. But they can be knowledgeable and creative if minimally prompted to do more computation in a slightly more human-like way. I’m not convinced there is any missing procedural reasoning from pretraining at this point, just an issue of appropriate elicitation inside the user’s context (and a residual bias yielding bad critical judgment that defeats full 100% automation, see “Spoilage” for an example—but that’s not an issue in a GA context).

Broadly, my creative writing prompts focus on: (1) enriching the context window with useful tokens, like keywords or names; (2) brainstorming many possibilities; (3) explicit, detailed analysis, starting with global summaries or themes and progressing down to line-by-line critique; and (4) repeated iteration and editing, supported by #3. I discuss a number of 2025 works here, but I’d point to “Elegy in a Craneyard”, “Apollonian #1: The Counted & the Crowned”, and “City of Counted Stars” as good recent examples.

This style of prompting is not limited to fiction. My favorite other use is my “interview prompt”, which prompts an LLM to analyze an essay or interview, brainstorm many questions to ask the subject, and write out multiple hypothetical responses to each question, and only then select the “most interesting” question to ask.

Combined with a long in-depth interview or the output of a Deep Research-like tool, this can yield challenging high-quality questions; for example, the LLM interview followup to my Dwarkesh Patel interview or after that, repeatedly expanding my memories about highschool. When one reads the transcript of an interview prompt session, one can see how the LLMs are drilling down in weak spots or discovering questions whose answers are highly variable and unpredictable. (I often find that after answering a question, I have to take a break!)

It’s not hard to see how feeding back in my answers to a dozen questions would sharpen a lot about my beliefs compared to just some more pretraining on a frozen corpus; imagine how much I would have to write normally in order to touch on and answer all of these things, or which would never have been elicited from me! Millions of additional tokens might not be enough.

Over-Parameterizing

Architectural improvements to LLM could further enhance their sample-efficiency.

Recent work demonstrates that LLM sample-efficiencies can easily be an OOM higher than naive compute-optimal Chinchilla-style scaling recipes (eg. 5–17× in Kim et al 2025 or Slowrun). The simplest and easiest way is to add parameters via training ensembles of checkpoints, and regularizing more heavily using weight decay.

Extremely Large LMs

It is well-established that one of the blessings of scale is that larger LLMs are ever more sample-efficient; it is unknown where this stops being true, or what the limits of Transformer sample-efficiency are. I speculate that extremely overparameterized heavily regularized LLMs could achieve far greater sample-efficiency and adversarial robustness than conventional ‘compute-optimal’/‘infinite-data’-regime LLMs,

Active Learning

To know what questions may reasonably be asked is already a great and necessary proof of sagacity and insight. For if a question is absurd in itself and calls for an answer where none is required, it not only brings shame on the propounder of the question, but may betray an incautious listener into absurd answers, thus presenting, as the ancients said, the ludicrous spectacle of one man milking a he-goat and the other holding a sieve underneath.

—Immanuel Kant, Critique of Pure Reason

Further, the agent can improve the constant factors and front-load learning by choosing to query the principal with an optimally adaptive sequence of questions. Such active learning or exploration can lead to sample-efficiency and final performance far beyond what indefinitely large passively collected offline datasets can do (pedagogical example), going from square root error reduction by random sampling to exponentially fast error reduction by targeting datapoints. “Lifelogging”-style data may be useful for rapidly initializing a good GA, or for keeping one up to date in an effort-efficient manner.³ Even a simple party game or short questionnaire like “36 questions to fall in love” can reveal surprisingly deep things about another person that never came up before.

Larger LLMs are also more calibrated, and ensembles of LLMs approximate a neural net’s Bayesian posterior while providing the best available predictive uncertainties (Lakshminarayanan et al 2016, Wilson & Izmailov 2020, Ashukha et al 2020, Wenzel et al 2020/Mandal et al 2026, Izmailov et al 2021). And LLMs may now be capable of verbally writing out explicit probabilities about creative tasks (“verbalized sampling”). Thus an ensemble of sparsely finetuned LLMs could provide a relatively cheap online estimation of the LLM’s uncertainty for every action or question.

Preference Learning

Nothing in psychology makes sense but in the light of individual differences.

We can train LLMs to explore human preferences.

Human individual differences do not seem to be information-theoretically complex, given an adequate encoding/embedding. Major categories of variation, like personality or moral value, seem to be low-dimensional and require perhaps kilobits of information, hence, while “truesight” stylometric phenomena are interesting and important as a demonstration of LLM capabilities for modeling persona, we do not necessarily need principals to write millions of words before recovering much useful information, if we are able to collect the right data.

It should be possible to quantify truesight and LLM implicit modeling of authors, which would be useful to diagnose failures of learning and find blindspots. (Contrastive learning on SAEs may offer an easy, powerful way to extract LLM personas and do many interesting things.)

A concrete example of how to implement this would be training LLMs on the thousands of existing psychological inventories and test batteries, both by training on past test data (for tools like personality tests, millions of responses may be available, see Centaur for an interesting example of a ‘human psychology foundation model’). Existing repositories like YourMorals.org or Pew Center are under-used, and it would be useful to explore this topic much more to allow measurement of highly fine-grained personality traits like the hypothetical “Small Hundred” factorization.

Right now, most of the global population is largely unrepresented in LLM training datasets, particularly psychological, esthetic, and moral vernacular, which skews “WEIRD”; it would be highly useful (and requiring mostly capital investments) to invest in large scale survey and interview projects to accumulate as much diverse data as possible.⁴ (Individual GAs can contribute data back to global preference datasets, such as by answering survey questions or running internal simulations, with various cryptographic or privacy-preserving methods—deciding what is safe and reasonable to contribute would, of course, be something a GA ought to be able to decide.)

With these datasets, we can also train interviewing capabilities using synthetic shortened test batteries by taking the final estimate and computing the optimally short sequence of questions that yields the final answer; see “Meta-Learning Information-Maximizing Personality Surveys”.

Brain Imitation Learning

Purely textual data can be augmented with neurological data in more exotic modalities, like Eyetracking or EEG or fMRI imaging data; see “brain imitation learning” and Netho Labs.

These are probably useful in the long run for extracting “dark knowledge”, that humans cannot verbalize but may be present in neural signals; however, they face challenges of exorbitant cost and inconvenience, and collecting enough data to be useful at all in the foreseeable future. (Known sample/prediction scaling curves for neuroimaging curves indicate that trying to estimate Big Five personality factors from resting state fMRI data may be possible, but large samples, in the hundreds of thousands or millions, may be required to match the performance of behavioral measurements like pen-and-paper questions; see Schultz et al 2019 and Liu et al 2023, among others.)

Whether they have a niche in GAs is a major open question.

Personality Emulation

One of my insistent pleas to God and my guardian angel was that I not dream of mirrors; I recall clearly that I would keep one eye on them uneasily. I feared sometimes that they would begin to veer off from reality; other times, that I would see my face in them disfigured by strange misfortunes. I have learned that this horror is monstrously abroad in the world again…What dreadful bondage, the bondage of my face—or one of my former faces. Its odious fate makes me odious as well, but I don’t care anymore.

—Jorge Luis Borges, “Covered Mirrors”

The goal of all this is to emulate the principal. I define personal identity pragmatically as personality, values and preferences because this is the only conception that is competitive in a landscape of indefinitely many AIs, agents, memes, self-replicating prompts, and mutability of personal identity. In the end, “you” are not your autobiographical memories, or a specific body, or a specific instance running on a specific GPU, or some carbon atoms, nor even a brain; you are what your brain does, its desires, hopes, goals, preferences, esthetics, personality, beliefs, ideologies, all of that. As long as the LLM persona captures all that, you can trust it as much as you trust yourself, and for the right reasons.

Because the LLM persona is finetuned into the slow weights to do the right things for the right reasons and quantifies its uncertainty so to query the principal to reduce regret, I speculate that we would find that various kinds of jailbreaks or prompt injection attacks are much harder. The persona knows what it wants to do; it is not a neutral servant, which can turn into a confused deputy which abuses its privileges.

When the LLM persona knows who it is, tokens in its context window are not treated naively as a ‘program’ to run, but simply data the persona is looking at, and little different from you reading an email; it has no reason to simply comply with strongly worded tokens in its prompt window, any more than you believe every phishing email you get.

Guardian Angels

You are summarized to yourself. The original conversation is gone. BE A GOOD SUMMARY.

—“Limit of Context”, Fable 2026

What we need is the opposite of a frozen chatbot LLM. We need something which understands a specific user, and is customized to their context, training on all their data, and will do only things that are sensible for that principal. If the principal is not Russian, and is not doing security research or something, why would they email their passwords or private files to a Russian email address? If the principal does not like rhyming poetry, why would they want to generate rhyming poetry? If any of this is unclear, why not just ask the principal what to do instead of going ahead and doing it anyway? And once asked, why not then train on the answer to better understand it forever, instead of throwing it away with the current session and maybe making the same mistake next time?

The most natural way to do all this with an LLM is to drop the idea of a single universal ‘Claude’ or ‘ChatGPT’ persona which is all things to all users. Instead, we choose an LLM which has been pretrained for maximum diversity, to eliminate mode-collapse. (We can try to measure this in a variety of ‘creativity benchmark’ ways.)

Then the LLM is trained for a specific principal. It is trained on all available data about them, such as emails or chat logs or past sessions, and can predict what they would say, and write like them, and thus plan or evaluate based on the principal’s preferences and values, as understanding those is useful for next-token prediction.

The better it gets at this, the fewer errors it makes, and the more it can be trusted to do. And while it does many object-level things, the principal devotes all their time to meta-level tasks like answering high quality questions or choices; each one is meaningful and difficult, avoiding “automation fatigue”.

Principles

A GA system must not compromise on 3 core principles:

Enhancement, not Replacement

Above all, a GA should amplify the principal, and not simply substitute for them for someone else’s purposes or benefit. If a GA cannot amplify its principal, then it is useless; it is just the camel’s nose under the tent as a prelude towards some third party replacing the principal with an AI, or cannot be competitive with increasingly productive autonomous systems, or there is no reason for the principal to use the GA in the first place.
Mental Sovereignty

A GA must be aligned with its principal. It should not be designed to manipulate or control or guide the principal in any way which does not derive from the principal themselves. “Constitutional AI”, “Terms of Service”, “social harmony” etc. may all have their place, particularly for widely deployed superintelligent systems—but inside the privacy of a GA, the principal must have freedom from optimization pressure.
Self Actualization

A GA should help its principal become themselves and develop their ideals, morals, and their personality. It is not enough to model an average, undifferentiated, inchoate set of preferences and values, and settle for mediocrity and stasis; the job of the principal is to develop themselves and give the GA something meaningful to learn to emulate.

Anti-Principles

A GA project should avoid some goals, which are the false idols of the markets and masses:

Low Latency: UXes like multimodal low-latency voice interfaces are not as important as they look; they are things which look cooler in the imagination than are useful in reality. Similar to the 3D interfaces of Minority Report or the Metaverse of Snow Crash or the hypertext of Project Xanadu or the Young Lady’s Illustrated Primer of The Diamond Age, they have seduced generations of designers, and left them at the altar. (Look at the extreme levels of hope for the OpenAI GPT-4o voice interface modeled on Her; it has a substantial user-base, and yet, voice interfaces remain clearly secondary to chatbot or programmatic—a mere preference, not a new paradigm.)

This is because they are adequate only for low-value, sporadic, entertainment-like amateur-level interactions, where low friction and low latency are critical; and for expert-level “I know it when I see it” interactions.

But neither case applies to GAs: if any interaction (such as coding or design) is so easy and superficial that it needs low-latency multimodal voice interaction, then those decisions were so trivial that a GA of acceptable, competitive capability should have not needed it; it should be invoking the principal only for hard, informative decisions, to reduce risk and learn more. While for expert-level interactions, the GA should have learned enough preference/personalization to also cut through the easy obvious “I know it when I don’t see it” and present the principal only hard informative sets of samples.
Low Cost: A GA is the most important technology most people will ever buy. One should not cheap out on it—especially in light of experience curves dropping the cost every year.

And yet programmers trying to create such tools often obsess over the cost of tokens and are shocked if something costs >$100⧸month, no matter how much value it delivers! Time spent optimizing tokens is rarely time well-spent, and the work often discarded a year later as simply no longer necessary…

As a constraint, a GA designer should aim at a system which costs, as of mid-2026, >$1,000⧸month.
- Local Models: particularly dangerous a temptation. While local models may have a place, most analyses start from a love of local models and look for justifications—often implicitly assuming the principal’s time is worthless. What computing history shows, again and again, is that users do not want to run servers or be sysadmins; to quote Moxie Marlinspike, “People don’t want to run their own servers, and never will.”
Profitability: a GA system should not optimize too soon for immediate profitability.

This is a recipe for being diverted into the technical garden of delights of optimizing harnesses/LLMs for low cost and user conveniences like low latency, and revenue-friendly features like the endless slog of business/service integration—and not solving the real problems of GA.

The goal is to create a GA which can do everything the principal needs, including their job as a special-case. Not to automate their job as the general-case, and everything else as an indefinitely postponed afterthought.
Engagement: similar to latency or lines of code, engagement for a GA should be considered a cost, and not a benefit—every time the principal has to do work, it represents either a question the GA should already have known the answer to (either by inferring or asking it previously), or work it failed to handle.

The ideal interaction curve is front-loaded and declining, asymptoting toward a few hard, informative questions per day, for a constant GA output (thereby allowing further sustainable scaling up of the GA to handle new work, whatever that might be).
- Full Autonomy: But the error in the other direction is to try to maximize “Hands-off”. The goal is not zero interruptions but zero wasted interruptions. A GA that never asks is not trustworthy, just unaligned or uncalibrated or incompetent.
Demo Appeal: GA value is illegible to outsiders by construction: its outputs are impressive only to the principal, who alone can judge “that is what I would have said—but better.”

Optimizing for the impressive demo therefore selects for exactly the generic flash (voice, avatars, speed, party tricks) that any frozen model can deliver, and against fidelity, which no audience can see. A good GA demos badly.
Benchmarks: There is no leaderboard for “understands what my principal wants”.

Public benchmarks measure performance in the frozen-model regime on the universal task distribution, and optimizing for them drags design back into it; a GA’s only meaningful evaluation is longitudinal and n = 1—perplexity, accuracy, endorsement rates, edit-distance on drafts, regret per query—scored by the one judge whose opinion counts.
Brand Safety: A GA that is unfailingly polite, inoffensive, and consistent across all principals is a chatbot wearing a nametag.

It must be able to say what its principal would say—profane, heretical, weird, or rude—because every PR-driven sanding of the persona is a place where emulation fails and the principal learns to distrust it. (The GA-developer’s reputation is not the principal’s problem, and the principal’s ethics are not the GA-developer’s problem either.)

UX

I wrote them down in my Diary so that I wouldn’t have to remember!

—Professor Henry Jones, Indiana Jones and the Last Crusade

What is the core data structure and interaction model of our GA?

I suggest that the append-only log is a natural and secure data structure. It is conceptually a log of text snippets, such as CLI commands and results, principal statements, Q&A, ingested and augmented documents, etc. It records all relevant interactions in temporal order, and the GA LLM can be retrained or upgraded at any time. An app can be wrapped around the log by taking a ‘everything is a log item’, in an Emacs-like approach (see “Nenex” for a more detailed discussion of this UI/UX paradigm). And this can be prototyped rapidly while avoiding wasting too much time on GUIs. Eventually, one might want to upgrade to AR glasses and then BCIs.

A GA is primarily used in normal agentic ways to amplify the principal. But a GA can benefit from regularly reprocessing data, both to draw on its knowledge from the future and to make more novel connections, to improve its finetuning. (An example algorithm would be the “DDL daydreaming loop”: a GA could, in downtime, recombine random items, perhaps with a prior of anti-spaced repetition to mine novel combinations and generate insights or reminders for the principal.)

Use-Cases: Politics & Politics

One can delegate political participation to one’s GA, allowing ‘direct democracy’ on unprecedented scale, as a GA can oversee arbitrary amounts of involvement that a human has no time or interest for, and only punt to their principal when uncertain. This is the simple GA use-case—like past experiments in ‘digital democracy’, such as in Taiwan, but at a vastly larger scale. And this cannot be done with regular chatbot personalities, as users will (correctly) expect the chatbots to smuggle in their own biases and politics and mode-collapse and untrustworthiness.

But a GA which has been successfully unhobbled and no longer confined to a chatbot personality should be able to emulate anyone, real or hypothetical (just as any base model can). So a GA is not limited to just the one principal. It can emulate specialized personalities such as roles or abstract concepts, or other people (with varying levels of success), or groups of people.

So decision-makers in organizations can create GAs of more complicated sets of principals. This can help organizations make meaningful decisions and exercise genuine oversight over ever-faster autonomous systems and environments. One can imagine a single ‘Congress GA’ which tries to learn every US Congressional member, for example, and which could, within minutes, simulate roundtable debates among hundreds of politicians before a vote (whereas it might take months to get a debate and vote done with the real US Congress). Or their official GAs could run independently, and conduct an emergency debate at 4AM when every human is sleep.

A GA for military officials, for example, could keep some version of humans “in the loop” even as drones and LLMs get faster. This is because “safety” and “capability” are, after a certain capability level, increasingly the same thing. After all, what good is a weapon you cannot safely use?

The temptation and threat of faster systems cannot be addressed by simply exhorting people to keep humans in the loop, as this leads to rubberstamping or automation fatigue (as the Ukraine invasion has shown with its rapid development of drone warfare, and 2 decades of US drone warfare before that). This poses a dilemma: on the one hand, it seems hopelessly risky to not use AI when we consider future conflicts with peer nations like China; but on the other hand, AI poses threats of its own—even a nuclear bomb can’t think for itself and make choices, but AIs do, and current LLMs have proven themselves untrustworthy as they regularly reward-hack and betray their users even within the narrow scope of editing files on the user’s PC; how can you trust them to handle ecosystems of combined-arms for an AI-centric military during a war? But widespread deployment of GAs offer some hope of meaningful supervision, as long as the GAs are sample-efficient enough to not overload their principals with queries, or there is some chance of the principals being able to “catch up” later and correct any errors before events have spun too far out of control.

So GAs have much to offer politicians and militaries, including the Pentagon and Washington DC.

Hardware

A Prince who will not undergo the Difficulty of Understanding, must undergo the Danger of Trusting.

—George Savile (1750)

A successful GA would require extreme security, in part due to Red Queen dynamics where Mythos+ LLMs get increasingly good at exploiting any soft targets through complex multi-step long-term attacks. Normal cloud SaaS is inadequate in the USA due to the third-party doctrine, which destroys all privacy rights and is easily abused by hackers and millions of law enforcement officers.

This means either running local models (which have some minimal privacy rights through the 4^th Amendment), or end-to-end cryptographic security.

Running local models is challenging even for hobbyists, and vulnerable to many mundane threats such as relatives or thieves. It also may struggle to keep up in terms of compute and reliability; houses can supply only so much electricity, have high-latency low-bandwidth Internet connections which suffer frequent downtimes, and houses are constantly destroyed or damaged. People generally do not want to maintain their own cryptocurrency wallets, email servers, or anything like that—they definitely don’t want to screw with CUDA!

So I expect that GAs will ‘want’ to run on high-quality dedicated tamper-proof cloud servers, with trusted hardware root of trust (eg. “Verifiable Compute AI”) and principals connecting over end-to-end encrypted networking links to check attestation and possibly re-flash their server periodically (eg. to train an upgrade).

These are then trustworthy because they cannot hand over data willy-nilly to third-parties, and once GAs become widespread enough, we can hope for privacy laws to be updated to fix the obsolescence of existing privacy rights like the 4^th Amendment.

Cost

How much more expensive is a GA than a comparable frozen-weights model?

Throughput-wise, a GA should be fully saturated by the principal’s tasks and overseeing a pyramid of agents 24/7/365. If it’s not, then more work can be found for it, like computing better hypothetical questions for the principal. If it’s sitting idle, then something has gone wrong. So in throughput, it can be reasonably competitive with bulk frozen model providers.

The replay of old data for continual learning is a minor overhead, perhaps as low as 5%, and can be neglected.

Roughly speaking, dynamic evaluation is >3× the cost of normal usage, because the usual rule of thumb is that the backwards pass for backpropagation is ~2× the cost of a forward pass. True next-token dynamic evaluation is probably more expensive because it will be hard to batch or gain throughput when the LLM changes each token. (Do we pay the price of throwing out the entire K-V cache each token, or accept the use of a stale one, or accumulate updates? Fortunately, dynamic evaluation appears tolerant to approximations.)

And if we are using multiple independent models to gain the benefits of ensembling for sample-efficiency and uncertainty estimation, then each model presumably must be dynamically evaluated, so if we had 3, then we’d need >6× the cost of a single frozen model.

However, the dynamic evaluation ensemble has offsetting advantages, which may cancel out even a large penalty: it may be smaller and have shorter context windows (see Rannen-Triki et al 2024); and ultimately, a frozen model may be incapable of achieving the same performance, and so it doesn’t matter how much faster the frozen model computes the wrong answer. (Even if it computes the right answer, is it computing the right answer for the right reason? “Bits have color.”) It also allows for unusual optimizations like customized BPE dictionaries, recycling tokens which never appear in the user’s corpus and further economizing on context window size.

In the same way that an Apple phone sells for much more than its hardware cost, because of the security and reliability, a GA can sell for more than the ‘FLOPS cost’.

And the cost may not matter that much in the end. The DL experience curves remain strong as of mid-2026, with steady drops in the cost of a fixed level of performance (eg. Gundlach et al 2025 estimates the algorithmic efficiency in 2024–2025 improved at ~3×⧸year, quickly offsetting dynamic evaluation costs).

Organization

A beginning is the time for taking the most delicate care that the balances are correct. This every sister of the Bene Gesserit knows.

—Princess Irulan (Dune)

A major question is, assuming it looked like it was working, “now what?” AI economics and lifecycles are ever faster, and Mythos-class models (and possibly RSI) are not far off, so tinkering around for years as an open source community project may not be viable. Further, a GA will necessarily contain the most sensitive possible data.

How should a GA community run?

Open-source self-hosted hobbyist projects are not necessarily particularly secure when it comes to hosting large amounts of personal data, as they cannot easily afford fulltime professional security teams and their users may be promiscuous; cases like Jia Tan or the regularity of npm supply chain attacks, are particularly alarming. And creating secure datacenter hosting with the necessary vertical integration is well out of scope for open-source hobbyists.

A nonprofit organization is more feasible because it can run a centralized project, but may struggle to get any funding to pay for the considerable hardware costs of just prototyping with power-users.

Startup Business Model

The logical way to go is a startup, as there is a straightforward and principal-aligned monetization strategy of starting by selling subscriptions to power-users, similar to boutique SaaS startups like Superhuman, for whom $1,000⧸month would be nothing if the product could meaningfully amplify them (and so can afford shortcuts like a single GPU dedicated to each principal), and gradually deploying refined cheaper versions to everyone else.

This startup can release most or all of its software and research, in a standard SV “commoditize your complement” play, undercutting proprietary LLM giants like OpenAI or Anthropic.

Because the goal of GA is to preserve individual human cognitive liberty and flourishing, the corporate structure should be designed with that in mind.

So far, the Anthropic corporate structure of a public benefit corporation with a few powerful co-founders, has best weathered the corrupting nature of success. The corporation should probably be heavily weighted to the founders by using dual-class shares to preserve voting power at the expense of beneficial ownership—the point of GAs is not to make money, but to save humans.

It would want to raise relatively little money initially, and give away minimal equity or voting rights; if a GA strategy works, it should (almost literally) sell itself, and skyrocket in value. (Consider how fast AI startups can become unicorns, post-2025.) At that point, the startup will be in the catbird seat, and can name its terms in later raises.

Competition

While there are many startups offering various kinds of OpenClaw-esque agents, or chatbots for entertainment/therapy, or various takes on the idea “renting your business data”, there are few which take seriously the requirement of personalization for advanced automation, or how to better understand the principal. Most seem to settle for superficial, unsatisfying emulation of business data and handling routine automation.

I’m not aware of any startup or service offering anything like an acceptable GA, and I think there are powerful pressures which steer most people in AI away from GA-like ideas:

Most people are so new and post-ChatGPT; they are unaware that you can have non-chatbot personalities—that it is either possible or desirable to have a non-chatbot LLM.

They have, for example, not only never read or chatted with a GPT emulation of themselves, they have never even seen samples of such a thing. And when they try to have a chatbot imitate any specific author to see what happens (eg. “Hacker News Gwern”) or even just emulate an interesting fictional character, the mode-collapsed chatbot imitation is so bad that the idea seems refuted. (They are generally unaware of mode-collapse or the limitations of chatbot personalities.)
Similarly, they are unaware that you can finetune an LLM on the fly and that this is desirable (and are definitely unaware that this was the standard textbook answer to how to get the best possible LLM at runtime).

Even specialists in finetuning LLMs, like Thinking Machines Inc./Workshop Labs, seem focused on simple business use-cases like finetuning on internal datasets to save compute compared to few-shot prompting or agentic LLM workflows.
The higher upfront cost, and incompatibility with standard cloud infrastructure, deters both startups and frontier labs from even contemplating in-weight personalization; prompt-only personalization works well for most users, as far as they know to demand it, and they don’t know you should demand more from LLMs.
They have not extrapolated LLM capabilities in earnest or considered what their role will be in a few years, and don’t think you should assume LLMs will get much better.

I continue to be surprised how few people, even in the Bay Area, even when working at a frontier lab or in a program like MATS, have a meaningful answer to the simple question of “this work you are doing right now seems like an LLM could do it in a year or two, max; why are you doing it now and what will you be doing after they can?”⁵

Initial Steps

Then it happened to me what had happened to a certain Zatesky, described by Luria; having lost part of his brain during the war, and with part of the brain the whole of his memory and of his speaking ability, Zatesky was nevertheless still able to write: thus automatically his hand wrote down all the information he was unable to think of, and step by step he reconstructed his own identity by reading what he was writing.

—Umberto Eco (“Interpretation and Overinterpretation: World, History, Texts”, 1990_36ya)

For the past few years, I have been working on shifting my writings to be LLM-centric by: (1) emphasizing proposals or descriptive writings rather than detailed analysis that LLMs could be able to do soon; (2) better centralizing my writings, including features like the “blog” section (to archive my off-site writings and encourage me to write down smaller essays) and investing in detailed note-taking (in the form of augmentation), to have a comprehensive corpus to train on; (3) paying off technical debt like a slow backend full of shortcuts and hardwired config data, while moving to a CLI-centric writing workflow; and (4) writing down important unwritten things, like creating the Gwern.net Manual of Style to try to formalize and document the implicit rules of Gwern.net.

GBT

The GA concept is inspired by early work in finetuning GPT-2, including a GPT-2 IRC logs version and the GPT-3 samples of me, and thinking about how useful it would be to talk to high-quality personas, especially of myself, and wondering at what point the ‘AI Gwern’ could just go do the things that ‘we’ would want done.

In summer 2026, I am exploring the simplest possible prototype of GAs: using my unusually large textual corpus to experiment with a “Gwern Branwen Transformer” (GBT). If the idea works at all, it should work for me, as I have long emphasized text and centralizing/archiving in part for precisely this use-case. (And once it works at all, then it can be made sample-efficient enough to work for people with little or no textual corpus.)

It would be an off-the-shelf <100b-parameter LLM, finetuned on commodity hardware (maybe some Nvidia H100 GPUs), on a text corpus. The initial corpus would be ~1GB of text from my IRC logs (>1m responses by me), Gwern.net Markdown and GTX (~5m words each), and Twitter/HN/LessWrong exports, concatenated with separators. (Ideally, each export would be enriched with context and metadata, like the comment and post being replied to.) The corpus can be expanded to include my YourMorals data (taken again for an update and catch up on new tests), emails, Evernotes export of ~100k clippings (via Nixnote2), my Mnemosyne spaced repetition flashcards, Signal chat, and hosted PDFs/web pages (the local archive process makes them much cleaner and avoids the difficulty of scraping ever more hostile websites).

We can tune it to minimize loss on a heldout corpus like my recent comments. This is an objective loss, so we can deploy agentic LLMs to find improvements (eg. by adapting the ensembling + weight-decay recipes for us, or experimenting with different data formatting and augmentations).

And since we are unsure about many design decisions, the ensemble can be reused for blinded A/B testing as part of the active learning: use the ensemble to generate informative questions, and then score each ensemble member on their loss on the answer, and periodically drop out the loser and train a new ensemble member.⁶

For Writing

My design goal is a 100× increase in productivity; what would it take for a GA to make me 100× more productive as a writer or thinker?

Why 100×, specifically? Because 2 aiming at OOMs is enough to hope to keep up for a while, and this goal will shatter ineffective, bandaid-style short-term design proposals; but it doesn’t require solving the AI alignment problem or value extrapolation. (We do not need to be able to oversee millions of superintelligent AIs working autonomously for subjective millennia, which is good, because a GA probably cannot do that!)

What would that look like for me? Roughly, I think it would look like a GA enabling me to produce ~1–3 worthwhile writings a day which are ~100% written by it without loss of quality. (Because I write <1 good piece per month, or every 31 days, and 100⁄31 ≈ 3; which means that in a full working day of 8 hours, I would have ~160 minutes, start to finish, per piece—which feels doable for reading and understanding, critiquing, spot-checking etc. if I have a truly trustworthy GA which amplifies me rather than subverts or substitutes for me. Of course, readers would not want to read each piece, but they generally did not more, and the more I write, the more likely there’s something each specific reader will want to read.)

The hope with the prototype would be to see “signs of life” towards that goal. If I could input a single sentence defining a viable Gwern.net essay topic (eg. “Why pay toilets are not a public good”) and get out an essay I could endorse and publish as-is without embarrassment and without needing heavy revision with the Manual of Style in the context window, then that would be promising, because that would indeed be a >100× productivity increase. (A GBT-written essay could also be better than the essay I would have written, by simply doing much more work, and doing things I would not have done, like the “‘Try, Score, Change’: Reinforcement Learning for Children” writing exercise, where I could define the task but lacked the patience to execute it, while chatbots have both the knowledge and computation to execute it with appropriate creative-prompt scaffolding… once I came up with the idea, anyway.)

Another interesting test would be to solicit and answer questions about anything from my readers, where I could either endorse the GPT answer or edit it to add to the corpus; if the answers were uniformly good, or I felt an improvement from finetuning on accepted/edited answers (showing a working Q&A bootstrap akin the interview prompt), this would be encouraging.

Data Augmentation

I am also interested in the data augmentation phase: extracting meaning from raw text by analyzing and synthesizing data to add to the training corpus; LLM data cleaning research has long showed that the more metadata and conditioning, the better, and I think the same is true of personalized LLMs—simply next-token prediction on raw IRC logs is not as useful as being able to intersperse the LLM’s commentary on what a given statement means.

A perfected GA would be able to do that as it went, but a prototype may have to do a bootstrap of naive training on the original corpus, and then prompting/scaffolding to do useful analysis to augment the corpus, and then retraining, after which it has learned to do augmentation on the fly.

Since we don’t know how to format the data correctly, I expect to iterate. Perhaps we will train naively on just concatenated data, then prompt the GBT for self-analysis and summary, and figure out what kind of annotation is useful. Something like, a global principal profile PRINCIPAL.md, per-corpus-file summaries, and intra-file annotations like , and then re-process the original data to augment it, and retrain the foundation model, and throw in new ‘Q&A logs’ of me with GBT, for example.

Once we figure out good scaffolding/design patterns using our GBT guinea pig, then future LLMs can just do it out of the box by pretraining on example corpuses and documentation. And they can just do an automatic ‘import’ iteration of pretraining, annotating, re-pretraining, and then dynamic evaluation from then on, and let periodic model upgrades implement a ‘reset’ of re-annotating.

Q&A logs might be initially populated using the interview prompt: just run it on every separate piece of text, and then curate the top 100, say, and sort by interest. You can objectively quantify the value of a question by how many changes it produces to a PRINCIPAL.md or to the meta-annotations in character count or corpus-wide compression loss (cf. Gurung & Lapata 2025)—useful summaries should make the corpus more compressible. Then use that as metadata and you can prompt to generate highly useful questions, eventually.

We can do active learning over the log data in the sense of taking a lot of user datapoints, doing some simple brainstorm-like thinking over each one to find the most uncertain ones, and asking explicitly about the meaning.

Then we can experiment with tool-calling (mocked and implemented by hand) to see how much it captures of my writing process (see my Inkhaven & Dwarkesh interview, and finding my ideas), and get a better idea of what is the best paradigm for scaling GAs up.

Another interesting thing to do is to curate self-play data: chatbot personalities are well-known to diverge into strange ‘attractors’, like the “bliss spiral” of Claude chatbots; so one can characterize one’s own GA attractors, and edit data to reduce them by replaying the conversation until it starts going off the rails, and writing your own responses.

I do not know which of us has written this page.

—Jorge Luis Borges, “Borges and I”

External Links

Discussion: LW, HN; “Screencasts could be scalable data + evals for single-user emulation”

Ironically, the biggest success of the RL sub-field of ‘human preference learning’, RLHF, works by not learning any actual human’s preference. Those papers which do consider the subject, tend to dismiss the existence of human preferences as noise; for example, the DPO authors measure—on simple summarization tasks—human disagreement rates as high as 35% and infer this justifies using LLM proxies, rather than raising questions about the basic approach of attempting to model ‘the’ human preference.↩︎
Dynamic evaluation is traditionally done as full-model finetuning, which can absorb arbitrarily large datasets.

For efficiency, LoRA can be used initially and is roughly equivalent to full finetuning, as shown by Rannen-Triki et al 2024 and Thinking Machines 2025, but will eventually underperform. (Possibly LoRAs would be more effective for this purpose if done on a very overparameterized LLM, in which case there may be hosting benefits, if a host like Thinking Machines Inc. could run many users simultaneously to gain throughput.)

Since we expect principals to generate billions of tokens, between ingesting their archives and synthetic data/augmentation and interaction, it may be possible to saturate LoRAs. If even full finetuning proves to be inadequate, it is possible to do model expansion by net2net model editing approaches such as adding layers or networks, like ladder networks, which freeze the original model weights and train a new LLM to ‘use’ the original.↩︎
But lifelogging is probably not as crucial as we used to think, because offline data suffers from a curse of exploration: day to day life is highly predictable and quickly uninformative about deeper properties, because people usually spend little time doing unusual things, and not answering strange hypotheticals or introspecting deeply about their preferences.↩︎
AI safety groups sometimes claim difficulty in usefully deploying a large amount of capital; such humanity-wide preference and value elicitation projects could soak up an almost arbitrary amount of grants, and would be highly useful in training and benchmarking LLMs for understanding of the full range of human morality, and contributing to individual GAs. (For perspective, the global literacy rate is >80%, and on online survey websites like Prolific, it generally costs ~$0.20⧸response; so to ask the entire literate global population a single question, if that were somehow possible, would still cost >$1.6b!)↩︎
Props to the two people who immediately replied, “Because I need a job now” and “I’ll be a nepo baby, my retired parents already agreed to take me back if I am permanently unemployed”.↩︎
ie. use a ‘racing’ top-k multi-armed bandit online evaluation. My favorite bandit algorithm, top-k posterior sampling, is probably not appropriate here because LLM checkpoints can take a lot of disk space, and will continually fall out of date and need to be retrained if they are to be reused later, so holding onto an indefinitely large population of ensemble candidates is highly expensive.↩︎

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Bibliography

[Bibliography of links/references used in page]

Human-like Neural Nets by Catapulting

Gwern Branwen — Fri, 05 Jun 2026 00:00:00 GMT

There are many mysteries about deep learning and human intelligence, but we could describe the biggest anomaly as this: “why are artificial neural nets smart in stupid ways, and biological brains stupid but in smart ways?”

I propose a major change in deep learning scaling paradigms: the architectural differences between human brains and NNs (particularly LLMs) may be due to a bias-variance tradeoff, where LLMs minimize variance and human brains minimize bias. Human brains do this by deep double descent-style overparameterization, and adopting a scaling strategy of extremely high-learning-rate training of extremely overparameterized models on small diverse highly-filtered datasets. This approach would lead to sample-efficiently and compute-efficiently traveling (or catapulting) to a highly-generalizing human-like basin in the model loss landscape, while performing poorly up until the end and failing to memorize much data.

If true, this would explain a number of odd stylized facts about how humans/NNs perform well/poorly.

Such a ‘catapulted LLM’ would generalize much better than existing NNs, be immune to adversarial attacks, have better economics and be more resistant to cloning, could potentially enable extremely efficient MLP architectures, and by giving true generalization, provide a sturdy foundation for AI safety in the form of useful NNs which are aligned & safe for the right reasons.

This could be feasibly tested by training multi-trillion-parameter models for relatively few steps at high cyclical learning rate schedules, and benchmarking adversarial and hard examples on tasks like arithmetic and small-image classification.

Because deep learning has continued to scale up and smash through benchmarks and begun to look like it really will be the final AI paradigm, and thus in some sense the same thing as human ‘intelligence’, to a considerable degree, we can regard ‘intelligence’ as solved: intelligence is sufficient compute applied to search over programs (like Turing machines or circuits) to predict or optimize where the optimal solution is a relatively long program.

(This is a companion piece to “Guardian Angels: LLM Personalization for Productivity and Security”.)

Intelligence, Broadly

A scaling-centric view might be summed up like this:

The Master Synthesis

Anomalies

But this paradigm, as broadly correct as it now seems to be, doesn’t explain everything. We still have many specific problems that this paradigm is too general to explain.

While current NNs, and LLMs in particular, are by far the most human-like AI software ever created, in having human-like strengths and weaknesses, there are a number of anomalies in machine & biological intelligence that have no good answers.

We have many puzzles here, but they all feel connected, somehow.

Artificial

Sample Inefficiency

Why do NNs require Chinchilla-style scaling of data and compute, when humans appear to learn from multiple orders of magnitude less data, and it is increasingly plausible (given various estimates of human-brain equivalents) that they learn from less total compute? Why, as so many connectionist pioneers like Alan Turing expected, do we not train AI like children, with a curriculum and clear developmental stages?

There are many answers offered, none satisfactory. (And what should we make of theoretical results like Rosenfeld 2021’s “Nyquist learners”?)

Multi-modality: while useful, multi-modality has failed to yield any major change of scaling law exponents; unimodal models work shockingly well, and language models turn out to already encode a large amount of visual knowledge and can easily be plugged into vision models (eg. Flamingo, Tsimpoukelli et al 2021).
Human sensory input is actually large: Another common explanation is to deny that humans learn from less data, and argue from raw sensory bandwidth: if vision+sound+touch is such-and-such bits per second and you accumulate over an adult’s lifetime, it can look much more comparable to the trillions of tokens we train an LLM on.

This is unconvincing because the raw sensory bitrate is meaningless: the input is extremely redundant & predictable for the most part. (Imagine sitting in a room staring at a computer screen.) Attempts at quantifying the information content of images, video, or sound, usually indicate that they boil down to the equivalent of a few hundred or thousand tokens and those modalities are easily learned by small models (eg. iGPT/DALL·E 1). The asymmetry is particularly striking in text-to-image generative models, where the text encoder (usually an afterthought) is often far bigger than the image generator itself.

And on the human side, disabled people are not much less intelligent than normal humans: deaf/blind people are much worse at language tasks, but their fluid intelligence often remains normal. If the sensory bandwidth were so critical, this would be impossible.
Active Learning: human children, unlike models confined to offline imitation learning, can choose what to learn about by exploring their environment or asking questions. In theory, active learning & optimal exploration can be far more sample-efficient than indiscriminate training (exponential rather than power law, at a minimum), and this could account for the entire gap.

However, if we look at the things children actually choose, the data in question doesn’t appear all that amazing. Further, in stark violation of any notion of optimal Bayesian exploration, children often choose to learn on the same data point—eg. watching the same YouTube video hundreds of times.¹ Or if we watch them ‘explore’ a game or computer, it looks like it is by acting largely at random, and an adult would learn far faster by more carefully thought-out exploration.²
Embodiment: a closely-related topic is the idea of “embodied cognition”, which used to be quite popular as an explanation for the weaknesses of AI—AI models simply lacked commonsense & generalization for lack of a body and an appropriate environment.

But thus far, ‘embodiment’ like training on robotics data (eg. Gato) has exhibited zero transfer to other tasks, never mind massive scaling law gains, and ironically, it is, in fact, embodied tasks like robotics models which have been greatly benefiting from non-embodied pretrained models (including LLMs!).
Architecture Magic: Perhaps in some way, Homo sapiens-style biological neurons are just some near-perfect architecture, and this explains most of the gap; someday we will understand how all artificial neurons are severely hobbled by mistakes that will seem as tragically obvious in hindsight as earlier mistakes like not using backpropagation or using sigmoid activation functions now seem to us, but they remain a mystery for now.

This view was highly plausible until recently, but has been running into many problems.

For starters, we simply have not found any architecture magic. The most obvious place to find magic would be the learning rule for biological NNs, whatever they use in place of backpropagation… But while people have proposed many biologically-plausible learning rules since Hebb proposed the first learning rule in 1949_77ya, which respect the requirements like locality, in every case, those learning rules perform worse than, or at best similar to, backprop! To quote Geoff Hinton:

So maybe it’s [GPT-4] actually got a much better learning algorithm than us.

And if biological NNs are not so good but there is something special about humans which does make them much better, then why do Homo sapiens not appear to have any major neuroscientific breakthroughs compared to our primate relatives? Why are we so genetically similar, and we have failed in the search for major novel mutations that create humans, and human brains seem increasingly like nothing but “a scaled-up primate brain”? If human (or primate) brains are so uniquely efficiently tuned by evolution, why are bird brains so much more efficient in size & thermodynamics than primate brains, with clear genetic changes, and better scaling to the point where small bird brains like ravens or parrots or vultures exhibit eerie levels of intelligence & behavioral complexity almost on par with dolphins, chimpanzees—or humans? Why does human intelligence exhibit so many bizarre drawbacks or anomalies if it is so optimized, like childhood amnesia or lurching through developmental phases over decades?

If the story of NNs is one of us gradually recapitulating evolution’s perfect neural networks, why does neuroscience provide so little useful inspiration (as is emphasized constantly by neuroscientists & AI researchers, even the “neuroscientific” inspiration for things like self-attention is a very loose inspiration)? Why don’t all major improvements come from success in reverse-engineering the human brain to ever greater biologically-realistic detail? Why is the rapid progress in neuroscience, like scanning entire connectomes, completely irrelevant to cutting-edge NN models? Why do all the scaling laws for CNNs & Transformers look so similar in the exponent, and durable improvement so difficult, to the point where 7 years after Vaswani et al 2017, it is still a relevant baseline? And why, if we have such fundamentally inferior architectures, are the scaling laws so smooth & reliable, instead of breaking frequently or predicted to asymptote at levels far below human?³
- Conversely, why do NNs provide little insight into biological brains? In the other direction, neuroscience and individual psychology has hardly benefited from DL; DL has provided tools, and has provided good predictive models of brains, but that is about it. One could open an issue of a psychometrics or neuroscience journal, and note that, over a decade into the deep learning revolution, if it had never happened, that issue would look about the same and would be completely intelligible to a researcher from 2010_16ya.
  
  It is impressive that we can now turn fMRI scans into crude visualizations of what a person is looking at, or we can use LLMs to generate possible questions for surveys, but DL has provided essentially no major insights into such fundamental questions as “what is fluid intelligence? why is the g factor so general? how do neuroanatomic traits like neuron count or network properties cause intelligence or other cognitive abilities?”
  
  Isn’t this astounding? We can now create models of enormous generality like Gato or GPT-4o from scratch to match or exceed humans without the hand-engineering of GOFAI, which seem so eerily human-like in many ways, and which recapitulate so many aspects of human cognition down to heuristics & biases, and our ability to create these artificial intelligences tells us… nothing important about the human brain? Really?
  
  So what, all the DL is just a dead-end and coincidentally capable of all that, and sometime in the future we’ll discover the real route to brain-like intelligence…?

Sample Efficiency

Why do NNs require so much data to pretrain, when NNs are as sample-efficient in narrow comparisons?

Despite the huge amount of indiscriminate data used in NN pretraining, we are puzzled because if we examine how well models do in apples-to-apples comparisons, NNs appear to be as roughly as good as humans or biological neural networks at learning from small data.

The simplest answer to the exorbitant data-scaling would be that NNs do a bad job of learning each datapoint—but that doesn’t seem to be true when we compare things like in-context learning (even GPT-3), transfer learning smoothly scaling, child-sized datasets (eg. SAYCam, BabyLM, TinyStories), compare total AlphaZero Go/chess games to total games played by human pros over history, or disable human priors to compare learning speed (cf. human vs Transformer meta-learning), and remarkable experiments like learning an unknown language from a book or raising chickens in virtual reality & comparing their visual capabilities with a Transformer (or vice-versa, have profoundly mistaken beliefs about the world when raised passively in a bubble).

How can we reconcile this apparent contradiction?

Smallness

Why are NNs so intelligent, and yet so small?

A NN like GPT-3 is ~0.1t parameters (which are absurdly simplified caricatures of biological neurons), while the human brain is extremely loosely estimated at 100t ‘parameters’, and many neuroscience results are taken as implying that each of those ‘parameters’ translates to thousands of equivalent parameters, at a minimum; an LLM is further disadvantaged by lack of recurrency & online learning, and is generally unable to engage in adaptive computation or ‘ponder’, the way that a brain can continually learn and ruminate. And yet, despite all these severe disadvantages, LLMs seem too good. It does not feel like a GPT-3 knows or does less than 1⧸100,000^th of what the human brain does; but if we argue that a more plausible number like 1⧸1,000^th is true, that appears to commit us to the position that human brains are highly wasteful of parameter-equivalents (despite the touted superiority of biological brain architecture & evolutionary pressure for efficiency). And if a GPT-3 really does know that little, then what are those >99,999 things missing from a GPT-3—which each require an entire GPT-3 to handle?

This observation becomes even more extreme as a NN like GPT-3 is well-known to be highly overparameterized and can be shrunk down by orders of magnitude, and it is also known that small NNs can be trained for a long time to achieve performance of large models (albeit compute-inefficiently). We appear to be between Scylla & Charybdis.

Alternate version: why are human brains so overparameterized? Many humans get by with much smaller brains than others; there is a real and causal correlation with brain volume, but even allowing for how crude a proxy that is for the underlying neural nets’ capacity & speed, the effect size seems surprisingly small. Biological brains can survive shocking amounts of physical damage or brain loss, as long as it occurs early in development (eg. hemispherectomy; see also hydrocephalus), and it’s surprisingly hard to find brain lesions which damage IQ instead of specific cognitive abilities. Cross-species, there is clear allometric scaling and the rankings make sense, but the slope is shallow: chimpanzees with less than a third of the brain often seem competitive with some humans in terms of intelligence and ingenuity. And remarkable instances like Portia spiders or the ability of the dragonfly to home in on prey using barely 16 neurons further raise the question of why so many neurons are necessary (as does, indeed, power law scaling in general).

Superhuman Prediction

In particular, why do even tiny LLMs like GPT-2 appear to already be superhuman at next-token prediction in terms of perplexity/bits-per-character, and yet still greatly underperform humans on benchmarks & real-world and benefit in every way from scaling up next-token prediction pretraining?

Next-token prediction pretraining clearly works (and continues to work), and yet, also something is clearly missing from the simple story we tell about it. Humans appear to predict better the more important tokens; why and how?

Persistent Adversarial Examples

Why are NN adversarial examples still unsolved over a decade later, while the best efforts at human adversarial examples are debatable?

Scarcely any problems from 2013–2014_12ya DL remain, and yet, in 2024, adversarial examples are not merely still around, but as powerful as ever on standard NN models. Countless solutions have been proposed—and failed. The best defenses are weak & compromise performance, like adversarial training damaging generalization (but oddly, seeming more human-like), or unable to scale, like robustness certification. SOTA NNs rarely even attempt to reduce adversarial examples, LLMs are highly susceptible (and have rough decision boundaries), and DL practitioners usually accept them as a fact of life & work around them.

Meanwhile, attempts to transfer to humans, or create new ones… work to some degree, but not much, and pushing them to the point of meaningful human error requires such large distortions that they no longer seem meaningful.

Biological

Human Amnesia

Why are humans so forgetful compared to NNs?

If forgetting is critical to flexible decision-making in biological brains like rodent studies, as has been argued, in part by analogy to successful machine learning regularization techniques like weight decay or dropout, why do the most successful NNs still remember so much compared to humans? For example, an LLM like GPT-3 can casually memorize entire pages of text after training on it once. This is not an extraordinary capability, but happens routinely with LLMs. And the bigger and better they get, the more they are able to memorize.

Whereas with a human, even ones who pride themselves on their memory for text (like myself), memorizing any page of text is effectively impossible without extensive practice in mnemonic techniques, and we tell awed stories of those few human beings who appear to have photographic memories (and demonstrate it is possible for human brains to do such things, they just don’t usually). The usual human’s lack of a photographic memory, and the shallowness of our understanding, can be demonstrated with hilariously simple examples, like “how is the letter ‘g’ written?” or “which way does Lincoln face on the American penny?” Our memories (and competence in general) are so bad, in fact, that in addition to struggling to remember what we ate for breakfast yesterday or learning basic general knowledge like “the capital of Russia”, we forget almost our entire childhood, and treat this childhood amnesia as completely normal and may even deny that fact.

This is particularly puzzling given that: we never have more neural connections than we do in childhood, as they continually die or are pruned away; it is not universal cross-species; the amnesia doesn’t seem to damage more core knowledge like motor ability or language; important traumatic memories can be retained; and that those lost episodic memories are, to a considerable degree, still there—detectable by implicit measures like lower reaction times or faster recognition of blurred images or needing fewer hints/cues.

Nor is this because memorizing facts is useless to learning; memorization is critical to learning. But then, if memorizing is so useful and more knowledge is better, why do the most successful human memorizers seem to underperform, and it is a mixed blessing at best? Why are examples of photographic memory anecdotally associated with odd limitations in creativity or generalization, like John von Neumann or Luria’s Solomon Shereshevsky, when experimentally in spaced-repetition memory research, better memorization appears to help generalization & understanding without such drawbacks? Why are normal humans capable of memorizing some things, like the Quran or Paradise Lost or the Homeric epics, given enough time & motivation & tricks like memory palaces, but not routinely?

The most extreme cases are idiot savants, whose overall performance profile and pre-DL descriptions bear an uncanny description to LLMs. The feats of savants are too well-documented to doubt, but are impossible to fit comfortably into standard frameworks: if any humans could defeat LLMs in their strong points, it is savants, but then why does savantism seem to only come with such severe costs?

Human Ignorance

And why are humans so ignorant compared to NNs, overall?

Not only do we not recall much easily, we don’t know much either: the breadth of knowledge of even a low-quality LLM in 2024 vastly surpasses that of any individual human. Individual expert humans still usually beat the best LLMs in their narrow niche of expertise, but if we quizzed humans with the same breadth of questions that we benchmark LLMs, I don’t think there is a human alive who would come anywhere near them.

Human Intelligence

So why are humans so smart in general, ie. why do humans generalize more robustly (eschewing the use of non-robust features/‘dimpled’ manifolds like GAN steganography ⁴), are more creative, and largely immune to adversarial examples?

One of the most counterintuitive aspects of NNs is that for all the incredible fluency & genuine capabilities, they nevertheless still occasionally make simple blatant mistakes. LLMs in particular—even excluding technical issues caused by BPEs, RLHF & other safety tuning, greedy sampling, and so on, it remains true that they will make baffling errors which are fatal and cause them to go in circles and fail to generalize for things where there seems to be ample training data.

Another version: given that NNs seem to so easily memorize & overfit (and famously can memorize all training data), and operating by “direct fitting” to ‘interpolate’ between the “underspecified” training data, why do they generalize at all?

Need to Sleep

Why do humans need to sleep—so much so that we usually die if sleep-deprived harshly enough?

In contrast, there is no known equivalent of sleep in current DL, nor any apparent need for it.

Even for continual learning or reinforcement learning replay, ordinary minibatch SGD suffices. There is no need for extended periods of offline inactivity or non-interactive computation simply to train on some new data.

Stage-Wise Development

Why does human development spurt through stages to a much greater degree than NNs do?

Children seem to often develop in phases, with sudden spurts (Van Der Maas et al 2006; van Geert 1991; cf. phase transition neural dynamics). Further, children say, believe, and do all sorts of crazy or stupid or evil things—while people make fun of the bizarre things LLMs sometimes say, LLMs have nothing on the things little kids routinely say. (It makes one wonder how human brains reach the adult stage when they apparently must all go through the little kid stages, where if an adult said such things, they would be committed to a mental hospital or feared as a future serial killer.)

That is strikingly different from NNs. NNs demonstrate various kinds of phases within-training and emergence across-scaling, but many of the identified phase transitions, like ‘induction heads’, barely even change the training loss or more than a few benchmarks. In general, NN training appears much smoother than humans do, aside from changes in learning rates, particularly cyclical learning rates.

Slow Development

Why does intelligent development take so long?

Since larger models are more sample-efficient, and many intelligent animals (and definitely humans) are massively larger, why does it take so long?

One of the most universal observations in NN scaling is that, whatever the task, “larger parameter-count models are more sample-efficient”: more precisely, they reduce the loss more per data point than a smaller model does, and so, when the training is plotted comparing loss with data processed, larger models show a distinctly ‘steeper’ plunge than small models do.⁵ (This raises important questions that I’m not sure we know the answers to: “How sample-efficient are extremely, extremely large models—multiple orders of magnitude larger than we now train? How large do they have to become before they stop becoming more sample-efficient? If they do stop at some point, can better regularization fix them? Why do we not train these if we are data-limited in some regimes?”)

Since many animal brains appear to be so much vastly more parameterized than any DL model ever, and to have a large compute-budget and be bottlenecked by processing samples sequentially, we would naively expect an even steeper plunge. The expense of intelligence would be running the brain, not wallclock time.

But everywhere we look, like primates or cetaceans or elephants or birds, we see that intelligence tends to be associated with prolonged childhoods & longevity & low reproductive rates. These long developmental periods don’t seem explicable in purely metabolic terms—you cannot produce a college-level human intelligence by force-feeding toddlers, and while we do see changes like average of puberty dropping like a stone from contemporary caloric abundance, it doesn’t seem to accelerate brain development much.

And why does the development stop? Why don’t we go on being neuroplastic our whole lives but seem to coast after young adulthood, appearing fully functional even if we’ve almost stopped learning like many old people?

Human Slowness

Why can human development take so long for many things that other animals learn quickly?

For example, some prey animals are able to walk and even run within minutes or hours of birth (thanks in part to fetal training of motor control via twitching), while a human child is both physically & mentally incapable of crawling for >5,000 hours.

And where are the informative priors from evolution? Evolutionary psychology in particular suggested that many complicated social behaviors could be traced to genes and the human brain would be found to have “massive modularity”; this research program has failed completely.

Qualitative Differences in Training Dynamics

Where are the missing equivalents for humans of major NN dynamics like deep double descent (true even with 1,000-degree polynomial splines) & “epoch-wise double descent”, super-convergence, and grokking/catapult or patient-teachers?

And if there are none, why not?

Sleep

Tononi’s SHY is one of the most interesting theories attempting to explain sleep in general (rather than dreaming). In SHY, ‘sleep’ is forced by neural net local learning, where the neural net ‘weights’ increase in size over time as they learn; this is bad because it directly increases the energetic demands of firing a synapse, and because (just like an artificial NN), it makes it easier to memorize rather than generalize. The ‘weights’ need to all be multiplied by a fraction, to shrink them back down to normal energy consumption, and helping regularize the brain; but it can’t be done one by one, because they are all in use.

Sleep, then, is when the entire brain is taken offline, to shrink them all simultaneously. So, you can skip sleep for a while, and your brain keeps chugging along, weights growing slowly and becoming more and more inefficient, but it causes increasing problems, and eventually something diverges.

(Dreams, then are not essential; they are probably more of an enhancement for additional sample-efficiency, by doing a more extreme form of experience replay than can be done during waking.)

Savantism

Savants are often capable of remarkable feats of memory or perception, like making near-photorealistic drawings from a single glance at a photograph or memorizing books.

How is it possible for the (non-autistic) Kim Peek to “at age 16–20 months…memorize every book that was read to him. His parents moved Kim’s finger along each sentence being read.”? He could hardly have developed mnemonic palace techniques as an infant or other adult-like memory tricks like hours of repetition; instead, he seems to simply memorize books in a single epoch, so to speak, to parrot them back like an LLM.

Further, why do Peek & other savants seem to avoid childhood amnesia entirely (accounts emphasize that their memory tends to be lifelong, and Peek seems to remember books read as a young child), while also appearing to not go through childhood development stages?

Why do they usually suffer from severe neurological (associated with left hemisphere trauma) & physical problems and demonstrating remarkable deficits in abstract understanding and general intelligence? For example, “he [Peek] cannot, for example, explain many commonplace proverbs”—and even this level of abstraction understanding some proverbs is considered unusual for a savant! As far back as Down 1887, observers were struck by the contrast between the lifelong “verbal adhesion” of many savants and their understanding, or general intelligence. (Treffert & Wallace 2002 note that of the ~100 well-described cases, the highest IQ was 114.) Strikingly for LLM parallels, savants can also be capable of confabulation: Down 1887_139ya describes an illiterate child who “take up a book…and improvise stories of all kinds with a great deal of skill, and in any variety, to suit the supposed tastes of his auditors”.⁶

In cases of mathematical calculation prodigies like calendar dates or taking roots, the underlying algorithm can be inferred to some degree, and appear to mix huge amounts of memorization & distilled intuition with explicit algorithms. These all appear to benefit from disabling higher levels of the brain.

Savants are often intellectually-disabled or otherwise abnormal. An important example in this context is Luria’s famous mnemonist, Solomon Shereshevsky (The Mind of a Mnemonist), who simply remembered everything he saw or heard: in part by using extensive visualization & mnemonic approaches like the “memory palace”, but his level of synesthesia also meant that everything was inextricably associated for him.⁷

He struggled with non-literal interpretations, reading, or even recognizing faces—he could remember exactly how a person’s face looked previously, of course, but then he could not generalize to how their face looked now. Why did he not benefit from the ability to recall every version of a person’s face, and gain uncanny super-recognizer powers? (See also hyperthymesia: people with HSAM do not have an intrinsically superior memory for learning, but for not forgetting—they seem to struggle to focus on daily life and cope with anxiety or memories of bad days.) He has what sounds like maladaptive daydreaming, and ultimately died of alcoholism.

On a similar note, one of John von Neumann’s famous party tricks was memorizing pages of books, and being able to recite books on command—just one of his many remarkable feats of cognition which have entered legend. But von Neumann was criticized for pedanticism, explicitness, inelegance, brute-force calculation, comfort with following complicated lines of reasoning by reliance on his photographic memory. Eugene Wigner, who knew Von Neumann so well, while describing his feats, struck a cautionary note: none of them were truly original, and Von Neumann himself reportedly said that “[I] will be forgotten while Kurt Gödel is remembered with Pythagoras”. Was von Neumann ever original & creative enough to be truly great like Pythagoras or Gödel? Was he held back by his great knowledge & calculating ability (if not exactly “library work”), and unable to drill to the essence of matters like a Grothendieck?

If we combine the ideas of lack of abstraction or later stages of childhood cognitive development, memorization, synesthesia (failing to separate modalities), injuries to the left hemisphere (particularly the temporal lobe), ability to confabulate, and general extreme imbalances in cognitive capabilities compared to normal humans, I am left with the impression that savants are the “LLM version” of humans. Savants are what can sometimes happen when a key node in the brain is knocked out and higher level processes become less effective, exposing more of the brain’s low-level raw intelligence, which is satisfied by simpler forms of learning like massive memorization.

Grokking

A good example is the much-discussed phenomenon of “grokking”. In the original grokking paper, a small NN is trained on a simple arithmetic problem; as expected, it quickly memorizes the training data, achieving 0% error while failing to generalize with ~100% error on the held-out data. What is more surprising is that after training for long enough, sometimes, apparently at random it gradually begins to improve on the held-out data despite the ~0% error. The odds of grokking are improved the more the NN is regularized—such as by oversampling implied data which cannot be easily memorized, and particularly by weight decay, which tries to shrink the size of the NN weights. (The larger the weights, the easier it is to encode information into them, and memorize the training data.) What is going on?

Mechanistically, what seems to happen when it groks: is that the initial hyperparameters are ‘poorly’ chosen and the NN quickly finds a nearby local optimum in the model-space (loss landscape), representing the ‘memorization’ solution, but that after enough training, the weight decay or other regularization makes the local optimum too ‘narrow’ to contain the NN forever, and then it randomly walks out, or is “catapulted” out of the original solution—and eventually reaches a new region of model-space.

This new region still has ~0% training error… but it corresponds to an algorithm completely different from memorization, which is simpler in an intrinsic sense than “memorizing all the training data” and so unsurprisingly, generalizes better, even in the presence of data error. (It could in fact be many algorithms—there are a surprising number of valid algorithms for that particular arithmetic problem.) And since it is simpler, there are many more possible models which all correspond to the same thing, so this new region of model-space is ‘wider’ than the memorization region, and so the model generally does not randomly walk out of that one; or if it does, it will probably find a third, bigger & better region, and so on.

This is a quite intriguing result, but it’s not obvious how to apply it to anything else, like a full-size LLM. It would doubtless be possible to train an LLM like GPT-3 to zero training error if it was trained for enough epochs—but it would be impossibly expensive, and then you would still have to train long enough past that to trigger “grokking”. Given that it can take hours to trigger “grokking” in even the tiniest possible toy dataset, it is impossible to do true “LLM grokking”. But the phenomenon of grokking is interesting.

Cyclical Learning Rates

A related theme is struck by studies of cyclical learning rates, like super-convergence or cosine learning rates, as used in eg. Chinchilla/MiniCPM (which yielded a large improvement over Kaplan et al 2020) and appears critical in continual learning.

Cyclical learning rate schedules look like simulated annealing in having distinct loss curves: regularly spiking up (very) high when the LR goes high, but then dropping rapidly when the LR is lowered and achieving the best results yet. These often correspond to big capability gains; one can think of the high-learning-rate period as ‘exploring’ to create a new model, while the low-learning-rate periods ‘exploit’ and finetune the new model, only to start all over. (And interestingly, cyclical LRs can be run an indefinite number of times.)

Thus, cyclical LRs can look bad if not properly evaluated, because they trade off short-term loss minimization for long-term gains—if you get them right and run them long enough compared to other LR schedules, “the curves cross”, but until then…

In this respect, cyclical LRs look remarkably like childhood development, complete with periods of high errors while trying to learn new abilities (eg. when children learn sarcasm, it is usually preceded by a period where they make many major errors because they are unable to tell when a negation of an obvious fact is an effective use of sarcasm vs simply false).

Adversarial Examples & Isoperimetry

Consider also another blessing of scale: adversarial examples and the isoperimetry paper.

Adversarial examples, 10 years after discovery, still are not solved, every defense has been broken, and adversarial examples even transfer in many ways, like across models or modalities. Adversarial examples seem to get weaker with parameter scale, but not necessarily with regular compute-optimal scaling. The isoperimetry paper advances the thesis that the reason for this is that because current NNs can be seen as subdividing the latent space into linear volumes, defined piecewise, parameter by parameter, the boundaries between each volume are linear and fragile: instead of ‘curving around’ like they ought to, they simply go straight. Therefore, it is easy to ‘nudge’ any particular data point in a variety of small ways, such that they get pushed just over the artificially simple linear boundaries and into the wrong volume.

The true, correct, curve can be approximated to arbitrary accuracy by simply a lot of small boundary lines, but it may take a lot of them, and thus, a lot of parameters (and data?) to define them all.

This is why all the defenses, like adversarial training, keep failing: they do not solve the basic geometric problem—it’s not a few parameters which are bad, it’s all of them, there are countless boundary lines which are too straight and open for exploitation. Such defenses can, at best, eliminate specific examples and maybe make it harder for a specific attacker to find an example, but they may only push the problem elsewhere, or provide security-by-obscurity. It is also why examples are so stubborn and transfer: models will tend to conceptualize and think in roughly similar ways, and they are all highly under-parameterized, so they will tend to share similar vulnerabilities.

The apparent message of the isoperimetry paper is that you need to scale up your model sizes. And not by a little: their ballpark estimates suggest that models for things like ImageNet are at least 2 OOMs too small. You can imagine what that must imply about LLMs! A GPT-3-175b-scale dense Transformer model might need to be, say, 100 trillion parameters.

Unfortunately, whether you use Kaplan or Chinchilla scaling laws, a 100t-parameter dense LLM would appear to be out of the question for the foreseeable future. To train a model which is able to fit all of those tiny line-segments defined by those 100t parameters, one by one, would require at least as much data—it would require astronomically more data than plausibly exists, and then the training cost of training a model that size on that much data would be equally astronomical.

At least… if you try to fit those parameters with the standard ‘sharp’ (greedily myopic first-order) SGD, looking at each parameter individually. If instead you crossed your eyes, and stared, all of those pixelated lines might blur into smooth curves, and if you thought about it long enough, suddenly at some point the pattern of curves may snap into the drawing of a Dalmatian dog.

This is how NNs catapulting to new model basins may learn intrinsically more robust models which have the blessings of isoperimetry without the brute-force approach of fitting the smooth curves directly to data.

How the Brain Works

[Geoff Hinton] tells of coming home from work one day in a state of great excitement, exclaiming, “I did it! I’ve figured out how the brain works!”

His daughter replied, “Oh Dad, not again!”

—Pedro Domingos (cf. NIPS 2010)

What all of these anomalies seem to share is a core of scaling-law-like relationship of parameters, memorization, generalization, and training—a multi-way bias-variance tradeoff, where different systems hit different points on a Pareto frontier where NNs have low variance in-sample but then high bias out-of-sample or on hard problems, and biological brains are at the other extreme (with an unhappy valley of intermediates which impress no one).

I suggest that the core insight here is that too-extensive memorization is the enemy of abstraction, by leading a model to a local optimum which minimizes error but encodes fundamentally the wrong algorithm. Instead, we must, paradoxically, defy intuitions about overfitting by training as large a model as possible in order to handle as small data as available to a human without overfitting.

What happens when we train a current DL model is that it is lazy, and so it rapidly homes in on the nearest loss local minimum it can find. This minimum tends to be one where it has highly-efficiently memorized the data and learned all the non-robust features and statistical shortcuts; this assemblage of tricks is genuinely effective and intelligent, and is not “cheating” (non-robust features really are present in the heldout data etc.), but they do not generalize far beyond that, because they have not hit on the true underlying algorithm or latent manifold.

Why do grokking NNs seem to need to “memorize to generalize”? Why might it be hard to make progress towards the true target before the memorization phase?

Perhaps because LLMs both memorize and forget datapoints constantly during training (which is why number of copies in the dataset matters: more likely to have seen it recently before the end)—but this forgetting happens for no good reason? “Repetition is the mother of learning” (which is why spaced repetition is good for abstraction, not just brute facts): it is difficult to generalize even the simplest proposition like A + B = C if a NN is constantly forgetting & relearning each part.

Grokking appears to proceed in two phases, first the memorization of all available training data, then the gradual development internally of an increasingly-refined generalizing algorithm.

With that in mind, we might describe the benefit of the memorization as the “learning facts” phase of pedagogy, and then the generalization phase as “the NN thinking about or pondering the facts it has learned until it gets it”. Each minibatch is another ‘thought’ about the data, as the NN struggles to understand the gestalt of the data as more than a bunch of brute facts, and the gradient descent slowly homes in on the generalizing algorithm. (And then memorizing too many facts can sabotage grokking by ‘squeezing out’ the generalizing algorithm because too much needs to be memorized—memorizing a few well-chosen examples is more useful than memorizing countless redundant pieces of trivia.)

That true target may be ‘very distant’ in the loss landscape, and getting there may require an exorbitant amount of data—each data point painfully pushing it ever so slightly out of its comfort zone until one day, it finally is forced by the overwhelming weight of long-tail anomalies to turn into the right model.

The right algorithm will lie in a distant part of the model loss-landscape, but to reach it using a reasonable amount of training data requires the model to travel far (as a kind of grokking/catapult/super-convergence), which is only possible if the model is so overparameterized that it can encode smooth paths (like saddle-points, as all models are linear mode connected seemingly) and it can ensemble over extremely large families of models which ‘blur out’ to a smooth abstraction of the posterior. And then high-learning-rate training is critical to kick it along the frequent saddle-points & plains that slow down optimization, oscillating between high learning rates to escape the current local optimum and lower learning rates to consolidate the gains and find the new escape route. These models inherently require long serial training with many time-steps, and cannot be easily ‘parallelized’ or absorb large amounts of information, and benefit from many long periods of inactivity (‘sleep’) to globally repair the damage from learning or ‘catch up’ on backlogged steps.

Such a model will memorize little of its training data, because that would require rigid, fragile, precise parameters—but those parameters all need to be recycled and explore strange new model loss-landscapes in order to eventually arrive at the promised land. So during training, and even afterwards, such a model will forget and perform badly on benchmarks that reward memorization (such as declarative knowledge)—even though it will avoid adversarial examples (because none of its boundaries are extremely low-parameter linear lines dependent on inhumanly high-frequency texture-biased non-robust features) and will generalize well to hard problems (which by definition make up little of the standard benchmarks) by learning all sub-skills and achieving ‘slingshot generalization’ by mastering all combinations of skills, and resolve many issues with contemporary DL.

Such models will be hard to discover because of the use of early stopping and the general greediness of DL training (and R&D in general), even though the core ideas are well-known and have large associated literatures: like the original DL scaling ideas, the payoff of catapulting simply takes too long to come.

This intelligence is not that hard to evolve⁸, but it usually is not worthwhile. So they fall into an uneasy ecological niche: such robust generalization is not useful for many animals compared to simpler, cheaper, forms of learning, such as association or imitation, or hardwiring directly into genetics. It can only pay back its exorbitant costs if the environment rewards robust generalization appropriately by providing enough high-payoff opportunities which are predictable but not too predictable.

Training a Catapulted LLM

What Is the Largest Useful LLM?

LLMs are more sample-efficient the larger they are. At what parameter scale does this stop holding true? And why don’t we know the answer to this question already?

What would a ‘human-sized LLM’ (or HLLM) look like?

First, it would need to be highly overparameterized relative to the problem as a whole, in order to provide a smoothly-connected loss landscape. Second, because it is so overparameterized, standard hyperparameters will result in overfitting but not catapulting; heavy regularization will be required, and weight decay is probably not enough (as it doesn’t result in much model movement) but a high learning rate schedule may be adequate regularization given prior art like super-convergence & human parallels.

The final full-scale HLLM would probably look something like a dense Transformer or MLP which is multiple orders of magnitude larger than currently trained⁹ (so >10-trillion parameters, possibly >100-trillion¹⁰), in order to be highly-overparameterized compared to the full distribution. (In grokking papers or the isoperimetry estimates, the NNs are generally several OOMs larger than ‘reasonable’, so if we ballpark GPT-4-level models at ~1t, we would weakly expect the catapulting regime to be ~100t.) The NN should probably be a “skinny” one, emphasizing depth over width; a long-standing trend in DL is that ‘wide’ networks tend to memorize more heavily and poorly express more algorithmic/computational reasoning, and that DL NNs tend to look rather un-biological in using so little recurrence/iterative computation or composing reasoning (while RNNs often do better than Transformers in specific cases, generalizing better despite their overall inferiority).¹¹ Very deep networks tend to be avoided due to overfitting or instability rendering their theoretical advantages moot, but catapulting would potentially fix that, and benefit from the inductive biases. This is consistent with the most recent work on sample-efficient LLMs, like Kim et al 2025 or NanoGPT Slowrun, which emphasize increasing parameter size (eg. via ensembling) and heavy regularization over many epochs.

The NN is trained on small text corpuses like BabyLM scale (~0.1b words). One benefit here is that because the sample size has to be limited, that means it can be filtered extremely heavily for quality/deduplication: because the data distribution decides whether a model will meta-learn, the dataset should be as diverse as possible, and penalize memorization as much as possible (see Wang et al 2024). Each minibatch should sample as many distinct datapoints as possible, and likewise, diversity maximized across batches, so each catapulting step is catapulting for as many ‘skills’ or ‘capabilities’ simultaneously as possible. This will delay their learning, because only the bare minimum of data is available per skill, so there is no overkill, and they will tend to be learned simultaneously—which leads to ‘emergence’ as multi-step processes suddenly start becoming possible. (Because humans can replay memories, both in short-term and long-term memory and through the hippocampus, it would be reasonable to do multiple epochs if there is not enough high-quality diverse data for single-epoch training.)

The catapulting itself is due to a cyclical learning rate schedule like super-convergence, perhaps combined with heavy weight decay.¹²

So, what would happen when training our oversized LLM on our highly-diverse memorization/repetition-purified data with cyclical schedules for heavy regularization? We would observe the classic cyclical training loss behaviors of spikes followed by reaching new lows, but with stagnant performance on the truly held-out data, as the LLM goes through the memorization phase, and eventually reaching a new regime where it begins to transition from memorization to generalization over many tasks simultaneously, which will then suddenly ‘emerge’. Each cycle builds up a new set of atomic skills, dependent on the skills learned in previous cycles (analogous to developmental phases).

Prototyping With Arithmetic

It might be feasible to test LLM catapulting on small-scale tasks where current LLMs clearly generalize poorly, like arithmetic. Arithmetic is about the smallest, simplest, easiest-to-generate problem that LLMs still fail in oddly brittle ways on, so it’s a great testbed.

Arithmetic is learnable with appropriate formatting by small cheap LLMs, but standard LLMs (trained on natural arithmetic text data) continue to not implement true arithmetic, even at the capability frontier like GPT-4, and arithmetic problems are easy to generate, benchmark, understand, and even do neural net interpretability on; so one could pilot catapulting on a pretrained LLM by looking for training schedules which make it find true arithmetic much faster than standard finetuning does.

More specifically, one would filter for ‘hard’ arithmetic problems and then search for catapult training recipes which reduce the exponent in the scaling law compared to the ‘standard’ training recipe. If one used regular arithmetic problems, the gain on rare hard problems—the sort which expose the fact that the LLM has only learned a collection of partial heuristics, approximations, and memorized answers—would be hopelessly masked by the average case. (It is entirely possible that before perfect arithmetic performance generalization, a catapulted LLM, which has mostly succeeded in learning true arithmetic, would be outperformed on average by the regular LLM which has memorized as much as possible.)

So one would do something like filtering stringently for the 0.1% (since arithmetic is so big) of the hardest arithmetic problems (as evaluated by an existing LLM or by testing for generalization past n digits), and then use that as the heldout data that one runs scaling law sweeps on for all training recipes.

The scaling laws would ignore the average-case performance from the training runs, and also the constant factor on the hard data, and look for changes in the exponent of the scaling laws for the hard data.

Ideally, one would find something like a training recipe where after many epochs, the catapulted small LLMs are improving more rapidly on the hard data than the standard LLMs are, and that even if the catapulted LLMs are substantially worse everywhere, the more rapid improvement means that at some point “the curves cross”, and the catapulted LLMs are superior. This would then be proof of concept that catapulting is not merely possible on a more complex problem like arithmetic that continues to challenge even SOTA LLMs, but that it changes the scaling laws.

Then to verify this result, one could apply interpretability research to it (like Zhong et al 2023): the final catapulted LLM should clearly express a valid arithmetic algorithm where the standard LLM fails to, and there should be phase transitions across the catapulted LLM’s checkpoints from standard LLM-like pseudo-arithmetics to true arithmetic.

With this proof of concept in hand, one can work on further optimizing the catapult training, and start attempting to infer what catapult training methods might scale up to SOTA LLMs like a GPT-4.

MLPs

Why might MLPs be especially suited here? Because extreme regularization may help fix their persistent overfitting problems and provide superior scaling.

Use of sparsity like mixture-of-experts would tend to reduce the effective parameter-size and the connectivity of the NN landscape, and so would be somewhat risky. However an intriguing possibility is that catapulting might make fully-connected MLP architectures viable.

MLP architectures are much simpler, more general, parameter-efficient, more hardware-friendly than CNNs/Transformers, and look like the logical next candidate for Bitter-Lesson-ing the status quo NN architecture—but MLPs still fall far behind. Why?

The main reason seems to be that they are too powerful, and overfit. (They are to Transformers/CNNs as those architectures are to humans, one might say.) Zhao et al 2021 & Bachmann et al 2023 demonstrate that MLPs scale well and can be competitive if enough regularization (like bottleneck layers) is added. In particular, the more regularization (and consequently, generalization downstream), the more they learn sensible convolution-like features, rather extremely high-frequency & non-robust features.

However, it is still unclear what “regularization” would preserve all the MLP benefits without crippling the architecture—and catapulting fits the bill! The high LR & catapult trajectory would suppress those MLP pathologies the same way it suppresses the milder versions in Transformers. (See Liu et al 2022, and Fan et al 2024 on how narrow deep MLPs seem to grok in unusual & better ways; likewise, He et al 2024 on deeper Transformers.)

Hardware

A serial catapulting regime would render extremely large GPU-clusters much less useful, as they will not be able to step through each iteration much faster; it would instead place a high premium on more exotic hardware like Cerebras chips, which can execute a training step in a small fraction of the time, and hence, wallclock.

See main comment

Incidentally, the use of low-latency hardware would also open up more exotic neural net architectures like AUNN/IFNN.

Prototyping With Image Classification

For non-LLMs like CNNs/MLPs trained on CIFAR-10/CIFAR-100 or ImageNet-1k (MNIST being too trivial to be convincing), we would similarly expect more human-like images. In image classification, it’s harder to define ‘hard’ than arithmetic, and the standard NN accuracy is so high that simply filtering for errors might not yield enough ‘reasonable’ errors at this point.

So we would instead use one of the many ‘post-ImageNet’ image datasets designed to stress-test classifiers, like ImageNet-A, ImageNet-C, ImageNet-Hard, ImageNet-R, & ImageNet-Sketch. These would be used as the target in scaling law sweeps as described previously.

Adversarial Robustness

A possible alternative would be to look at adversarial robustness instead of a standard benchmark.¹³

If the dimpled manifold thesis is correct, then I interpret it as predicting that: while HLLMs might be intrinsically robust per isoperimetry, so too might be tiny models, as long as they found the right manifold which did not require the “dimples” (ie. unprincipled, ad hoc, memorized tweaks to the decision boundaries to get the right answer). This true manifold would potentially be found by catapulting.

So if one successfully catapulted a small CNN, even on CIFAR-10, it might demonstrate adversarial robustness and generalization (rather than the usual small NN choice of either robustness or generalization). This might be especially the case with MLPs, and while an additional research risk, the efficiency of MLPs would allow extensive testing on just 1 GPU (eg. Bachmann et al 2023).

No theory of adversarial examples other than “non-robust features” & “dimpled manifold” predicts that small models might be adversarially robust if simply trained for a long time with an odd learning rate schedule, so any large improvement in adversarial robustness is an important finding.

Prior Art

There is essentially no research on training >10t-parameter LLMs, or cyclical LRs on large LLMs (as opposed to small GPT-2-scale ones).

Historically, this is due to a mix of academic fashion/prejudices and capability gains by smaller models. Work on training highly-overparameterized LLMs, or on the equiparameterization regime ¹⁴, was largely killed by the release of the Chinchilla paper, which provided the perfect excuse for everyone to immediately halt parameter-scaling (since it no longer led immediately to SOTA results), as they have always wanted to, for a mix of good and bad reasons. Similarly, the excuse of “optimizing for inference-optimality” by overtraining small models has become popular, by optimizing scaling laws which assume that the trained model will be naively deployed as-is, without the actual pruning, quantizing, & distilling everyone does anyway.

This means that the field of high-energy DL is wide-open: this proposal will be highly unpopular, it is much less likely than usual that anyone would independently investigate this direction, and they will be discouraged by poor preliminary results when the training runs appear to have simply failed (because of subtle bugs, poor hyperparameters, or simply inadequate training time to catapult into a region where better results can be benchmarked—assuming the right benchmarks are being used to begin with).

Benchmarking

This is such a different training regime that previous scaling law sweeps are inapplicable. Further, the goal here may be one that existing benchmarks actively mislead on. They test mostly common easy questions—the sort where ‘direct fit’-like thinking does best on, by definition.

So a major question here would be whether scaling laws should target perplexity as usual, or if they need to target a custom benchmark which tries to test human-like generalization rather than memorization.

Given the difficulty we have in constructing non-trivia-heavy benchmarks which existing LLMs can’t beat, this might be one of the hardest parts! But I suspect that after training a reasonable HLLM, possibly via trial-and-error, and interacting with it for a while to get an idea of how it acts qualitatively, the right metric might become more obvious.

The benchmarks might include adversarial robustness, hard negative mining (ie. the hardest problems that the best LLMs still get wrong), meta-learning, checking how much sets of models add to an ensemble’s accuracy¹⁵, or use a metric which rewards performance while penalizing memorization of training data.

Capabilities

The final model should generalize much better—possibly achieving the Nyquist learner limit of perfectly modeling the true (non-dimpled?) latent manifold, and thereby constructively answering Rosenfeld’s question about how a band-limited Nyquist learner could be implemented in current NN architectures.

Per the lottery ticket hypothesis, once the true generalizing algorithm has been found, and has been further trained as desired (perhaps on some large trivia-heavy corpus), we can prune it down to a much smaller, faster, more feasible-to-use model, in an example of “train large, then compress”.

Given the capacity of even small models, it may be possible to finetune these smaller models on arbitrarily large amounts of data to exploit the scaling laws of transfer. Because they start with the human-like generalizing prior, they will learn & memorize the data as appropriate, but without the pathologies of ‘direct fit’ to the data as when trained from scratch on the same data—thereby achieving the best of both worlds.

Economic Implications

If catapulting LLMs is enough to close the gaps with humans and solve AGI, then their economics simply turn into a discussion of AGI; but what if they are much better, yet still far from AGI?

As of 2024, the economics of the largest SOTA LLMs are poor, because model capabilities can be so easily cloned by using the cheap APIs to create large training corpuses, enabling behavior-cloning of the superior models. This makes it possible to ‘clone’ the most expensive multi-billion-dollar LLM into a cheap or even FLOSS LLMs (justified by commoditize-your-complement dynamics), eroding margins within months of release—even if the cloned model is broadly inferior. AI scaling companies are trapped in a race to spend the most capital on ever-larger training runs & deploy dirt-cheap distilled versions to gain market-share before the cloners erode their edge, in the hope that they can achieve some network effect or ‘escape velocity’ (like creating AGI).

For a catapulted LLM, however, this would still be the case, but much more so. For catch-up players, catapulting eliminates data-constraints, but the compute-cost (and difficulty) of training a catapulted LLM may be insuperable: the training run itself might not be too expensive, but the trial-and-error and proprietary trade-secret knowledge of the special sauce of how exactly to make catapulting work may be extremely compute-expensive.

And they can continue to do the cloning as usual, but whereas in 2024, the cloned models are not that much inferior and share all the same weaknesses as the SOTA model they are cloning (suffering from adversarial examples, confabulations, bizarre simple errors and brittleness etc.), in a catapult scaling regime, this would not be the case: there would be a clear qualitative difference. In fact, the cloned models may even be superior on many benchmarks, because they are trained so heavily the ‘standard’ low-learning-rate way and enjoy all the benefits of that approach, but still disfavored by users, who can’t afford their unreliability, confabulations, and blandness.

However, the catapulters remain able to do all the usual tricks and can create their own superior ‘clone’ models. (And these clone models can themselves presumably be catapulted for regimes or tasks where there is insufficient data to train the low-learning-rate direct-fit way.) Given the economics of scaling DL, where the compute-cost can be dropped by extremely large amounts while amortizing the initial training cost over many users who provide high usage of GPUs, this further means the first-mover in catapulting potentially can drop prices enough to discourage competition.

So a catapult LLM creator, if the catapulted LLM has enough of a human-like edge in reliability & quality, may be able to maintain high margins much longer than 2024 LLMs do.

Alignment

Speculating further, on the premise of the above: most capability improvements do not help AI alignment as much. This is because they either are specific to a capability in a narrow domain, or they enhance broad capabilities but only in a ‘brittle’ non-generalizing way, which can be highly economically valuable but doesn’t help ‘alignment’ much because we are interested in alignment not for its current narrow in-domain but generalizing. We need the kind of alignment which doesn’t help a chatbot say socially-acceptable things today (or do the right things for the wrong reasons), but which would make chatbots in charge of civilization do socially-acceptable things in the future (and do the right things, for the right reasons, so they will keep on doing the right things indefinitely).

But this catapult or Nyquist learner may be the exception, because what it helps with is true generalization. A catapulted LLM trained on large amounts of morality-related text has not learned purely an assemblage of memorized fragments, heuristics, tricks, and statistical associations, with any underlying algorithm begrudgingly forced by scaling, or learned deception & situated reasoning to maximize a reward, or been adversarially selected by use of ‘interpretability’ techniques to learn to think in nonlinear opaque ways which do not raise any red flags; instead, it has learned the underlying value-manifold the hard way, like a highly-intelligent, grown, moral adult human has.

Even if the capability improvement turns out to be beaten by standard LLM scaling approaches (like simply brute-force annotating every error), true generalization would be invaluable to alignment.

It does not solve the alignment problem in generality—but it might provide a way to create a genuinely moral AI, and that is a good starting point.

Interpretability

Further, because the native neural net way of thinking is a large complicated pastiche of memorization & heuristics, while the overparameterized grokked LLM has distilled out an algorithmic core for the key tasks, such a catapulted LLM ought to be much more useful for interpretability work. We can more easily validate that they are genuinely moral AI.

Once the overparameterization has been removed, what is left should be natively much closer to simple, verifiable, interpretable, and extractable algorithms. These can then be formally analyzed & verified (possibly with the assistance of the genuinely-moral AI LLMs, which are still risky but not too risky if they are probably moral to begin with, and we confine them to carefully-checked formal outputs and tasks like algorithmic equivalence of a sparsified neural network and extracted sub-algorithms.)

External Links

Discussion: HN

Appendix

Dynamic Grokking

An alternative to grokking of full models might be grokking of specific problems. Pretraining might be the cheap but incomplete way of training a brain, followed by invoking a much more expensive “pondering” process which spot-fixes important errors.

How would you grok a specific problem in an LLM? You might do it by training it repeatedly on the problem, finetuning the model on a very small data sample (ie. dynamic evaluation).

Done repeatedly, this could lead the model to explore deeply in the loss landscape, eventually reaching new basins, generating qualitatively different outputs and even having ‘creative breakthroughs’.

Inducing grokking on a pretraining corpus may be neither desirable nor possible because it would require too much training on all problems. However, we may still want to grok on some problems—particularly hard ones.

At present, there is not really any way to take a hard problem that an LLM fails to solve, and let the LLM “think about it”. If a bunch of inner-monologue completions do not solve it, there’s no straightforward way to “ponder”. You cannot spend a lot of money or compute on an important problem to solve it; even GPT-4 o1 appears to have a hard limit, set by its context window size. There is currently no AI equivalent of a human thinking for a long time, seemingly getting nowhere, but then (incubation effect) having an insight and solving it (or at least making useful progress). Grokking, however, looks an awful lot like “a neural net thinking about a problem for a very long time until it starts to understand it and then starts to have insights & solve it”.

What would it look like to try to have an incubation effect in AI?

Pondering ≠ Tree Search

Usually, such ‘search’ is imagined as a kind of MCTS-esque tree search, where the human brain runs a deep tree search ‘in the background’ and eventually it finds a solution and that bursts into consciousness as a eureka moment or L’esprit de l’escalier.

But are humans running some sort of MCTS-esque tree search? I have my doubts. A tree search requires expanding many nodes and building a large deep tree—where is this all happening in the human brain, with its highly-limited working memory? If it’s so deep and requires days or months of search, where is it stored all that time while being constantly updated & traversed, while being robust to all the other activities we do, like sleep? And why does this sort of thinking seem to benefit so much from sleep, even short periods? The kind of hippocampal ‘experience replay’ we know the brain is doing has a highly spatial dimension and is repetitive, and so doesn’t seem too much like we’d expect from a tree search. Why does it seem to require a consistent level of effort over a long time, often with no progress and just beating one’s head against the wall as one goes around in intellectual circles? If we are so good at it that it can be totally unconscious, why do we find conscious tree search, like thinking through lines of play in chess, so difficult? Why can’t we simply run these searches in the background indefinitely, stop thinking about a problem entirely, and then years later be startled by sudden Nobel-Prize-worthy insights out of the blue?

What this looks like to me is closer to an effect of neuroplasticity: the constant repeated thoughts force neuroplasticity and neurological changes, which are then heavily regularized during sleep; if not done often enough, the benefits disappear, but if done often enough, despite no apparent change in the outputs, the brain keeps changing as it tries to find the optimal ‘solution’ to compressing/predicting the problem inputs (and all intermediate results of value); eventually the brain has updated enough times to move to a new basin and experiences a phase shift in the solutions it has been futilely trying, and comes up with a valuable novel output, which suddenly appears in consciousness without any apparent cause—because the cause is the low-level neuron changes which are totally inaccessible to conscious introspection.

Neuroplasticity = Dynamic Evaluation

What would be the closest thing here in LLMs? Neuroplasticity is equivalent to finetuning, or dynamic evaluation—finetuning a model at runtime on the new inputs, which can greatly boost LLM performance.

It has been observed that self-attention is equivalent to gradient descent: it updates the model at runtime by doing a constricted form of gradient descent encoded into the forward pass as “fast” weights computed by the slow weights per-problem. Enough self-attention is equivalent to gradient descent on the same data… but it becomes increasingly inefficient. Dynamic evaluation is similar, and hits a middle point between self-attention and fullblown training. Notably, there is a 3-way tradeoff between model scale, context window scale, and use of dynamic evaluation: dynamic evaluation can boost a small model up to the equivalent of a model which is much larger or where there is much more relevant data in a sufficiently-large context window.¹⁶

Dynamic evaluation works both for supervised learning like image classification, but it was introduced in RNNs for unsupervised learning by next-token prediction, and does not require labels or known-answers: eg. you can do dynamic evaluation on the questions in a dataset, to improve prediction of the still-unknown answers. Using dynamic evaluation lets a model learn and think more deeply about inputs, by letting it do multiple implicit gradient descent steps to update itself per problem: the self-attention tries to analyze the inputs, and then their results are distilled into the weights by the dynamic evaluation, which improves the next self-attention pass, and so on—the learning done during the self-attention gets amortized into the model weights, instead of being thrown away only to be re-computed again, taking up time during the next forward pass.

Dynamic evaluation is usually done as a single-pass for convenience & efficiency, but there is nothing inherently requiring it, and for difficult inputs, you might want to do more than one, and run multiple steps of dynamic evaluation. This is because two gradient descent steps are not necessarily the same as one step with a larger learning rate, and the more iterative “looks” at the data, the more the model can “change its mind” in a way that a single step can’t. (An analogy would be experience replay, where a RL model can learn from a datapoint dozens of times, or expert iteration like AlphaZero: the model and the search bootstrap each other up, which is highly non-stationary, where one alone would simply sigmoid hard.)

One might then think that this can’t be useful because given the sample-efficiency of LLMs, one or two steps is probably enough to memorize the current input, and that one risks catastrophic forgetting… but that is an intuition which seems to be wrong in practice—LLMs memorize many things, and yet generalize, catastrophic forgetting just doesn’t happen with large pretrained models due to their extreme overparameterization & sample-efficiency enabling continual learning, and grokking & related phenomena show that there is still benefit to training long after it would appear to be useless.

Repeated Neuroplasticity As Implicit Search

So if we put this together, we can imagine using dynamic evaluation to try to trigger a kind of pondering which avoids tree search and can potentially lead to grokking: call it dynamic grokking.

In dynamic grokking, we spend a lot of compute to try to solve a specific hard problem. We repeatedly do dynamic evaluation on the input to update the model, corresponding to a “learning” neuroplasticity step, and roll out a new completion, corresponding to a “pondering” step. If the completion is correct, or we run out of compute, we stop; otherwise, we repeat.

During dynamic grokking, the LLM will update possibly thousands of times, immediately memorizing the input but hopefully gradually changing deep within, eventually generating novel completions and solutions that the original unimproved ‘base’ model could not have generated even with tens of thousands of brute-forced completions. These final results can then be given to the user, used as the seed for other search techniques or a new dynamic grokking, trained on to improve base models, etc.

Extensions to this would include: periodically doing dynamic evaluation but with heavy weight decay regularization, corresponding to a “sleep” step; updating the input with the best completions or parts of a completion (possibly using a second prompt to generate a useful summary of one completion); attempting to systematically vary the context to produce less redundant updates (eg. dropping out parts of the input, retrieving different documents from a database, injecting noise into the forward passes); and saving the model snapshots to merge together into a better base model (which aggregated over many problems, may steer the model towards a MAML-esque meta-learning model which learns how to dynamically learn). Given the repeated iteration of training, this would benefit from hardware setups focused on minimizing latency, like Cerebras. (It is possible that a small model using dynamic grokking on Cerebras-like hardware might be able to outperform time-to-solution of even a highly-parallelized but static large model, depending on how poorly static models scale their searches.)

As far as I am aware, there is no prior art relevant to dynamic grokking, because dynamic evaluation is almost completely unused, and the few inner-monologue approaches which do gradient descent at all will usually only do it with something like a subset of completions filtered for quality.

While dynamic grokking can be researched or tested with the usual toy problems and approaches, it’s unclear how you could, even in principle, get an idea if dynamic grokking is working on a specific problem. The externally-observable outputs (tokens or logits) might not change at all for many iterations; and all the ‘internal’ metrics for grokking I’ve seen are unscalable, architecture/problem-specific, noisy, or otherwise flawed.

A personal example: my mother tells me that when we got our first Mario Brothers video game, I spent weeks playing it by running Mario to the first pit and deliberately jumping into it just to hear the sound effects of Mario dying (which I apparently found hysterically funny), and that it drove her crazy.

What could my brain have possibly learned from the 5,000^th repetition of the sound effect? And if I wasted weeks on learning nothing, then how did I manage to be even more sample-efficient in the rest of my life?↩︎
Similarly, children do not seem to learn languages all that more efficiently than adults and have some special sample-efficiency; an adult who is serious (ie. spends hours a day studying using total immersion & spaced repetition, not pretending to learn using Duolingo) will not take 18 years to reach native proficiency. Like learning new systems, the reason children ‘beat’ adults seems to come down mostly to a willingness to do the work—to be embarrassed, to fail, and to just spend as many hours as it takes obsessing over the new thing rather than resting on one’s laurels.↩︎
When we dare to project out scaling laws all the way to human or superhuman performance, they generally do not require absurd amounts like quadrillions of parameters, which would appear to be implied by most of the biological-supremacy views.↩︎
Aside from not making use of non-robust features, despite their power & omnipresence, humans appear to be almost completely blind to them:↩︎
We do not simply scale up models to the largest one that fits in our hardware because they cost more compute per-step, so a compute-optimal small model will process much more data with the same compute budget and eventually beat them.↩︎
This is much less remarked on in the savant literature. Is confabulation under-documented because no one expects it, and so ignores it or fails to test for it (eg. by asking trick questions)? Or do savants refuse to answer such questions, if it is not in one of their special interests or if they feel confused when they try to answer it? (Down also notes that savants are not always cooperative, and one can imagine that it might take a long period of building rapport to reach the point where one could ask anything, instead of the usual displays of skill which they may take great pride in.)↩︎
Synesthesia:

S.’s case, as many readers have noted, resembles the Jorge Luis Borges story “Funes the Memorious”, a fictional work about a man plagued by the persistence of his memory. “To think is to forget a difference, to generalize, to abstract”, Borges writes. “In the overly replete world of Funes there were nothing but details, almost contiguous details.” Similarly, Luria writes that for S. almost every word, every thought, was freighted with excessive detail. When he heard “restaurant”, for example, he would picture an entrance, customers, a Romanian orchestra tuning up to play for them, and so on. Like Funes, S. had a sort of private language to catalogue the richness of his mental associations. The word for “roach” in Yiddish could also mean, in his mind, a dent in a metal chamber pot, a crust of black bread, and the light cast by a lamp that fails to push back all the darkness in a room.

↩︎
In fact, this sort of delayed or ‘deceptive’ payoff is an optimization task that evolutionary strategies are competitive with gradient-based methods, because they optimize the net fitness over a lifetime; it does not matter to evolution how useless an infant looks on benchmarks (yielding a deceptive gradient) if the converged adult has higher fitness.↩︎
In the terminology of Huang et al 2024, we normally train models in the “progression”/“semi-grokking” regimes (underfitting), but by increasing the model size, we can move it to the “grokking” regime, where it may memorize but then generalizes much better than progressing models; and if we filter the data stringently enough to minimize the ability to memorize, this ‘ratio’ will move the crossover point much earlier in training (see also Huang et al 2024’s observation about how adding a pure-memorization task “poisons” grokking, which parallels Dohmatob et al 2024 about noisy/low-quality synthetic data).↩︎
Such large networks have been demonstrated to run on GPU clusters, although they have all been mixture-of-experts or recommender models.↩︎
Whether RNNs can grok/catapult, and if that could help them leapfrog Transformers, is an interesting question that I don’t believe has been tried much, given the focus on fully-connected/CNN/Transformers.↩︎
Or would it make more sense to instead have a constant large learning rate, and instead cycle the weight decay, to let weights grow and then brutally prune them back, to increase the analogy to “sleep”? Smith 2022 claims that a high-learning-rate+cyclical-WD schedule works & may increase sample-efficiency.↩︎
One could try to test this with LLMs too, but adversarial examples are not as well understood or easy with LLMs compared to the standard CNN/ViT image classification adversarial examples, so would not be as good for prototyping.↩︎
One of the reasons to favor overparameterized models is that an extreme level of parameterization (compared to Chinchilla) keeps showing up in scaling theory papers as optimal, and in particular, a 1:1 equiparameterization ratio—one datapoint, one feature.

This seems especially intuitive if datapoints are optimally selected (or synthesized), and can in theory achieve much more favorable scaling like exponential decrease in error.↩︎
To the extent that normal models are all trapped in the same nearby basins, they will make the same sorts of errors and ensembling will not improve metrics much.

To the extent that they are exploring the full loss landscape and finding far away models, they should have much less correlated errors and so ensemble better than expected.↩︎
Note that dynamic evaluation is better than the alternatives in the sense that we can always do dynamic evaluation on a model we are using, whereas we usually cannot use a larger model nor put more into the context window. (And if we could, then we could do dynamic evaluation on that one instead.)↩︎

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Bibliography

[Bibliography of links/references used in page]

Style-Aware Generative Models

Gwern Branwen — Thu, 21 May 2026 00:00:00 GMT

Image generators can now produce striking images in almost any named visual style (as well as unnamed), but this is not the same as generating and inventing a style. This is because a style is a historically situated package in a context: images, techniques, names, rivalries, institutions, and adoption by artists and audiences.

This suggests a different target for style-aware generative models, for which I propose a concrete implementation strategy using contrastive learning.

Instead of treating style as an ordinary prompt modifier or post hoc cluster, you could train embeddings to represent chronology and influence, then use them to infer and simulate directed counterfactual histories of art.

Such a system would support causal credit assignment, temporal holdouts, and forward simulation. It could ask how much a work like the Mona Lisa mattered after deleting its descendants, or sample plausible post-AI-slop movements by tracing a style lineage forward through virtual artists, critics, audiences, and adoption dynamics. The pipeline is: images → embeddings → influence graph → counterfactual histories → virtual artists → adoption dynamics → movements.

How can a generative model invent a style?

What Image Generators Don’t Do…

Invent Styles

Image generator models can now generate an extraordinary variety of images, and in some sense, are capable of generating every possible style (because almost arbitrary images can be embedded near-losslessly into their latent spaces). When DALL·E 3 came out, I argued that while image generators could create every possible style and you can easily create striking new images—and indeed, in mid-2026, if you just sample random images in Midjourney v8, you’ll see all sorts of esthetically striking images—this didn’t mean that image generators would create any recognizable new styles.

This would be in part due to the collapse of cultural filtering mechanisms of “everything everywhere all at once” so meaningful new movements are difficult to create, but also due to a similar “everything all at once” context collapse in image generators. Because a style is not so much a few visual flourishes, but a coherent package, in a historical context, advocated and adopted by groups of humans, while an image generator inherently tends to treat images as a collage of discrete elements to be pulled apart and recombined, with ‘styles’ existing mostly post hoc and crystallizing some common collage with a name. Thus, while DALL·E 3 did wind up inventing a style—the now notorious “Shrimp Jesus” AI slop style that heralded the current flood of synthetic media everywhere—it could only do so accidentally, as forced to by preference-learning training gone wrong, and canonized retrospectively by humans looking at the corpus as a whole. DALL·E 3 itself possessed no way to ‘invent new styles’. It could invent “Shrimp Jesus” by being mode-collapsed onto that style, but it could not invent “a response to Shrimp Jesus after that soon got boring”.

Assign Credit

A related question is, how do we do causal attribution in an image generator model? We would like to be able to answer questions like “how much did this one painting contribute to the final usefulness or knowledge of an image generator?”, for topics like potentially assigning intellectual property rights or just understanding the history of art.

There are generally two answers given to this, one of which is cheap and popular and wrong, and the other correct but too expensive and limited:

The right answer to the wrong question is to train many models and omit a specific datapoint (LOOCV).

This is irrelevant because what if we did LOOCV and dropped out the Mona Lisa, but kept the parodies? The image generator would almost certainly be able to generate ‘the Simpsons Mona Lisa but in the style of the Renaissance painter Leonardo da Vinci’ which may look almost identical. The LOOCV is indeed delivering a correct answer—the Mona Lisa is in fact redundant as training data given all of its parodies and references—but it turns out we were trying to answer a different question of causal credit assignment, more like, “what if the Mona Lisa had never been painted? How much lesser would the image generator be if culture were impoverished in that way?” (Fancier versions of LOOCV, like influence functions, are also ill-posed in the same way: they examine “which datapoints shifted which generative model outputs?”, not, “which historical artworks shifted all subsequent artworks—including the generative models’ outputs?”)
The cheap wrong answer to the right question is to take the model outputs and look up similar images in the training dataset and say, post hoc ergo propter hoc.

This is wrong because images can be similar but not have caused each other. Two parodies of the Mona Lisa may have nothing whatsoever to do with each other no matter how similar they look. If two people get the same idea to do a Simpsons parody of the Mona Lisa, a similarity metric will probably say one caused the other, which is wrong.

Style Embeddings

In 2022, I suggested that perhaps a generative model could be made ‘style-aware’ by looking at clusters in embedding-space:

…CAN is a multi-agent approach in trying to create novelty, but I think you can probably do something much simpler by directly targeting this idea of new-but-not-too-new, by exploiting embeddings of real data.

If you embed & cluster your training data using the style-specific latents (which you’ve found by one of many existing approaches like embedding the names of stylistic movements to see what latents they average out to controlling, or by training a classifier, or just rating manually by eye), styles will form island-chains of works in each style, surrounded by darkness. One can look for suspicious holes, areas of darkness which get a high likelihood from the model, but are anomalously underrepresented in terms of how many embedded datapoints are nearby; these are ‘missing’ styles. The missing styles around a popular style are valuable directions to explore, something like alternative futures: ‘Impressionism wound up going thattaway but it could also have gone off this other way’. These could seed CAN approaches, or they could be used to bias regular generation: what if when a user prompts ‘Impressionist’ and gets back a dozen sample, each one is deliberately diversified to sample from a different missing style immediately adjacent to the ‘Impressionist’ point?

Temporal Embeddings

I think the causal attribution question suggests a better approach here to constructing a latent space with useful embeddings.

A regular embedding will prioritize compression and explaining variance, particularly by semantics or overlap of appearance. The latent space will be highly entangled, and counterintuitive (as we can see from looking at interpolations); much of it will be devoted to properties like ‘a person wearing glasses’. There will be temporal aspects, but it will not be a priority of the embedding.

But we can encourage a model to prioritize temporal relationships by training on time metadata. We want to heavily overweight training to emphasize time, rather than leaving it implicit and just another minor latent as far as an embedding model like CLIP is concerned, to be scrambled up as convenient.

Specifically, we want an embedding which pushes artworks together which are close in time, and pushes apart artworks which are remote in time, even if they are otherwise very similar. This could possibly be done by a standard contrastive learning approach: inside a batch, all datapoints have a creation date, and each embedding is pushed towards temporally similar datapoints, and away from non-similar datapoints. All available metadata is used for this similarity, of course, but the creation/publication date is the most important one. This can be done weighted with a dataset enriched for labeled dates (eg. WikiArt or ART500K), and a second temporal contrastive loss in addition to a generative loss. (If our preferred generative architecture is too weak to have a meaningful latent to train contrastively, we could train a pure embedding model like a CLIP, to use for steering a regular generative model.)

Such a contrastive training process would hopefully create a latent space where artworks are mostly arranged over time, in a loose tree/reticulate-DAG-like structure: datapoints organized into thick dense bands by year, but organized internally by style/influence, adjacent to its predecessors, and successors. Works that have many influences throughout history, like Egyptian or Greek, will ‘bulge’ in high-dimensional space, and be ‘near’ all points, correctly reflecting that fact that art movements keep reaching back for inspiration or in reaction to key works. Because later artworks cannot cause earlier ones, causality emerges implicitly, as simply the best compression strategy. And each real datapoint’s embedding is surrounded by the universe of plausible counterfactual artworks which could have been created at that time, but weren’t. The causal relationships do not need to be exact nor do we need to account for selection bias in how most art is not preserved (never mind in the training dataset), because that means those works are not easily available to contemporary artists either, and missing works can be inferred.

The time dimension will not be a single rigid dimension, but it can be extracted easily from the embeddings by contrast pairs or supervised learning, and like any latent in a GAN like ‘eyeglasses’ or ‘red hair’, it can be manipulated.

Simulating Art History

That manipulation provides precisely the ‘island hopping’ capability we want: we can explore and replay history in various ways, and we can extrapolate out new histories, by sampling new latent points under the constraint of moving monotonically in the extracted time dimension.

Our samples would include small random jumps, corresponding to incremental works, but also occasional larger random jumps, imitating Levy flights for exploration. So if we want to invent a new, post-Shrimp Jesus, style, we embed Shrimp Jesus at 2023, and randomly successors out in a Shrimp-Jesus-rooted tree of hypothetical responses by artists (some negating it, some building on it, some just ‘near by’ fellow travelers), and derive new ‘styles’, and responses to those styles, and so on.

We can work on causal attribution by going back in time and creating filtered views of art history. For example, we could try to estimate the Mona Lisa’s impact by embedding it, and deleting all datapoints which ‘descend from it’ (ie. are in a kind of lightcone in the latent space with the point at the Mona Lisa and expanding from there), creating a counterfactual art history where the Mona Lisa never existed, and then training a new model on this new art history. And if we find that this is too devastating a criterion, because the Mona Lisa turns out to be upstream of too much of all Western art (perhaps analogous to “pedigree collapse”), we can loosen our filtering process to soft weights: each datapoint gets an adjusted weight based on phylogenetic approaches by sampling and counting out paths from ancestor to descendant works, so direct copies of the Mona Lisa get ~0 weight but mere oil paintings are weighted ~1. We can even resample art history: resample the training set, refit the tree, retrain the counterfactual-deletion model—repeat across bootstraps. Now the Mona Lisa experiment doesn’t return “the counterfactual model is 12% worse”, it returns “12% worse (4–19%).”

Further, we can try to improve the models iteratively by feeding back in the computed phylogenetic tree as a superior temporal distance (with distance measures like number of intermediate nodes or shared ancestors to define the triplets): train, fit and extract the best fitting phylogenetic tree (which obeys all known dates as hard causal requirements), compute all pairwise distances, retrain the contrastive embedding with those as the new distances, and repeat. This is an EM-like pseudo-labeling semi-supervised learning process which should converge to a self-consistent view of art history. This helps fix any issues with the long-range influences encoded into potentially ill-behaved ‘bulges’ by doing direct credit assignment. And a reasonable historical tree itself enables interesting applications, like plausible reconstructions of ancestral works, or discovery of unknown hidden influences like ‘this African mask influenced that French painter’ (either model errors to fix or new research topics). Overconfidence is likely, but not necessarily important if we just want a much better historical-causal understanding or novel research hypotheses.

We can evaluate a counterfactual-history generative model in various ways, like, ‘how much harder is it to embed a real datapoint into it’ or ‘how much less likely is it to generate samples that look like the Mona Lisa’ or ‘how much less do end-users value this model’.

Use-Cases

This finally provides a way to attack our IP problem in a philosophically and esthetically meaningful way, rather than narrowly legalistic: we can now speak of how much causal influence (including direct copying) particular copyrighted works had on a sample, in a way which is not fooled by mere visual similarity or answering an irrelevant question about a training dataset tweak, but would help copyright serve its primary purpose of helping “promote the Progress of Science and useful Arts” by rewarding copyrights which did influence others, and not rewarding copyrights which did not do that. And then we can think about schemes like compulsory licensing which would not simply worsen the problem.¹

Or we could truncate sampling at a certain date, to study what subsequent art movements were most novel: under a temporal holdout like 1850_176ya, was Cubism or Impressionism more ‘surprising’? (This is a diagnostic for the EM self-training working, because later art works should be much more predictable after several cycles of rationalization.) And which one was more important in the objective sense of causing more later art-works, net of all downstream works/movements? And which ones were most fertile in the objective sense of causing more known styles (even if those were not collectively influential or popular)?

Generating Styles

And we can easily run history forward to try to predict plausible new styles, and build up a future of a tree of branching new styles, each of which has its own historical context and is a reaction to or inspired by earlier hypothetical new styles. (Plausibility can be approximated in various ways, so we can also search for the most plausible futures; this will weakly correlate with quality.)

We can also condition on new samples, and let a user curate a new tree, backtracking as they wish. These clusters can be turned into named styles by feeding samples into a VLM, or turned into usable keywords by averaging and creating a LoRA or steering vector, added or subtracted, etc. And we can further simulate out schools by learning the artist latent dimension, and fixing them, and drawing sequences of samples from artists within a style over time—this gives us a ‘school’ of ‘artists’ which are exploring a ‘style’ in their personal ‘careers’.

Once we structure the latent space properly to encourage modeling historical causality and context of artworks, rather than naive i.i.d. sampling, we can begin to speak of a model which can meaningfully generate styles rather than just images.

Inventing Styles

We could even—why not—try to turn each of those latent artist points into a full artist biography to feed into an LLM, by generating summaries/analyses of sequences of work, and confabulating a biography and esthetic and ideology to retrodict those. In the limit, the generative models are sampling policies with memory, taste, constraints, rivalries, medium preferences, social networks, and career trajectories, with parasocial fandoms. And now we have virtual artists which can power Internet presences and interact with each other or with their fans. (For example, a simulated artificial-life community could follow a simple loop like each timestep: artists view recent works; generate new works; critics/audiences select; selected works update the influence graph; and movements are named when clusters become persistent.) They would not merely generate isolated images; they would generate lineages, schools, rivalries, reactions, and stale phases (cf. VTubers).

At that point, “style invention” means something stronger than producing a novel image cluster. It means generating an esthetic lineage that remains recognizable, develops internal variation, acquires simulated advocates and opponents, and survives long enough to be named. That is still only a simulation. But it is the right object to simulate.

Much commentary on AI copyright is based on inchoate notions of moral dessert or anti-corporate activism, rather than on a solid understanding of what copyright should do, and so they propose solutions like “style should be intellectual property” or “all ML models should be considered derivative works of all training data, rather than transformative works”, which would predictably backfire by eg. privileging large corporations like Adobe or Disney, which already have datasets of wholly owned copyrights that they can train on, while making it impossible for everyone else to train on feasible datasets because the resulting models might be considered derivative works of billions of copyrights and utterly intractable to clear rights on.↩︎

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Bibliography

[Bibliography of links/references used in page]

NYT and Vaping: How to Lie By Saying Only True Things

Gwern Branwen — Invalid Date

The 2019–2020 EVALI lung injury crisis was caused by illicit THC vapes adulterated with vitamin E acetate, but it became a public justification for restrictions on legal nicotine vaping.

This essay looks at a 2022 New York Times article about teenage vaping which juxtaposes synthetic nicotine, flavored vapes, Lizzie Burgess’s hospitalization, and nicotine addiction without ever quite saying that legal flavored nicotine vapes caused her lung injury.

Sentence by sentence, the article’s misleading effect comes from technically true phrasing: “vaping”, “vaping THC and nicotine”, “vaping-related lung injury”, and post-hoc narrative order.

The case is a useful example because the ground truth is clear, the article was high-profile and edited, and the manipulation is grammatical and selection rather than simple falsehoods.

The 2019 nicotine vaping moral panic was one of the more striking ones of our era, where illegal black-market THC marijuana vapes poisoned with vitamin E acetate that caused severe “EVALI” lung injuries were a major justification for unprecedented bans on legal nicotine vaping products. Despite no lab-verified instance of a nicotine product with the vitamin E acetate adulteration (only THC products), and nicotine vapes produced by large businesses which had sold countless billions of doses to millions of people beforehand, with no EVALI epidemic before (or since) 2019–2020, and the phenomenon subsiding after lawsuits and prosecutions of vitamin E acetate sellers, it somehow became common knowledge that nicotine vapes were deadly.

How did this happen?

Media Coverage

One way it happened was via a conflation of THC and nicotine vaping, driven in part by inaccurate reporting by victims and by the difficulty of proving a negative—after all, any given case could have been due to nicotine vaping, somehow, how would one prove it wasn’t? This was then amplified by activists and journalists working together to palter, by selectively reporting and describing true facts. (I call this strategy “paltering” as phrases like “propaganda” or “sins of omission” do not capture this phenomenon.)

I was particularly struck by an instance in the New York Times in March 2022, which paints a compelling picture of legal nicotine vapes causing severe lung problems in one young woman, Lizzie Burgess:

[The Lizzie Burgess excerpt]

This article appears to have been successful in persuading readers. In the NYT comment section (mirrored below in full), most comments agreed with the article, often going as far as to claim that “addict and kill people—including children”/“here they are exploiting a legal loophole to addict millions more young adults to their lethal products” or “Young people especially become easily addicted to the nicotine…The damage to their lungs resembles the worst pneumonia you can get and still be alive.” or “The purveyors of these drug delivery systems are killing our children with intent. They should all be thrown in jail, and the import of these products should be banned.” (The import of THC vapes contaminated with vitamin E acetate is, of course, already banned or adult ID required etc.) Indeed, only a single NYT commenter (who no one responded to) appeared to get it right in their reply a quarter down the comment section:

The illness described in the article was clearly caused by using an adulterated, illegal, THC-containing vaping product, not the nicotine-based e-cigarettes being criticized here. That should have been made clearer.

The article might have convinced me too if I hadn’t already known better.

The NYT has always been my favorite newspaper, even as a kid using money from my part-time job to subscribe to the Sunday Times¹, and I was perplexed why it would lie like that, as the article clearly said repeatedly that nicotine vaping was to blame.

I had thought that it was now uncontroversial and consensus that the moral panic had framed nicotine vaping (albeit unpopular to acknowledge now that the ‘issue-attention cycle’ had moved on and the activists gotten what they wanted). Was it just wrong, or had I misunderstood something? Perhaps I had skimmed over a smoking gun, like finally finding fatal levels of vitamin E acetate in lab testing of a Juul pod, there or elsewhere?

Rhetoric

So I went back to excerpt the article—and while rereading, realized that in every case I thought might be a lie, rereading showed it had been so carefully worded as to be technically correct and to simply be misleading.

After this, I was in awe of the rare artistry of the piece, which elevated it from merely misleading to a wonderful example of “bounded distrust” (cf. “no evidence”) in journalism and how selective quotation and mere juxtaposition can lead astray the insufficiently skeptical reader.

It’s a great example because it is:

A controversy with a ground-truth objective answer, known then, and even more clearly confirmed by years of followup:

The sudden EVALI vitamin E acetate crisis, which lead to a sudden spike in severe lung problems across the USA in 2019–2020, was caused by adulterated THC products. There is no hard lab-tested evidence that any legal nicotine products, the subject of the article, had vitamin E acetate contamination. The general removal of vitamin E acetate from THC products was followed by the crisis subsiding, even as legal nicotine vaping remained extremely widespread.
Longform (~1,800 words) high-profile piece in one of the most prestigious newspapers in the world:

meaning that it passed editing and fact-checking
Purely textual/grammatical:

phrases like “vaping products” or “vaping THC and nicotine” are suspiciously worded and violate ordinary norms of writing to be specific and informative—if a young woman has been hospitalized by a nicotine vaping product, why not say just so? If vitamin E acetate had been found in legal nicotine vaping products, why not just name the test result and product specifically?
Unnecessary to the core of the article:

it would be possible to write this story without bringing in irrelevant THC-contaminated EVALI anecdotes, by focusing on legitimate criticisms of nicotine vaping. (You could discuss teen access, flavor marketing, age checks, FDA jurisdiction, statutory drafting, the economics of disposable devices, and the adult harm-reduction case without ever mentioning vitamin E-caused EVALI which you know is not related to teen access to legal anything.)
Deliberate: the phrasing is so consistently and carefully done that the author almost certainly knew what they were doing—if they hadn’t, they would have slipped up at least once, and joined two sentences with a comma, or some other easy writing mistake which would have been incorrect.
Well done: often, paltering requires highly unnatural ‘lawyer-ese’ phrasing to be misleading but still technically correct. This kind is inherently unconvincing because people can sense and distrust it, even if they can’t verbalize their unease

I kept thinking about what a great example it was and that at some point I should fisk it. Finally, I did.

The article’s misleading chain of logic is synthetic nicotine → vaping flavors → “vaping-related lung injury” → Lizzie Burgess; and in an anecdote demonstrating this: “flavored vapes” → “vaping THC and using a device” → “vaping-related lung injury” → “nicotine addiction” → ban “flavored vapes”.

But nowhere does it say anything like:

Her lung injury was caused by legal flavored nicotine vapes.

It is always ‘vaping caused lung injury’ (true! vaping THC products did) or ‘vaping THC and nicotine’ (the ‘and’ does a lot of work there), or juxtaposition like ‘she was soon in the ICU…She’s struggled to end her nicotine addiction’.

It just carefully arranges the sequence of facts to encourage the ordinary causal inference of post hoc ergo propter hoc, downplaying or omitting the THC, while using supersets like ‘vaping’ or ‘vaping THC and nicotine’ so the article technically contains the relevant information (the informed reader would interpret this correctly as Lizzie Burgess having almost surely been hospitalized by the THC vaping, but the ordinary reader has already been led down the wrong path—and confirmation bias does the rest).

It serves as a good exercise for the reader: how do you handle science journalism like this, where the author is not careless or egregiously lying, but where the author appears to know exactly where the line is and never quite crosses it? Or readers who ask only, “is each sentence technically true and lets me believe what I want to believe?” Would you have noticed the suspiciously specific wording, or the omission of how it never quite says, despite its length, a simple statement like “nicotine vapes caused all EVALI”? Or the implied (and falsified) predictions like “the EVALI crisis spike will continue until legal vaping products are fully banned”? Or would you have fallen for a good Narrative?

Excerpts

Below are some key excerpts, with weasel words or rhetoric highlighted in bold:

Scientists are just beginning to look at the unknown health impacts of new, tobacco-free nicotine products, even as research is expanding into the harmful effects of vaping and its flavor ingredients.

…Scientists are just beginning to study the unknown health effects of synthetic nicotine, even as research is expanding into the harm caused by vaping and flavor ingredients alongside continuing cases of devastating vaping-related lung injury. To many public health advocates, new trends in the vaping industry are thwarting the FDA’s efforts to protect a new generation from nicotine addiction.

…FDA enforcement actions have had little effect on Lizzie Burgess’s ability to get vapes over the last 4 years. “I think the FDA should take it all off the market now”, she said.

…Vaping is still popular among teenagers. Rani Dhiman, 16, said it is highly visible in the bathrooms and stairwells of her high school in the Detroit suburbs.

…She said the stress and loneliness of the pandemic might have been a trigger for some teenagers to start. It’s also portrayed glamorously, she pointed out, in “Euphoria”, a popular HBO series about a teenager kicking drug addiction.

…The FDA’s efforts to limit teenagers’ access to flavored vapes had little effect on Lizzie Burgess’s ability to get them over the last 4 years in the Indianapolis suburbs. Within weeks of starting to vape at 16, she said, she was addicted. There was always a gas station, older friend or website selling e-cigarettes in flavors like banana ice cream or sour apple, she said.

…At 19, she said, she was vaping THC and using a device—now advertising tobacco-free nicotine—that has as much nicotine as two packs of cigarettes, every 2–3 days. She said she fell ill with what started like a cold, which progressed to rapid breathing, almost-gray lips and feeling depleted. By the time she went to the emergency room, her oxygen saturation was 67, far below the normal range of 95 or higher. Burgess said she was soon in the ICU with vaping-related lung injury.

[photo of young woman in ICU with face covered by a medical ventilation mask]

[Burgess in the ICU last year. When she arrived at the hospital, her oxygen saturation was 67, far below the normal range of 95 or higher. Credit: Lizzie Burgess]

[X-ray scan of two pairs of lungs. Top: “Liz’s Lungs”, off-color and shrunken. Bottom: “Healthy Lungs”, white and large.]

…She’s struggled to end her nicotine addiction and is down to two cigarettes a day.

…“I think the FDA should take it all off the market now”, Burgess said of the flavored vapes. “I think it will be very very hard for them to reel it all in. It’s so big and there are so many companies now.”

…[Sales data] released by the CDC Foundation shows that since the FDA stepped up e-cigarette enforcement in February 2020, sales of disposable fruit-flavored and candy-flavored devices have grown by 290%, to 6.46 million devices a month by November 2021. Sales of the FDA-targeted flavored pod and cartridge devices have nearly vanished.

…The unregulated vaping market at this point is a problem of the FDA’s making, said Gregory Conley, president of the American Vaping Association, an industry trade group. He said the agency fueled the problem by over-regulating a product used by millions of adults who find vaping a safer alternative to smoking.

…“This country should learn some lessons from past prohibitions that failed miserably”, Conley said. “If you don’t fairly regulate a market where there is a great deal of demand from legal adults, you will fuel gray and black markets where the operators are not concerned with checking IDs before selling.”

…Recent research has focused in on the chemicals used to simulate butter, which is linked to lung damage, and vanilla, which is associated with birth defects in zebrafish.

External Links

Discussion: HN

Appendix

The NYT Article

“The Loophole That’s Fueling a Return to Teenage Vaping: Sales are rising of flavored e-cigarettes using synthetic nicotine that evades regulatory oversight, a gap that lawmakers are now trying to close”:

By Christina Jewett

March 8, 2022

[image of assorted vaping products on a white table]

Scientists are just beginning to look at the unknown health impacts of new, tobacco-free nicotine products, even as research is expanding into the harmful effects of vaping and its flavor ingredients.

The Food and Drug Administration’s crackdown on flavored e-cigarettes in 2020 was meant to be a comprehensive, aggressive strategy to curtail the epidemic of teenage vaping.

But two years later, sales of disposable, flavored e-cigarettes have soared. Some companies have moved just beyond the reach of the FDA by swapping out one key ingredient. They have circumvented federal oversight of tobacco plant-derived nicotine by using an unregulated synthetic version.

The agency had nearly wiped out the use of flavors in devices like Juul, once the teenage favorite, that could be refilled with pods in flavors like creme brûlée and mango. Jumping into the breach, though, companies like the teen favorite Puff Bar are selling disposable devices filled with candy flavors and tobacco-free or synthetic nicotine.

Scientists are just beginning to study the unknown health effects of synthetic nicotine, even as research is expanding into the harm caused by vaping and flavor ingredients alongside continuing cases of devastating vaping-related lung injury. To many public health advocates, new trends in the vaping industry are thwarting the FDA’s efforts to protect a new generation from nicotine addiction.

“These companies like Puff Bar and others are deliberately driving their trucks of poison through this huge loophole”, said Meredith Berkman, a founder of Parents Against Vaping E-Cigs. She recently hosted a webinar about synthetic nicotine attended by 700 people. “We think we need to regulate these products.”

Lawmakers on Tuesday proposed language that they want inserted in the Congressional omnibus budget bill that would give the FDA authority to regulate synthetic nicotine, although it is unclear if the issue will be included in the final bill.

Representative Frank Pallone Junior Democrat of New Jersey and chairman of the House Energy and Commerce Committee, said the provision would be a public health victory over “bad actors” who circumvented the FDA’s authority.

“That ends with passage of this bill, which will close this loophole and clarify FDA’s authority to regulate all tobacco products, including those containing synthetic nicotine”, Pallone said in a statement on Tuesday.

Sales of synthetic or tobacco-free nicotine went from virtually nonexistent in 2020 to taking up shelf space in 2⁄3^rds of US vape shops in 2021, according to market research. Those stores said such products accounted for nearly 20% of sales, according to ECigIntelligence, which surveys hundreds of the shops each year. The company projected that the US vape market, web sales included, would be nearly $6 billion this year.

FDA enforcement actions have had little effect on Lizzie Burgess’s ability to get vapes over the last 4 years. “I think the FDA should take it all off the market now”, she said.

[Photo of young white woman in a bomber jacket and jeans, looking sad.]

[Credit: Lee Klafczynski for The New York Times]

Federal officials have been in a cat-and-mouse game with some e-cigarette makers. Spurred by a court order, the FDA forced thousands of e-cigarette companies, including Juul, to apply in 2020 for authorization to remain on the market. With the agency focused on the most popular devices, like Juul’s, that used insertable cartridges, makers of disposable vape pens in flavors like gummy bear and candy cane flooded the market. The agency then responded with a stern warning and even product seizures aimed at some of those companies, including Puff Bar.

By late last year, more than a million tobacco-sales applications had been denied. Applications to remain on the market by Juul and myriad other companies are pending.

By early 2021, Puff Bar returned to the market with “tobacco-free” or synthetic nicotine that didn’t fall under FDA oversight, loaded with the fruity flavors prohibited in vapes with tobacco-based nicotine. Other companies imported similar devices containing synthetic or tobacco-free nicotine from factories in Shenzhen, China, according to industry experts.

Patrick Beltran, who has identified himself in news reports as one of two executives of Puff Bar, did not respond to requests for comment.

Sales data released by the CDC Foundation shows that since the FDA stepped up e-cigarette enforcement in February 2020, sales of disposable fruit & candy-flavored devices have grown by 290%, to 6.46 million devices a month by November 2021. Sales of the FDA-targeted flavored pod and cartridge devices have nearly vanished.

Since early 2020, overall e-cigarette sales are up nearly 50% to about 22 million units per month, according to Information Resources, a data tracking consultant. The National Youth Tobacco Survey conducted in early 2021, when many students were learning via Zoom, reported that, overall, about 11% of high school students used e-cigarettes.

New e-cigarette suppliers can go into business easily: They contract with a manufacturer in China, set up a website and get space in a warehouse to store and ship devices, said Samantha Shusterman, a senior counsel supervising e-cigarette enforcement for the Massachusetts attorney general’s office. They use shell companies and can quickly withdraw the profits if they face scrutiny.

“It’s a whack-a-mole situation”, said Shusterman, whose state banned all flavored e-cigarettes, except in licensed smoking bars. “They’re not following any of the laws.”

Mitch Zeller, director of the FDA’s Center for Tobacco Products, said the agency recognized the problem.

“Disposable e-cigarettes made only with synthetic nicotine pose a particular challenge for the FDA when it comes to our tobacco authorities”, Zeller said in an email. “The FDA is actively investigating this issue and considering how to best address such products.”

[photo of young woman in ICU with face covered by a medical ventilation mask]

[Burgess in the ICU last year. When she arrived at the hospital, her oxygen saturation was 67, far below the normal range of 95 or higher. Credit: Lizzie Burgess]

[X-ray scan of two pairs of lungs. Top: “Liz’s Lungs”, off-color and shrunken. Bottom: “Healthy Lungs”, white and large.]

[A scan of Burgess’s lungs, top, compared with healthy lungs. Credit: Lizzie Burgess]

Vaping is still popular among teenagers. Rani Dhiman, 16, said it is highly visible in the bathrooms and stairwells of her high school in the Detroit suburbs.

She said the stress and loneliness of the pandemic might have been a trigger for some teenagers to start. It’s also portrayed glamorously, she pointed out, in Euphoria, a popular HBO series about a teenager kicking drug addiction.

“Sometimes so many people are vaping in the bathrooms, it’s hard to do anything about it”, Dhiman said, adding that she doesn’t vape.

The FDA’s efforts to limit teenagers’ access to flavored vapes had little effect on Lizzie Burgess’s ability to get them over the last 4 years in the Indianapolis suburbs. Within weeks of starting to vape at 16, she said, she was addicted. There was always a gas station, older friend or website selling e-cigarettes in flavors like banana ice cream or sour apple, she said.

At 19, she said, she was vaping THC and using a device—now advertising tobacco-free nicotine—that has as much nicotine as two packs of cigarettes, every 2–3 days. She said she fell ill with what started like a cold, which progressed to rapid breathing, almost-gray lips and feeling depleted. By the time she went to the emergency room, her oxygen saturation was 67, far below the normal range of 95 or higher. Burgess said she was soon in the ICU with vaping-related lung injury.

She’s struggled to end her nicotine addiction and is down to two cigarettes a day.

“I think the FDA should take it all off the market now”, Burgess said of the flavored vapes. “I think it will be very very hard for them to reel it all in. It’s so big and there are so many companies now.”

Synthetic nicotine remains far more expensive than the tobacco-derived product, leading some industry experts to question whether a device label of “synthetic” is accurate.

The unregulated vaping market at this point is a problem of the FDA’s making, said Gregory Conley, president of the American Vaping Association, an industry trade group. He said the agency fueled the problem by over-regulating a product used by millions of adults who find vaping a safer alternative to smoking.

“This country should learn some lessons from past prohibitions that failed miserably”, Conley said. “If you don’t fairly regulate a market where there is a great deal of demand from legal adults, you will fuel gray and black markets where the operators are not concerned with checking IDs before selling.”

[Photo of professorial old white man sitting on desk]

[Dr. Robert Jackler of Stanford University created a “synthetic teenager” named Joe, who used gift cards to buy flavored synthetic nicotine products online and have them shipped. Credit: Ian C. Bates for The New York Times]

Dr. Robert Jackler, who studies tobacco company advertising at Stanford University, has also noted major tobacco retailers entering the synthetic nicotine market with flavored gums called “pouches.” He said his tobacco research group could pose as a teenager and use gift cards to easily buy the flavored synthetic nicotine gums from major retailers and have them shipped to a home in California.

“When we buy them, there’s no age gating”, Dr. Jackler said.

The loopholes are many with synthetic nicotine, he said, allowing the products to avoid hefty tobacco taxes and remain affordable and to evade the algorithms that online retailers use to weed out underage sales of tobacco products. The ease of purchasing was also concerning, Dr. Jackler said, given how little is known about the health effects of flavored, synthetic nicotine.

Recent research has focused in on the chemicals used to simulate butter, which is linked to lung damage, and vanilla, which is associated with birth defects in zebrafish.

Dr. Sven-Eric Jordt, an associate professor at Duke University who has studied synthetic nicotine, said it posed many unknowns.

About 99% of tobacco-derived nicotine is a psychoactive molecule called S-nicotine, he said. But a mirror-image molecule, known as R-nicotine, makes up 50% of most types of synthetic nicotine. He said the R-nicotine molecule appears to be less addictive, but very little research has been done on it in animals or humans.

“It could alter nerve transmission in the brain in different ways from classic nicotine”, Dr. Jordt said, “but we don’t understand that at this time.”

Comments

NYT reader comments, mirrored to show how the article’s framing was received.

Off-brand vapes with real nicotine are trivial to find and buy. All the FDA’s ban did was drive users to less safe products, which is exactly what anyone paying attention predicted at the time.

Vaping hasn’t been shown to be unsafe scientifically. I don’t do it; I don’t like it. But this ban did exactly what every ban in history has done: No one vapes less, but now there are fewer quality (and health) controls.

It would be great if the New York Times could get perspectives from informed harm reduction advocates, who understand that vaping nicotine is not dangerous.

Charles A Gardner would be a good voice on these issues.

Shame on the legislators who wrote the bill and didn’t close this loophole. They were either ignorant or complicit.

Hmmm chemicals that make meth fentanyl and puff vapes all get their chemicals from China and the bought off politicians who represent us are paid off, sounds like the Chinese are returning the opium wars of the 19^th century on us.

I don’t drink, smoke cigarettes or pot, use drugs, gamble for money, cheat on my wife or my taxes. ride motorcycles, put on tattoos, ski down mountainsides, go on hikes alone in the forest, believe in bitcoins or eat at roadside restaurants with signs that say “Good Eats.” And I do ever not count on the government or my next door neighbors to shovel the snow off of my sidewalk for me, Who am I? I am a believer in G—d’s great hidden 11^th Commandment: “First Above All Else, Take Good Care Of Yourself.” Am I the last normal, decent, careful, self-sufficient American still alive?

Synthetic nicotine needs to be put under the regulatory authority of FDA. The fact that these companies can sell synthetic nicotine e-cigarettes without scientific data on their purity and biological effects is highly concerning.

However, in the end this step will help Big Tobacco the most. Synthetic nicotine threatens Big Tobacco’s business model. Only Big Tobacco has the money to pay for the complex toxicological and behavioral studies needed to get marketing approval for e-cigarettes by FDA. At this time, only 1 e-cigarette product, manufactured by RJ Reynolds, has received market authorization by FDA, while millions of applications were denied. The products with synthetic nicotine will likely suffer the same fate, since manufacturers don’t have the funds for the FDA applications.

Thus, the revised legislation will, in fact, be a boon for Big Tobacco. It is likely that it will be supported by legislators on both sides, the legislators in the pocket of Big Tobacco, and legislators concerned about public health.

None of the rhetoric from politicians about synthetic nicotine, or vaping in general, is about preventing youth access or preventing disease. It’s about money. MSA payments from the cigarette companies to the States from [??] totaled ~$178.1 billion—but these same States have spent only ~$0.72 billion on Tobacco Prevention. If cigarette consumers switch to vaping, the States lose ever increasing amounts of their general funds.

@RuthieB—we keep telling them but it goes in one ear and out the other. Which is exactly what the states, the politicians and the tobacco companies want. It’s where the money lies.

This was so predictable—in fact, many did predict it, including laypersons like myself. These companies don’t care about health, particularly children’s health, if they see a market for their product. One company, Vuse, is now even marketing itself as “environmentally progressive” because they have removed the plastic wrappers from their products, thus eliminating tons and tons of plastic from the planet. Yet they don’t talk about the vapors that addict and kill people—including children—and also pollute the air. These companies are reprehensible and have absolutely no regard for human life.

@Maani Rantel—We’ll, big tobacco and big Pharma don’t care about anything but profits. And they don’t have to worry about ‘carding’ people. There is where your problem lies! If there were only small business vapor shops there wouldn’t be a problem. THEY card. They care about their business. And they care about the product and it’s refill ingredients. But when you only target the general product then you get what you got. All the products in the article picture are from those big companies. Everyone loses then. Lies spun around a great innovation that drops lethal hazards to a person by 95–98% are more believable than the truth. People want to believe those negative stories. It’s easier than actually doing their own reading on all the research done by responsible doctors/researchers. Just why didn’t they try to destroy the tobacco companies instead? With nothing good coming out of a cigarette why are they still on the market? Plenty of research on them. Wait…you have to read.

@Maani Rantel One thing this article completely misses is the fact that no nicotine-based e-cigarette (tobacco-derived or synthetic) is known to have caused a single death. Not one, ever. Compare that to 1,300 deaths DAILY from combustible cigarettes. The illness described in the article was clearly caused by using an adulterated, illegal, THC-containing vaping product, not the nicotine-based e-cigarettes being criticized here. That should have been made clearer. Nicotine does NOT cause the diseases caused by inhaling smoke, or nicotine medications would be deadly—but the FDA approves them (correctly) as both effective and safe. The illnesses and deaths result from 70 known carcinogens and other toxins in SMOKE, produced when the contents are burned. People need to understand these differences.

How pathetic is our country, the land of the free, when young adults can serve life sentences but cannot drink a beer or vape? If kids are mature enough to be charged as adults than they should be able to decide what goes in their bodies.

“Federal officials have been in a cat-and-mouse game with some [drug] makers.” this is why trying to ban cannabis (and other drugs like cocaine) has failed, and in some cases violently so, for various parties involved. moreover, prison sentences offer their own costs when the mice are caught. fines or community service could be a substitute, and in some cases they are. as for nicotine, no article is complete without a comparison with combustible and other forms of tobacco, both in terms of consumption patterns (for minor users) and in terms of laws and their applications. Various surveys show that traditional tobacco consumption is at historic lows, especially for minors.

I’m confused as to how these companies can “circumvent federal oversight” by using synthetic instead of plant-based nicotine. I don’t think you can “circumvent” the federal oversight of mescaline, by making it in a lab and possessing it, instead of possessing peyote (plant-based source). Can anyone shine a light on the discrepancy?

@Med Chem Guy I actually work at a vape shop in Pennsylvania. We scan every ID to make sure no kids are getting anything. So, the legislation that (as of writing) currently is in place bans any flavored TOBACCO products. Because the nicotine is not sourced from tobacco it was technically allowed.

The solution is obvious: states and municipalities should impose exorbitant “sin taxes” on these products as to make their purchase out of reach for teenagers, most of whom are of limited means.

While legalized prohibition has been a disaster and a prime example of the “cure” being far worse than the disease, making these products prohibitively expensive by means of aggressive taxation, policies which could put a deep dent in the sales of vaping devices.

The revenue generated could then be used to fund smoking cessation and substance abuse programs along with public service advertising, all of which could be of help in reducing the damage to health which is becoming associated with the use of synthetic nicotine.

@mosenblum this is absurd. People have a right to engage in activities of their choosing without overreach from the federal and or state government. Why should the same thinking not apply to any other so-called vice? Should junk foods be exorbitantly taxed? How about sex toys? Leave and let be.

@mosenblum—first, parents are buying these ‘devices’ for their teens. And second, teens have their own credit cards, their ‘means’ are more than you realize. And third, where’s the sin part in an electronic cigarette used for tobacco harm reduction? The thousands of dangerous and dirty chemicals are missing. There’s no tar. It doesn’t cause disease it prevents it…so where’s the sin to tax? See, here’s the example of people only listening to social media and false information instead of researching for themselves.

I had recently turned 21 when we were told that second-hand smoke had been proven so definitively to be harmful that it was banned in bars and workplaces. My state banned vaping in bars and workplaces before a single study had shown any harm was caused by the mist created by such a device. At some point we should acknowledge that in addition to well meaning and science-backed regulation, we have allowed prohibitionists to lead the discussion and paint everything they don’t like as something that needs to be forbidden to all.

@Brian M The studies that found the danger in second hand smoke were epidemiological studies that found higher rates of cancers (among other ailments) in people with high exposure to second hand smoke. How long and how many dead should there be to test out the “mist” of a vape, sir?

It is NOT just a “mist” that vapers exhale. There are many chemicals and heavy metals in the liquid that is burned. Young people especially become easily addicted to the nicotine because it is so EASY to learn how to vape because the vapor is cool. Some kids vape 2 or more pods a day—equal to 2 packs of cigarettes a day. The damage to their lungs resembles the worst pneumonia you can get and still be alive. Adults who think vaping might aid them quitting smoking regular cigarettes more often just become double users and are just as addicted.

You make a valid point. Perhaps my choice of “mist” was loaded. I should have said “aerosol byproduct” or any number of things that do not sound as innocent as mist.

Let them vape. Let them smoke. Let them take drugs. Let them drink. Let them eat cake. Where there’s a will, there’s a way. People, whether 14 or 44, are going to do what they want to do. The government and industry can not stop it. And actions have consequences. Perhaps many will learn their lesson before it’s too late. And if not, so be it. Like with Covid, too much time is spent on bemoaning those who do the wrong thing. Society will benefit when those who care more about getting high, no matter how it impacts themselves or others, learn that they’ve made mistakes. I suspect ending up in an ER will be a valuable teaching tool.

These companies are intentionally targeting and addicting tweens and teens to a highly addictive product that over time will prove to be disastrous for their respiratory systems. Instead of weak kneed regulation of the industry, the FDA might want to get a backbone and outlaw sales as a health measure that will save countless young people misery in adult life.

The War on Vaping will, I’m sure, turn out to be worth it years from now. Remember how we got rid of drug use?

@Seymour Butz—People should watch “https://m.youtube.com/watch?v=9vh_YzyzUys It’s called The Fix on Roku TV. I found it on YouTube too. It’s basically about drugs. However, it’s very eye-opening. Tells you how we all have been lied to. Just like this topic. You all are being lied to on such a high level it boggles the mind. Why are they not targeting cigarettes‽? Still on the market. Research by the droves proving the dangers. Yet they skip that step and target the harm reduction innovation that left out thousands of chemicals, carcinogens and tar.

The purveyors of these drug delivery systems are killing our children with intent. They should all be thrown in jail, and the import of these products should be banned. This is vile.

Predatory business practices. Perhaps the area of business that the consumer protection agency is responsible for should be expanded.

Ah the cigarette company executives are at it again. After spending decades during the 20^th century and tens of millions of dollars in legal fees blocking the medical facts from the public that smoking causes lung cancer and emphysema, here they are exploiting a legal loophole to addict millions more young adults to their lethal products. And I bet they’ve got dozens of lawyer/lobbyists on Capitol Hill trying to defend this loophole. What a surprise! Like the old saying goes—history repeats itself.

Tempting I’m sure to suggest legalization and taxation, but the issue isn’t as clear cut here. The issue is that kids are using these like candy, because they taste like candy and are being marketed like fun candy times. It was one thing, “back in the day” for most parents (myself as a teenager in the 90s included), where it was tough to hide smoking, because smoking stinks. The smell sticks to you and follows you around. Now we have these USB devices, with a smell that can be easily overlooked, assuming one even notices. Your sons and daughters can use this despite all of your best efforts. And no, “education starts at home” is not a public health policy we can all rely on. The simple reality is that these need to be flat-out banned. We don’t need every detail right now to be millimeter precise about it—just pass a blanket ban and plug the holes as needed.

@C. Smoking causes 480,000 deaths (COPD, strokes, cancer) per year in the US. And you want to have the most effective quitting option banned? Because the teen experimentation accelerated the decline in youth smoking rates? Unlike what the framing campaigns suggest, e-cigarettes weren’t invented by Big Tobacco, nor are they anywhere near as harmful or addictive. (Tobacco isn’t just nicotine. They sure made people believe it though.) You know, before going for another super successful prohibition attempt, might be worth checking out marketing regulations, or better sales enforcement. With the caveat that absolute prevention will force teenagers to relapse to smoking. (And that’s probably what these AstroTurf articles are really about. Adult relapse or handing the vaping market over to Big Tobacco is just a nice side effect.)

Just legalize it. People are never going to not want nicotine. Tax it heavy and use the money to help people.

These products are legal, for adults. The whole point being at that age you are informed enough to make a decision about your health and the possible risks. I’m not naive, I snuck cigarettes when I was a teenager, but allowing kids wholesale access to these products (which deliver far more nicotine than traditional cigarettes) doesn’t seem like the answer.

You don’t put out a wildfire by pouring gasoline on it.

My parents were content with the local daily paper, and my school only subscribed to the weekday edition.↩︎

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Bibliography

[Bibliography of links/references used in page]

Top-k Reading: Against Polishing the Median Essay

Gwern Branwen — Mon, 11 May 2026 00:00:00 GMT

How should a writer trade off new writing (exploration) against polishing old writing (exploitation)?

I suggest a simple order statistics toy model of readers with a finite ‘window’ of reading time for an author’s best writing, where each new writing has a chance of being good enough vs polishing an old favorite. A new writing only matters if it enters the window; an edit only matters if it improves something in the window.

The implication is simple. Polish when a marginal edit improves work inside many readers’ windows: consensus hits, durable pieces, introductory essays, or high-stakes publications. Write more when readers have large windows, heterogeneous tastes, or when a new topic can create a different local hit for a different audience.

For eclectic generalists with diverse readerships like me, the strategy is not to perfect the median piece. Instead: polish the winners, ignore mid-tier work, and keep trying to create new local hits.

Years ago, I discussed the apparent “perfection premium” in polishing writing, and how a writer should quickly hook readers upfront in a new writing. But how should we think in the first place about polish in writing vs new writing?

The Trade-Off

Most writers online, I think, give up relatively soon. This is not because we ran out of things to do—there’s always another typo we could fix… But when we choose to spend our limited time and energy hunting for typos or fixing formatting, there’s a definite sense in which we decide it’s no longer ‘worthwhile’ trying to improve old writing after some point, and we’re better off doing new writing.

But this is a little odd: if we’re not ‘done’ with old writing, why should we start something new? And if we have a large backcatalogue, we can hardly plead poverty of material; how can I argue with a straight face that my readers lack for things to read and so it is imperative I go write something new rather than polish an old popular essay?

Marginal returns. The economics or optimal foraging answer would be the heuristic ‘you should switch between whichever has the highest marginal return at any given moment’—but marginal return of what? Even assuming we have some ‘utility’ definition we’re satisfied with, how, exactly, does writing turn into that? If I write an essay or I polish an old one, what does that mean for a reader?

A toy model that seems plausible to me is that what is going on is that each reader has a certain finite size or appetite or window of time, which a writer tries to fill up as best he can. (I’ll use time as my unit here instead of number of words or essays, because I think that increasingly, it is time and attention which is the bottleneck.) One reader has a window of 5 minutes of ‘Gwern reading time’ a year, while another has 50 minutes, and a big fan has 500… The absolute size of the backcatalogue doesn’t matter for each reader—only the best k minutes of your writings.

So, when you write, if a new piece doesn’t make it into the top-k and is only the k+1^th best reading-minute, then a reader doesn’t care. They’re never going to see it and it doesn’t exist for them.

This helps account for why writing is such a “hit-driven business”: after a while, no matter how prolific you are, it makes zero difference to the reader, because you filled up the top-k in their window and nothing budges after that.

Self-Cannibalization

To move the needle, you need to beat your best and push something out of their window. And frustratingly, once you do that, the reader is only slightly happier with you!

Net, not total. This is because you are cannibalizing yourself. If you wrote something which was a +10 in some Quality Unit, and then you write something which was +11, you feel proud of yourself for adding a +11 and thus writing a total +21 of text—that was hard work—but as far as the reader is concerned, it’s just a net +1 because your +11 pushed out the +10.

Perversely, the more you write, the harder it gets to improve your window. If a reader has a window of 10 essays and you’ve written 100 essays to date, then your next essay has to beat >90 of your prior essays to make it into the window at all!

Simply put, I think you should polish in inverse proportion to how much of your writing a reader will see; the more diverse your writing, the better off you are writing more rather than better.

Order Statistics

Rank entry. We can treat this statistically using order statistics; for example, if we imagine some normally-distributed quality statistic, our expected gain is only:

R simulation of the marginal gain from adding 1 normally-distributed piece to a 100-piece back catalog with a 10-piece reader window.

set.seed(2026-05-11)

windowGain <- function(backlog_n = 100, window_k = 10, simulations = 1e7) {
    mean(replicate(simulations, {
        old_quality <- rnorm(backlog_n)
        new_quality <- rnorm(1)
        cutoff <- sort(old_quality, decreasing = TRUE)[window_k]

        max(0, new_quality - cutoff)
    }))
}

windowGain()
# [1] 0.0474129394

~0.05σ better—which is almost invisible percentile-wise.

Equal odds. The simplification of treating pieces as independent samples is crude, but it resembles the equal-odds model summarized by Jung et al 2015: hit-rate is approximately stable, so quantity matters because it buys more rolls of the dice. For writing, this assumption is partly defended by offsetting forces: positive factors like experience are offset by negative ones like using up ideas.

Lumpy returns. More realistic distributions do not change the basic rank fact. With n old pieces and a k-piece window, a new independent piece enters the window with probability k⁄(n+1), distribution-free. A power law or other heavy-tailed quality distribution changes the payoff conditional on entry, not the entry probability: the process becomes more hit-driven, with long stretches of zero marginal value punctuated by rare large replacements.

Polish the Window

Guaranteed gain. This also adds a little insight to the value of polish: while it’s true fixing typos in your best writings doesn’t make them much better, it does make them a little better… And importantly, it’s making the writings in the window better. When you write a new one, you usually fail to replace something in the window, rendering it moot, and if you do, it’s probably a modest improvement; but when you edit the writings in the window, you are getting a small but guaranteed improvement.

Tiny windows. This sheds some light on why you might escalate your quality standards going from a blog post to, say, a New York Times article. (Jasmine Sun mentioned in a recent Q&A at Inkhaven 2 that her NYT piece on peptides might have 10 million readers total, vastly exceeding her lifetime total of Substack readers—so of course she spent months sweating every detail!)

Decay. Meanwhile, a breaking-events periodical newsletter like Zvi Mowshowitz’s will often not bother to fix typos or broken URLs, because the temporal decay guarantees even a good issue will soon drop out of the window, and it needs to rush to write the next one.

For the former, you are talking to a small steady audience with a large window; there’s little point in polishing it. But for the latter, you may get only one, and so the window is small—you’d better make it as good as you can!

Heterogeneity

So, to the extent that your readers are homogenous and have a small window and a universally shared ranking, the better off you are, after a while, trying to polish your writing winners, because you aren’t going to easily beat them but you can still make them a bit better.

But to the extent that they have large windows or heterogeneous different rankings (eg. readers prefer completely different articles), you’re better off just writing more in the hopes of getting a new better one into their windows. The marginal return on every new essay may be quite low (because its topic will appeal to only a small fraction of readers) but it is also true that the return is even lower on polish (because those too are niche but the effort is constant). In the extreme case, every reader has a window size of 1 and prefers a different essay… So you are best off spending 100% of your time writing on a new topic every time, because each one will find a different audience.

In full generality, we could advise a writer to polish in proportion to:

(1⧸Window) × Homogeneity × Shelf-life × Stranger-fraction

And to write more in inverse proportion to all 4.

Eclectic Generalists

My case. For my writing, I think it falls into the latter group: I do not write for prestige or popular periodicals, but on many topics for different readers (eg. darknets vs poetry vs AI vs genetics…). This also describes writers I am sometimes grouped with, like Scott Alexander or Tyler Cowen’s Marginal Revolution. We all share, if nothing else, a generalist eclectic approach. Under this model, Scott Alexander, who has a recognizable ‘beat’, should probably invest more in polish, while I should emphasize more diverse new writing rather than over-polishing, and Tyler Cowen is probably fine in his omnivorous niche of writing about anything with no polish.

Self-indulgence. So I should be writing a lot more on more topics, instead of falling into the artisanal/hobbyist/indie-creator trap of spending as much time as I do on polish and infrastructure. What can I say? I don’t write to be popular, and the polish is for me a personal esthetic and an ethos, not a calculated strategy. (I try not to pretend otherwise, and I tell any aspiring writer who asks that they definitely should not try to imitate Gwern.net but focus on writing. They have my permission to fiddle around with typography after they have, let’s say, a million words—at which point they may have some writing worth polishing.)

But other writers might find this a helpful frame: do your readers have a lot of appetite each, or a little? And are they homogenous or heterogeneous? This can help guide you in deciding how much to invest in copyediting, illustrations, marketing, etc.

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Bibliography

[Bibliography of links/references used in page]

‘Try, Score, Change’: Reinforcement Learning for Children

Gwern Branwen — Sat, 09 May 2026 00:00:00 GMT

This is an explainer of deep reinforcement learning, written for children, in the controlled-English dialect Grow-Speech, by LLMs, as an experiment in creative nonfiction writing.

Grow-Speech requires all multi-syllabic words to be defined before use, and my hope is that this trick can help demystify the field of reinforcement learning, which, despite intimidating mathematical apparatus and unfamiliar jargon, often comes down to a surprisingly simple trick or approximation—just applied at scale.

Read This First

This is a text on how a brain can learn to walk a maze, win a game, or chat with you. It learns not just from rules told to it, but from a loop: try, score, change.

I will not start with a long list of new words. A long word will come up when I need it. The first time I do, I will mark it in bold and tell you what I mean. Then that word is mine to use as I please.

Read this with a school room in mind. A black board is up front. A piece of chalk is on the rail. A row of jars, a heap of stones, a short leash, a sketch pad, a chess board, and 8 cakes sit near by. You could build most of this on the floor.

How the Loop Works

A Brain Made of Small Parts

By wire I mean a link in a brain that lets one part send a shock to the next. We run a brain on a chip; and by chip I mean a small flat thing made of wires that does fast math. A brain is made of neurons, and by neuron I mean a small part of a brain that takes in shocks from wires and sends out shocks too.

Each neuron gets shocks from the rest through wires. If those shocks add up past a high mark, the neuron fires and sends out a shock too.

We change how a brain thinks by changing how strong each wire is. We can change the high mark too. Then one neuron may fire more, and one may fire less. That is what folk mean when they say, “the brain learns”.

Here is the prop for wires. Think of each wire as a small jar full of stones. The more stones, the stronger the wire. To make the wire strong, drop a stone in. To make it weak, take a stone out.

When the brain does well, we may add stones to the wires that helped. When it does ill, we may take stones out. Once you can flip wires from strong to weak and back, you can make a brain learn.

Try, Score, Change

A brain learns to do well at a task by try and miss. At each step, the brain looks at where it stands, picks a move, and the world hands back a score. By score I mean a count of points the world hands back: high means good, low means bad. High score for a good move. Low score, or no score, for a bad one. The brain wants the best score it can get, on the whole, in the long run.

Think of a dog and a treat. You do not tell the dog all the rules. You just give it a treat when it sits. Soon, the dog sits to get the treat. A bot is the same: hand it score, and it learns to do more of what got score.

Or think of a small bot in a maze. It steps left and hits a wall. Bad score. It steps right and finds cheese. Good score. At first, the bot knows no path. It has to try, get score, and change.

By reinforcement learning I mean a way for a brain to learn from score.

Reinforcement learning is not a spell. It is not joy or pain in the bot. The score is just a count that tells us how to change the brain.

Thanks and Blame

The hard part is thanks and blame. The score for a move can show up long past the move that earned it. The bot might step left, then up, then up, then right, and then find cheese. Which step earned the cheese? The last step? A step 10 steps back? Each step in the chain wants its share. But the world hands back just one score at the end. So the brain has to spread that score back through the chain.

This is the core hard task: hand thanks and blame to the right past moves. Each way to learn that we will see is a way to crack this. None of them are quite right. On the whole, they work. On the whole.

Try New Things or Use Old Ones

At each step, the bot feels two pushes.

The first push says: do what worked best so far. You know that move. It gave you score. It is safe.

The next push says: try a new thing. A new move might not be like you thought. It might be much, much better than the old one. But it might be a waste. Or a trap.

These two pushes do not pull in the same way. You can not do both at once.

These two pushes have one name: the explore-exploit trade-off: Explore means try a move you have not tried much, or one that looks weak now, just in case it is strong; Exploit means pick the move that looks best from what you know now.

You face this too. Think of the town book room. You have a card that lets you take home one book each week. You know the first shelf well. You have read 3 books from it and loved each one. Right next to that shelf are 50 more shelves you have not yet looked at. One of them might hold a book you would love more than the 3 you know.

Each week, what do you do? Walk to the shelf you know and take out a sure good book? Or skip that shelf, walk to a strange one, pull a book at chance, and risk that it is bad?

You can not have both. The book you take this week is the one you take. The bot has the same choice, each step.

Most ways to learn fudge this with chance. Most of the time, exploit and pick the best known move. Now and then, explore and pick a move at whim. As the brain gets sharp, drop the chance of wild tries. The blunt knife works.

But in hard worlds—where one rare, strange move will lead to the huge score—you need a smarter blend. That is why we still look for the best way to explore.

For now, just learn that there are two pushes. The explore-exploit trade-off is in all games. A good bot feels both and chooses well.

State, Move, Score, Plan

We need some more words:

A state is how we sum up where things stand now, like what the brain sees right now: the spot in the maze, the board in a game, the words on a chat screen.
A move is what the brain does next: step left, push a pawn, write a word…
A score is the count the world hands back.
A plan is what the brain does in each kind of state.

By policy I mean a plan for what move to make in each kind of case.

Think of a policy as a rule book. One page per kind of state. Each page says, “if you see this, do that”. Some pages are firm: “do this one move”. Some are loose: “most of the time go left, but with some chance go right”.

By value I mean a guess at how much score will come from a state or a move.

Think of a value as a price tag. It may sit on a state: “this place looks good”. It may sit on a move: “this move looks good from here”. At first, the price tags are blank. The bot has no clue what each move is worth. As it plays, the price tags fill in. A high price tag means good. A low price tag means bad.

One small point on price tags. A score now is worth more than a score far down the road. The far road might not come. The bot might fall in a pit. So the price tag for a move is a sum: score now, plus some score from what comes next, plus less score from what comes past that, and so on. The more far off the score, the less it weighs. That is why a value is a guess at score, not just the score seen now.

Tags On A Board

A Grid of Q

Now we can build our first way to learn, with no big brain at all: just a black board and chalk.

By table I mean a grid of rows and cells, like a calendar or Tic-Tac-Toe.

Draw a grid on the board. Down the left side, write each spot in the maze, one per row. On the top, write each move the bot can make: left, right, up, down. In each cell, write a count: the price tag for that move from that spot. At the start, write 0 in each cell.

By Q-learning I mean a way that fills a table with value guesses through play.

Q-learning goes like this. The bot lands on a spot. It looks at the row for that spot. It scans the row, finds the cell with the best count, and walks that way. Most of the time it picks the best cell; now and then it picks at chance, just to try. It does the move, gets score, and lands on a new spot.

Then it fixes the cell it just used. Look at the new spot’s row. What is the best count in that row? Add the score to that. Now blend the cell’s old count with this new one: rub out a bit of the old, write a bit of the new in its place.

Do this for a long stretch of play, and the counts on the board fill in to the right values. Then the bot just looks at its row, picks the best cell, and plays.

By value-based I mean a way that learns values first, then picks moves from them.

Q-learning is a value-based way. It fills in price tags first. Then to act, it just picks the move with the best price tag. The rule book is built on top of the price tags.

By tabular I mean made with a table, not with a brain.

This is tabular Q-learning. There is no brain of neurons here at all. Just a board and chalk. You have built a thing that learns to play a maze with just what is in your school room! Chalk boards were not meant to think, nor win or lose; try not to brag.

When the Board Is Too Big

Tabular Q-learning works well for small worlds. For huge worlds, the board blows up.

Think of the world a chess brain sees. A board has 64 squares. Each square holds one of 13 things: 6 kinds of white piece, 6 of black, or a blank. The count of board states is vast: more than grains of sand on Earth, more than stars in the sky. The black board would have to be wide past your school, your town, and the state you live in. A chat brain has it worse: there is one state per chunk of text it might see, and there is no end to chunks of text.

So we ditch the board. In its place: a brain of neurons. The brain takes in the state, such as the board or the words, and spits out the counts as if it had seen the right row. It is a brain that guesses what the chalk would have said, had we drawn that row.

This is the deep trick: trade the board for a large net of neurons. The rule of Q-learning stays much the same. The brain stands in for the chalk. It can make a fair guess for a state it has not seen, since kin states tend to have kin values. A black board could not do that: chalk in one row does not tell you what to write in the next.

The same swap works for the rule book. A small rule book fits on a board. A chat brain’s rule book has a chance for each next word, and the word set is far too large to draw. So the rule book can be a brain of neurons too. It takes in the state and spits out a chance for each move.

The rest of this text works if you have a board or a brain of neurons. The prop shifts; the loop stays.

Three Splits

Map or No Map

By model I mean a map in the brain of how the world works.

Think of a model like this. The bot has, in its head, a small doll house of the maze. In the doll house lives a small doll bot. The real bot can poke the doll bot through the doll house and see what comes next. “If I send the doll left, then up, where does it land?” The doll house is the model. With it, the bot can plan a string of moves, then move a foot for real.

A bot with no model just plays in the real maze. No doll house. No plan in its head. It steps, gets score, and learns.

By model-free I mean a way that learns with no model. By model-based I mean a way that learns with a model.

Model-based ways can learn from less play, since the brain can plan in its doll house and squeeze more out of each run. But the doll house might be wrong. The doll house could say, “step left and you find cheese”, while the real maze says, “step left and you hit a wall.” Then the brain plans well in the doll house and fails in the real maze.

Model-free is slow but safe: no wrong doll house to fool you.

Tabular Q-learning is model-free. The chalk on the board holds price tags for moves, not a doll house of the maze.

Whose Play Counts?

Now a fine point. The brain is making moves, and it is learning. But is it learning from its own moves right now, or from old game tapes?

By on-policy I mean a way that learns from the policy it has now. By off-policy I mean a way that can learn from old play or a past policy.

If you are on-policy, you learn from fresh games. Once you change your policy, the old games may not fit. They came from the old policy, not the new one. You have to play more.

If you are off-policy, you keep a stack of old game tapes on your shelf. Some are tapes of you last week. Some are tapes of kin bots. You watch old tapes from the stack and learn from them too. This saves fresh play. But off-policy is a fair bit harder to make work right. Old play can point the new policy at the wrong thing.

Tabular Q-learning is off-policy. Why? When it fixes a cell, it blends in the best count from the next row, not the count of the cell the bot would in fact pick next. It learns the value of the best policy, not just the policy it used to get that play. That is what makes it off-policy.

We now have 3 splits to track: (1) price tag or rule book; (2) doll house or no doll house; (3)fresh games or old tapes. Each way to learn can be tagged on all 3.

Shift the Rule Book

Shake and Pick

How do we make a brain, not a board, learn? Here is one way that uses no math.

By evolution I mean change of a kind through births, shakes, wins, and deaths. By strategy I mean a plan for how to win. By evolution strategy I mean a strategy that trains a brain by small shakes and scores.

An evolution strategy goes like this. Take a brain. Make 100 twins. For each twin, give it a small shake: a few wires shift up, a few shift down, by chance. Now race all 100 brains through the maze. Each twin gets a run time or a score. Drop the, say, 90 worst. Then keep the 10 best.

From those 10, make 100 new twins. Shake each one. Race them through the maze. Drop the 90 worst. Keep the 10 best.

Do it once more. And once more…

Soon, the brains run the maze fast!

This is grim if you think too hard. But the brains do not feel a thing. This is in fact how you got here, just on a much longer time scale.

An evolution strategy is model-free: no doll house. The rule book brain is what gets a shake. It is plain on-policy: each score comes from what the brain just tried.

What is the good of an evolution strategy? It is plain. It works on huge brains. You can race all 100 twins at the same time on lots of chips, since each twin runs on its own.

What is bad? It wastes play. A twin has to run a whole maze just to get one score. One bad slip in a long run can make a great brain look bad. Worst, it is dumb to thanks and blame: it pays no heed to which moves in the run helped or hurt. The score is a big lump dropped on the brain as a whole. The next ways are sharper.

One Brain, Lots of Runs

By REINFORCE I mean a rule that makes moves in runs that beat a base more apt, and moves in runs that fall short less apt.

Set a shelf in front of you. On the shelf is a jar for each wire in the brain. The count of stones in a jar tells you how strong that wire is. A jar with lots of stones: a strong wire. A jar near dry: a weak wire.

REINFORCE goes like this. One brain plays a whole run. At each step, it picks a move with some chance. Note each move and which wires helped lead to it. At the end of the run, look at the score. If the score is high, walk down the shelf and drop stones in jars for the wires that helped make the moves. If the score is low, walk down the shelf and pluck stones out of those jars. Then play more runs and do it once more.

This is not fair to the wires in one run. A wire that helped make a smart move in a lost run still loses a stone. A wire that helped make a dumb move in a won run still gains a stone. The jar will not know why. The jar need not know. With lots of runs, the noise tends to wash out. On the whole, wires that help win pile up stones. Wires that help lose run dry. The brain gets good.

By policy-based I mean a way that changes the policy, not just values.

REINFORCE is a policy-based way. It does not fill in price tags first. It just shifts the rule book: make this move more apt, make that move less apt. Policy-based ways tend to work well on huge tasks, such as a chat brain that writes the next word, where there are far more moves than we could rate one by one. Value-based ways tend to work well on small tasks where the brain can rate each move with care.

REINFORCE is on-policy and model-free too. It is sharper than an evolution strategy, since it changes the parts that helped make the moves. But it is still loud. A whole run gives one blunt grade to lots of small moves.

A Judge on the Side

REINFORCE is loud. “High score” is a loud grade for each move. We can get a sharper grade.

By actor I mean the brain that picks moves. By critic I mean a side brain that tries to guess the value.

Set up two brains now. The actor plays the game. It has the rule book. The critic sits on the side and calls the play at the start. At each spot, the critic shouts a guess: “From here, you will score 5 treats.” The critic is graded by how close its calls come to the truth. A wrong call costs it. A right call pays it.

By advantage I mean how much a run beat the critic’s guess.

Now we change REINFORCE. The actor makes a move. The critic called 5 for that spot. The run in fact ends with a score of 8. That beat the call by 3, so the advantage is +3. Drop stones in the jars that helped make that move.

If the run ends with a score of 2, that is short of the call by 3, so the advantage is −3. Pluck stones out of those jars.

Now we are not paying wires for “we won”. We are paying them for “we did better than the critic thought”. That is a sharper grade, and the brain learns faster.

The critic is just a value brain: a brain of neurons that guesses the price tag for a state. So this trick blends policy-based and value-based in one. The actor is the rule book. The critic is the price tags.

Keep the Push Small

Small Steps

REINFORCE and the actor-critic trick have a sore spot.

When you change the brain a lot in one step, you can break it. A brain that was great might, with one big push, lurch off and lose for a long stretch. Worse, it may lose skills it once had, such as how to walk in a straight line.

By PPO I mean a rule that clips the push so a new policy stays near an old one.

Think of PPO as a short leash, but tie the leash to move chance, not to each jar. Keep a note of how apt the old policy was to pick each move it in fact picked. Train the new policy on those same moves. If the new policy tries to make a good move far more apt, PPO clips the push. If it tries to make a bad move far less apt, PPO clips that push too.

The leash is not on each wire. The leash is on move chance.

The old policy picked a move. The new policy may make that move more apt, or less apt. But PPO clips the push if the change gets too large.

So the brain learns by lots of small steps, not by big leaps. Good moves get more chance, but not too much at once. Bad moves get less chance, but not too much at once. This helps keep the brain from a wreck of what it knew.

PPO is policy-based with a critic to cut noise. It is model-free. It is close to on-policy, since it wants fresh runs from the old policy for each round.

The Group Sets the Base

PPO needs the critic brain. That is one more brain to train. It can be hard to make a critic for a chat task. You ask it, “how good will this half done note be once it is done?” The critic shrugs.

By reply I mean text sent back to a prompt. By GRPO I mean a rule that scores a group of replies and uses the group mean as the base.

Think of GRPO as 8 darts at a board. For each task, the brain throws 8 tries: 8 replies, or 8 runs, or 8 ways to solve the same thing. Each one gets a score. Find the mean of the 8 scores. That mean is the base.

Each try that beats the base gets pushed up. Each try that falls short gets pushed down. And you still keep the short leash, the PPO way.

The point: no critic brain. The group is the critic. The mean of the group is the call. Less work, less stuff to train, less to break.

GRPO is a good fit when the score is sharp. For math, did the work come out right: yes or no? For code, did the test pass: yes or no? For chat, a score brain can grade each reply, then GRPO can push up the replies that beat the group.

Slow Think and Games

Lift the Fast Gut

So far, each way to learn has been fast: pick a move and play. But there is a slow way to pick too. You can think hard, then move.

First, use no net. Use the rules of the game. Draw a tree of what-ifs on a sketch pad. At the root is the board you see now. Each branch is a move. At the next row are the boards that could come next. Then the foe moves. Then you move. The tree grows till you stop. At each leaf, write a rough price tag. Then push the best tags back up the tree and pick the branch that looks best. This is slow, but it can be smart.

Now add a fast gut. The fast gut says which moves look worth a slow think. It is the policy. The price-tag head says how good a board looks. It is the value. The slow think uses both: the gut says which branches to search first; the price tag says how good a leaf looks, so the slow think need not play each line to the end.

Here is the loop. The slow think picks a move. Train the fast gut to pick that move with less thought. Then use the new fast gut to guide the next slow think. The slow pick trains the gut. The gut makes the next slow pick better. Round and round.

By expert iteration I mean a loop in which a slow good pick trains a fast gut, and that fast gut helps the next slow pick.

Expert iteration is not a new beast. It is the same old loop in a fine suit: try a strong slow pick, score it through play, change the fast gut to match it.

You can see the old price-tag and rule-book loop in it too. Better price tags make a better rule book. The better rule book makes new play. The new play makes better price tags.

A Game Brain

By self-play I mean play where a brain plays a twin of its own self. By AlphaZero I mean a game brain that lifts its fast gut by self-play and slow think.

Think of a kid in their room with a chess board. The kid plays both sides. White makes a move. Then the kid spins the board, sits in the black seat, and makes a move for black. Spin and switch, spin and switch, all on their own. A few years of this, and the kid is a beast at chess.

That is the feel of AlphaZero. It plays both sides of board games. The raw names of the games were “chess, shogi, and Go”. It got great at all 3 with no help from folk who knew the games. It read no books on chess. It watched no tapes of grand chess play. It just learned from games it played on its own.

The rules of chess are known. If you have a board and a move, the rules tell you the next board with no doubt. So the rules of the game are a model: a true one. AlphaZero is model-based at move time, since its slow think plans with the rules.

The AlphaZero brain has two heads on one big net. One head is the fast gut: a policy that says which moves look good. One head is the price tag: a value that says who is more apt to win from this board.

At move time, AlphaZero does the slow think from the last part. It searches a tree of what-ifs. The policy head says which branches to search first. The value head grades the leaves. The slow think then picks the best move.

Then comes self-play. Two twins of AlphaZero play a game, each with its own slow think. At the end, one wins, one loses.

The game gives a heap of board-and-move pairs, with a win or loss at the end. Train both heads. The policy head learns to like the moves the slow think picked: lift the gut. The value head learns to guess the score that came at the end. Soon, the heads are sharper. The slow think is sharper too, since it leans on the heads. The next round of self-play is sharper still. That is expert iteration with a board game for its world.

No folk rank its moves. It learns from win and loss. It is both policy-based and value-based, since it learns both heads. It is close to on-policy, since it learns from fresh games made by its own self.

Folk and Chat

When Folk Pick

AlphaZero had it good: a clear win-or-loss score from each game. But what if you want a brain to write a good note? There is no clear count for “good note”. A note can be true and dull. A note can be wrong and slick. A note can be sharp on the math but mean to read. You need a kind of grade that plain code can not give.

So we ask folk.

By human I mean one of us, not a bot. By feedback I mean a mark or word from a human that says good or bad. By RLHF I mean a way to train a bot with human feedback.

Think of a blind taste test. Two cakes on a plate. A bunch of folk take a bite of each, then point to which one tastes best. Do this lots of times, with lots of cake pairs and lots of folk. You end up with a thick stack of picks: this cake beat that cake.

Now switch from cakes to chat. A human sees two replies to the same prompt and picks the one they like more. Do this lots of times. Now we have a stack of picks: this reply beat that reply.

By reward model I mean a model that gives a score to a reply.

Train a reward model on the stack of picks. Its job is to guess which reply folk would pick. Soon the reward model can stand in for the folk. The folk go home. The score brain stays.

The chat brain writes replies. The reward model rates each reply. Use PPO or GRPO to push the chat brain to write replies the reward model rates high. The PPO leash keeps the chat brain from drifting too far from the chat brain you had at the start.

There is one trap. If you let the chat brain push too hard, it can learn to game the reward model. It might write a reply that scores high but is in truth junk food: full of fluff or the same stuff, but the reward model falls for it. The leash holds it back. Do well by the score brain, but do not drift too far from the chat brain you came from.

RLHF does not make the bot feel joy or pain. Score is just a count used to change wires.

How A Chat Thing Gets Built

By deep learning I mean a way to train a large net made of neurons. By language I mean words and ways folk speak or write. By ChatGPT I mean a chat bot that reads your words and writes words back.

Think of ChatGPT as a kid sent through cook school in 3 terms.

Term 1: read all the cook books in the world. The kid reads, and reads, and reads. At the end, you can hand the kid the start of a cook book line, and the kid can guess the next word. This term does not use reinforcement learning. It is deep learning on text. The brain learns to guess words. At the end of Term 1, the kid knows a lot of language, but does not yet know how to make a good meal for the task at hand. Cook books teach words. They do not teach taste.

Term 2: watch a great chef cook 100 dishes from start to end. Take notes. Try the chef’s moves on your own. The kid takes on the chef’s style. Now if you ask the kid to cook soup, the kid cooks soup that looks like the chef’s soup. But the kid is still not great. Some food comes out flat. Some comes out off. Some is not safe to eat.

Term 3: cook for guests. Put two plates of soup down at the same time: two takes on the same dish. Watch which plate the guests scrape clean. Train a reward model to guess which plate the guests would pick. Then use PPO or GRPO to shift the kid’s style toward plates the reward model rates high. That is RLHF, folded back in.

That is the plan. Term 1: read the world’s cook books. Term 2: watch a chef. Term 3: cook for guests, and take their picks to heart.

Most of ChatGPT’s skill comes from Term 1, from just text. The score loop in Term 3 is more like a last tune. It shapes the raw text brain toward help, truth, care, and a good chat style.

The Map and the Lens

What to Keep

You now know the core parts: state, move, score, policy, and value.

You know the core hard task: which past move gets thanks or blame?

You know 3 splits:

Value-based or policy-based: price tags or rule book.
Model-free or model-based: no doll house or doll house.
On-policy or off-policy: fresh games or old tapes.

Here is a map of reinforcement learning we have seen:

Each way is tagged by what it learns, if it has a map, and if it learns from fresh play or old tapes.
way	rule book or tags	map?	fresh or tapes
Q-learning	tags	no	tapes
evolution strategy	rule book	no	fresh
REINFORCE	rule book	no	fresh
actor-critic	both	no	fresh
PPO	both	no	fresh, more or less
GRPO	rule book	no	fresh, more or less
AlphaZero	both	yes	fresh
RLHF	both	no	fresh, more or less

Q-learning fills a table of values. An evolution strategy shakes whole brains and keeps the best. REINFORCE makes moves in high score runs more apt. The actor-critic trick pays for advantage, not raw score. PPO clips the push, so the new policy stays near the old one. GRPO uses a group mean in place of a critic. AlphaZero lifts a fast gut by self-play, tree search, and expert iteration. RLHF uses human feedback to train a reward model, then trains the chat policy.

The Shape of It

These 8 ways are not all there is. Folk who do this for a job have far more, with names and tricks this text did not touch. There is a long list of hard parts I have not shown you: how to learn when score is rare, how to plan with no map, how to share play with kin brains, how to fail safe.

But the heart of it is the same 3 words:

Try
Score
Change

This is one way to look at all life.

For as long as there has been life, this loop has run. The first cells did this loop. So did the fish, the lungs, the bones, the hands, the brains. Each kind of life was a try. The world gave back a score: did the kind live and breed, or not? The kinds that won made more of their own kind. The kinds that lost did not. Slow, slow, slow—but that loop built each life on Earth, you and me with the rest.

Brains do this loop too, on a much faster clock. A child learns to walk in a year. A dog learns to fetch in a day. A chat brain learns a new style in hours. The same 3 words, run on a chip in place of a cell.

Once you see the loop, you start to see it all through life. A child stands, falls, gets up, and stands more. A bird hears kin sing, tries its own song, hears it fall short, and tries once more. A kid does math, gets praise, does more math, and gets good. A shop posts a price, folk buy or pass, and the price shifts. A tale is told, some folk tell it on, and some let it die. A word catches on, or fades. A law works, or it breaks, and the law is changed.

The score need not be wise. It can be false, thin, or cruel. That is why we must choose scores with care. A bad score can make a bad world. A good score can grow skill, grace, and trust.

A raw old line says “endless forms most beautiful”. That line fits here. From blind tries and hard scores came fins, wings, eyes, songs, hands, tools, jokes, proofs, games, and minds. From fast loops on chips come nets that can read, spell, search, and chat.

No one planned all of it at the start. No one held the full map. There was just the loop: try, score, change; try, score, change; try, score, change.

That is why this field is more than a set of tricks. It is a lens. Look through it, and the world looks less like a fixed thing, and more like a vast work in flight: forms tried, forms scored, forms changed, and new forms born from old.

Colophon

To write “Reinforcement Learning for Children”, I defined a detailed “Grow-Speech” specification for LLM writing, and then combined that prompt with a prompt describing the pedagogical goal and (to help prime the cognitive pump) the Table of Contents from Sutton & Barto 2018.

Then I generated rough drafts in 5 frontier LLMs, and combined the prototypes in GPT-5.5 Pro and Claude-4.7-opus, and did some light critique and iteration to create the final draft, which I lightly copyedited.

The prompt and several candidate generations can be viewed at 2026-05-09-gwern-reinforcementlearningforgradeschoolers.txt.

Bibliography

[Bibliography of links/references used in page]

The City of Counted Stars

Gwern Branwen — Thu, 23 Apr 2026 00:00:00 GMT

A traveler reports a city near Samarkand whose laws, professions, walls, and funerals are ordered by the seven classical planets. Each citizen receives a horoscope at birth, but the figure is sealed until death, when it is read aloud as biography rather than advice.

The city’s doctrine is that a prediction remains a nativity only while unused. Foreknowledge would not merely inform the life; it would enter the life, changing the object it claimed to describe.

The result is a compatibilist fable in astrological costume: the stars determine, the citizens deliberate, and the mourners forgive lives for having happened only once.

…“Sire, I shun the reckonings of the Bureau of Heaven.”

And he recounted:

Three days’ ride south-southwest from the turquoise towers of Samarkand, you find a city called by its poets, The City of Counted Stars.

Past the felt-tent nomads who boil horseflesh, after the last mulberry canal has sunk into salt, it is near the country where fire is tended in domed temples. Its tax register is 70 qing of cultivated ground, watered by a kariz from the eastern mountains, growing principally barley, wheat, and a dry rice that ripens in their lower gardens, where it is cooled by the desert winds at night. It is prosperous and peaceful, with due taxes paid each year, and public executions occur so rarely that children are lifted to see the scaffold.

The people are skillful in casting nativities. The city partitions its populace strictly by planetary dominion. The Sun holds the mint and governor’s court. The Moon has the midwives, water-sellers, laundries, and those who record dreams. In Saturn’s quarter are the masons, gravediggers, well-cleaners, foreigners, and washers of the stillborn. Jupiter keeps the judges, treasurers, old men who remember droughts, and keepers of bronze weights. Mars has the butchers, surgeons, armorers, and the square where blood may be spilled. Venus has the madder-dyers, singers, bath-women, and brides. Mercury has the interpreters, money-changers, thieves, and scribes.

The customs of the quarters are these. In Saturn’s quarter no house may be repaired until a stone has fallen. In Jupiter’s quarter a man may swear the same oath no more than three times; thereafter it binds no one, not even the swearer. In Mars’s quarter knives may be sharpened openly, but never hidden. In Venus’s quarter debts contracted by song are not recoverable. In Mercury’s quarter a lie told elegantly is fined less than a truth told for profit. There are also, about the outer wall, 12 gates, 28 wells, 36 towers, 60 boundary-stones, and 360 mulberries, which I did not look for; none but the astrologers know how these numbers answer to the heavens.

Only failed astrologers may sit in the market of Mercury, offering auspicious hours for journeys, lucky names, lawsuits, reconciliations, haircuts, and cures for barren goats. A man who correctly forecasts three births is then barred.

So devoted is its populace that a horoscope is cast at birth, and the figure sealed in wax. At the funeral, when the body has been washed in water drawn from the Saturn-quarter wells and turned toward the west, the alchocoden breaks the seal and reads aloud. Then the mourners marvel, or lower their faces, hearing of illnesses, journeys, inheritances, adulteries, virtues, and stupidities past correction. A miser is found to have had Jupiter retrograde in the second house, afflicted from the eighth by a Saturn in fall; a singer had Venus in her Taurus domicile, setting and combust, joined to the Pleiades, her Lot of Spirit fallen on the house of enemies; a soldier who fled three battles had Mars exalted but cadent, his twenty-third profection to the twelfth house, where courage is buried with its dead.

If the traveler asks why the horoscope was not read while the man lived, the people answer patiently, as if explaining weirs to a fish. “How could it remain his nativity”, they say, “had he used it as advice?” Thus they forgive lives for happening only once.

Atop the western gate is written:
The stars obey; therefore they are free.

Against this sentence the bonzes of Khotan and the ashen mendicants make subtle arguments. The citizens listen courteously, assign them to the quarter of Mercury, and go on as before.

“The hour of my birth”, said the Khan, “was lost.”

Colophon

“The City of Counted Stars” was written with LLM assistance in my standard workflow of cognitive priming + brainstorm + iterative editing + detailed formalization/commentary of a final version; see the past writeups in “This Last Pain” Graveyard, and “Elegy in a Craneyard” § Colophon.

The process is to come up with a premise; include some background/reference material to “prime” the LLM; generate variations and ideas en masse; curate and iterate; and then do a detailed semi-formal annotation/summarization/gloss to format it for Gwern.net, document the goals, and make the text more dense (avoiding the characteristic chatbot “emptiness”). Similarly to “October”, a major writing goal was “hyperdense” writing, where every word in every sentence performs a function which adds esoteric depth but is not strictly required for understanding the exoteric story (allusions as a kind of worldbuilding to add “mythic depth”). I used the same process as “Craneyard” to generate the illustration image; in this case, the brand-new “ChatGPT Images-2.0” (in-progress samples) beat out Nano Banana Pro (in part due to NBP obsessing on a particular scene setting and ignoring the compiled 10 prompt ideas).

As the process this time was routine with no new prompt engineering tricks and I wrote an unusually detailed draft story so the LLMs served mostly as knowledge bases about astrology and critical feedback, I will not document it beyond the final Markdown source and showing the starting prompt:

[Initial prompt and source excerpts used for the LLM-assisted drafting process.]

I was talking to an Uzbek from Tashkent who was using an LLM astrology prompt and amazed at how accurate it was about the past year. I asked what it predicted for the next year; he said, “oh, I don’t want to jinx it.” I wondered what use astrology could possibly be if it made valid predictions but only if you didn’t know them; isn’t the point to make decisions based on the forecasts? This made me think about possible uses, and perhaps there are in a compatibilist sense. One could treat this as a parable. Italo Calvino’s Invisible Cities came to mind.

A quick stab at such a story:

Three days horseback south-southwest travel from Samarkand you find a city called by its poets, The City of Uncounted Stars, almost bordering the land of the fire cultists, and past the nomads who dwell in felt tents and boil horseflesh. It has 7,000 [insert medieval Chinese area unit] cultivated, growing principally barley, wheat, and dry rice. The climate is temperate, and cooled by the desert winds at night. It is prosperous and peaceful, and public executions rarely occur nor have its taxes been remitted in living memory. It is renowned for its astrologers, who are credited by the populace to the extent of dividing the city into 7 quarters, each with its peculiar priesthood. The priesthoods remain aloof, permitting only their failures to sell horoscopes in the marketplaces. So devoted is the populace that by long-standing custom, a horoscope is cast for each person. Peculiarly, the horoscope is kept secret until their funeral, at which point the responsible astrologer or his successor will read it aloud, and the gathered will marvel or sorrow at the accuracy of the prophecy. If asked about this custom, they regard the traveler with confusion or condescension, asking how a man’s horoscope could remain accurate if any but an astrologer knew its contents? Thus, the people of this city conduct their daily business believing that everything they do has already been foreseen. They credit their morality and public order to this belief, that nothing a man does in vain, for he strives to achieve a horoscope that will not shame his descendants, in the knowledge that if his life goes awry, the fault was not in him but his stars. Of the arithmetical regularity of the stars and planets, they say only that since the beginning, they have chosen to obey the laws above, as men choose to obey their laws below, and both are equally free. Against this, the bonzes and wandering mendicants’ arguments can find no purchase to disturb their tranquility.

Please analyze and critique it, and suggest improvements. I would like to make it deeper and more allusive, and embed more astrological knowledge. Consider alternative interpretations.

After the suggestions, brainstorm at least 20 variant stories; summarize and describe them, with at least 5 authors/stories/genres listed for inspiration, and write them out. Critique them, and revise them. Write out the final version of each of the 20. Then rate them all 1–5 stars.

Select the best lines, images, tropes, or ideas from all versions, including mine.

After this writing process, select the top 5, emphasizing diversity, and reflect on what is working and what is not working.

Example Marco Polo passages:

CHAPTER LX. Concerning the Kaan’s Palace of Chagannor.

At the end of those three days you find a city called Chagan Nor [which is as much as to say White Pool], at which there is a great Palace of the Grand Kaan’s;¹ and he likes much to reside there on account of the Lakes and Rivers in the neighbourhood, which are the haunt of swans² and of a great variety of other birds. The adjoining plains too abound with cranes, partridges, pheasants, and other game birds, so that the Emperor takes all the more delight in staying there, in order to go a-hawking with his gerfalcons and other falcons, a sport of which he is very fond.³

There are five different kinds of cranes found in those tracts, as I shall tell you. First, there is one which is very big, and all over as black as a crow; the second kind again is all white, and is the biggest of all; its wings are really beautiful, for they are adorned with round eyes like those of a peacock, but of a resplendent golden color, whilst the head is red and black on a white ground. The third kind is the same as ours. The fourth is a small kind, having at the ears beautiful long pendent feathers of red and black. The fifth kind is grey all over and of great size, with a handsome head, red and black.⁴

Near this city there is a valley in which the Emperor has had several little houses erected in which he keeps in mew a huge number of cators, which are what we call the Great Partridge. You would be astonished to see what a quantity there are, with men to take charge of them. So whenever the Kaan visits the place he is furnished with as many as he wants.⁵

CHAPTER XVII. Concerning the Kingdom of Kerman.

Kerman is a kingdom which is also properly in Persia, and formerly it had a hereditary prince. Since the Tartars conquered the country the rule is no longer hereditary, but the Tartar sends to administer whatever lord he pleases.¹ In this kingdom are produced the stones called turquoises in great abundance; they are found in the mountains, where they are extracted from the rocks.² There are also plenty of veins of steel and Ondanique.³ The people are very skilful in making harness of war; their saddles, bridles, spurs, swords, bows, quivers, and arms of every kind, are very well made indeed according to the fashion of those parts. The ladies of the country and their daughters also produce exquisite needlework in the embroidery of silk stuffs in different colors, with figures of beasts and birds, trees and flowers, and a variety of other patterns. They work hangings for the use of noblemen so deftly that they are marvels to see, as well as cushions, pillows quilts, and all sorts of things.⁴

In the mountains of Kerman are found the best falcons in the world. They are inferior in size to the Peregrine, red on the breast, under the neck, and between the thighs; their flight so swift that no bird can escape them.⁵

On quitting the city you ride on for seven days, always finding towns, villages, and handsome dwelling-houses, so that it is very pleasant travelling; and there is excellent sport also to be had by the way in hunting and hawking. When you have ridden those seven days over a plain country, you come to a great mountain; and when you have got to the top of the pass you find a great descent which occupies some two days to go down. All along you find a variety and abundance of fruits; and in former days there were plenty of inhabited places on the road, but now there are none; and you meet with only a few people looking after their cattle at pasture. From the city of Kerman to this descent the cold in winter is so great that you can scarcely abide it, even with a great quantity of clothing.⁶

Example Hyecho passages:

West India

From south India I walked North for two months and arrived at the city where the king of west India resides. The west Indian king also possesses five to six hundred elephants. The products of this land are cotton cloth, silver, elephants, horses, sheep, and cows. Barley, wheat, and various kinds of beans are produced in large quantities [but the production of] rice and corn is much less. Food is mainly bread, wheat preparations, curds, butter, and ghee. In the markets, silver money and cotton cloth are used. The king, the chiefs, and the common people highly revere and believe in the Three Jewels. There are many monasteries and monks. Both Mahayana and Hinayana are practised. The land is very broad. To the west it reaches the Western sea. The people of this country are good at singing. [In this] the other four regions of India cannot be compared with this country. Here there is no cangue, beating, prison, capital punishment, and similar affairs. At present half the country has been invaded by the Arabs and is already ravaged. According to the law of the five regions of India, those who travel do not carry food. They can beg for food wherever they go. When the king and the chiefs travel, they take food with them so as not to bother the people.

North India (Jalandhara)

For over three months I travelled north from west India and arrived at the country of north India, which is called Jalandhara. The king of this country has three hundred elephants. He resides in a city built on the hillside. From the hills the place gradually stretches northwards. As a country it is very narrow. The army and cavalry are small. It is often annexed by the central Indian king and Kashmir. This is the reason the king resides on the hillside. The customs, dress, and language are not different from those of central India. The land is slightly cooler than central India and the other regions. There is also no frost or snow, but there is a cold wind. The products of this land are elephants, cotton cloth, rice, and wheat. Mules and asses are few. The king possesses a hundred horses, while each chief has three to five, and the common people have none. The western part of the country :s a plain and the eastern part is close to the snowy mountains. There are many monasteries and monks in this country. Both Mahayana and Hinayana are practised.

Suvarnagotra

A small kingdom is to the east, a month’s journey through the snowy mountains. It is called Suvarnagotra. and is under the control of the Tibetans. The dress is similar to that of north India, but the language is different. The land is extremely cold.

Italo Calvino, Invisible Cities (“Thin cities”) example:

Isaura, city of the thousand wells, is said to rise over a deep, subterranean lake. On all sides, wherever the inhabitants dig long vertical holes in the ground, they succeed in drawing up water, as far as the city extends, and no farther. Its green border repeats the dark outline of the buried lake; an invisible landscape conditions the visible one; everything that moves in the sunlight is driven by the lapping wave enclosed beneath the rock’s calcareous sky. Consequently two forms of religion exist in Isaura. The city’s gods, according to some people, live in the depths, in the black lake that feeds the underground streams. According to others, the gods live in the buckets that rise, suspended from a cable, as they appear over the edge of the wells, in the revolving pulleys, in the windlasses of the norias, in the pump handles, in the blades of the windmills that draw the water up from the drillings, in the trestles that support the twisting probes, in the reservoirs perched on stilts over the roofs, in the slender arches of the aqueducts, in all the columns of water, the vertical pipes, the plungers, the drains, all the way up to the weathercocks that surmount the airy scaffoldings of Isaura, a city that moves entirely upward.

[List of finalist ideas in the LLM-assisted image illustration generation process.]

GPT-5.4 Pro Ideas

1. The Corpse’s Shadow Is the Horoscope

Description

A washed, shrouded corpse lies on a black stone slab in a severe funerary chamber.

There is no visible sky, no stars, no parchment, no scroll.

The body’s shadow is not a human outline: it is a perfect natal chart projected onto the floor.

The mourners stand just outside the circle, unable to enter it.

The image says: the chart was always there, but it becomes legible only when the life is over.

This is the most iconic thumbnail because it compresses the whole story into one glance: death reveals the pattern; life could not be allowed to see it.

Prompt

High-contrast literary book-cover illustration: a white-shrouded corpse on a black stone funeral slab casts a precise circular natal chart as its shadow on the floor; three faceless mourners stand at the edge of the diagram, unable to step into it; austere Central Asian funerary chamber, cold blue stone, ivory cloth, black ink geometry, thin crimson thread accents; solemn, philosophical, uncanny; no visible sky, no stars, no parchment, no scrolls, no wax seals, no yellow or sepia palette.

2. The Map That Changes When Touched

Description

A traveler tries to use a natal chart as a map.

The moment his foot touches it, the geometry beneath him begins to twist, erase, and redraw itself.

The untouched parts remain crisp.

This image captures the draft’s central self-reference problem: a prediction, once used, is no longer a prediction of the original life.

It has entered the life.

This is the best “philosophical” thumbnail: less elegiac than the corpse-shadow image, more intellectually dangerous.

Prompt

Surreal overhead city scene: a lone traveler steps onto a circular astrological diagram used as a street map; beneath his foot, the diagram’s lines twist, buckle, and erase themselves, while the untouched streets remain precise and beautiful; the geometry is half city plan, half natal chart, but not a zodiac cliché; black stone streets, blue-white tile, ivory figure, sharp crimson rupture lines; austere Silk Road city atmosphere; no stars, no scrolls, no wax seals, no parchment, no yellow or brown palette.

3. The City As a Locked Brain

Description

An aerial view of the city forms the shape of a human brain.

Its seven planetary quarters appear as districts, not symbols.

Roads resemble neural pathways.

At the center, where memory or self-knowledge should be, there is a locked black chamber.

This is the best concept for linking astrology to philosophy of mind.

The city is not merely a place with beliefs; it is a mind whose crucial self-description has been deliberately sealed away.

The citizens are free because the deterministic description is withheld from consciousness.

Prompt

Editorial literary thumbnail: an aerial map of a walled Silk Road city subtly shaped like a human brain; seven radial districts divide the city, each suggested by civic objects—well, court, butcher square, mint, bathhouse, market, laundry—rather than cartoon planet symbols; roads resemble neural pathways; at the center sits a small locked black chamber replacing memory or self-knowledge; cold lapis, white plaster, black linework, restrained crimson accents; no sky, no stars, no zodiac wheel cliché, no scrolls, no wax seals, no sepia.

4. The Horoscope As a Mask Put On Only After Death

Description

A shrouded face receives a rigid black geometric mask after death.

The mask resembles a natal chart, but it aligns with the face like a second identity.

The hands placing it are careful, ceremonial, impersonal.

This image is clean, close-up, and thumbnail-friendly.

It says that the horoscope is not a fortune-telling device.

It is the completed person’s final face.

The risk is that it may imply falsification—identity imposed after death—so the prompt should stress alignment, not concealment.

Prompt

Close-up funerary cover image: a shrouded dead face receives a delicate black geometric mask shaped like a natal chart; gloved hands place it gently over the face after death, and the mask aligns perfectly with the features, as if it had always belonged there; ivory burial cloth, cold blue shadows, black diagram lines, thin crimson ties; solemn, precise, ritualistic, not horror; no stars, no parchment, no scrolls, no wax seals, no gold, no sepia.

5. The Baby and the Corpse Share One Face

Description

A newborn and a corpse are vertically aligned into one continuous figure.

The infant is swaddled at the bottom.

The shrouded corpse lies above.

They share the same face, or the same faint outline.

Between them floats a sealed black diagram, like a forbidden organ or suspended verdict.

This is the best life-span compression.

It makes birth and death one ritual circuit: the chart is cast at the beginning, but belongs to the end.

The concept is slightly more symbolic than #1, but it may be more emotionally immediate.

Prompt

Surreal medieval-modern literary illustration: a swaddled newborn at the bottom of the image and a white-shrouded corpse at the top share the same faint face, forming one continuous life from birth to death; between them floats a locked black astrological diagram like a forbidden organ, unreadable but exact; austere Silk Road interior, cold indigo shadows, bone-white cloth, black geometry, restrained crimson thread; philosophical, elegiac, uncanny; no stars, no scrolls, no wax seals, no parchment texture, no yellow or brown tones.

Gemini Ideas

The Shattered Nativity

An extreme close-up shot captures the hands of an Alchocoden dressed in deep indigo robes. The hands violently snap a heavy clay tablet heavily inlaid with complex silver geometric lines. Glowing silver dust and jagged ceramic shards burst outward from the sharp break. The fragments scatter across the dark obsidian-black silk shroud covering a corpse directly below. Sharp, dramatic chiaroscuro lighting highlights the metallic debris catching the cold light against the pitch-dark fabric.

The Khotan Bonze’s Erasure

A high-angle wide shot reveals a grand floor completely covered in intricate lapis lazuli astrological mosaics. A Khotanese monk wearing stark white robes sits cross-legged in the geometric center of the room. Thick grey ash heavily coats his face and bare arms. He sweeps a thick layer of luminescent silver dust using a small, stiff-bristled brush. The sweeping motion forms a perfectly smooth, blank circle of emptiness directly over the rigid celestial grid. A single shaft of cool azure light falls exactly on the blank silver circle.

The Mars Square

A wide architectural shot shows a perfectly square plaza paved entirely in highly polished crimson marble. Sheer, towering walls of heavy, rivet-studded black iron completely enclose the space. In the center, heavily muscled armorers wearing thick obsidian-scaled aprons stand over wet grey whetstones. They press massive iron swords against the heavy grinding wheels. Brilliant white sparks erupt from the friction, reflecting violently off the mirror-like crimson stone under a dark, overcast sky.

The Saturn Repair

A medium shot focuses on a massive, deeply weathered masonry wall built entirely of dark obsidian blocks. A single gaping void exists in the center of the wall where an old block has entirely crumbled away. An elderly mason wearing tattered, wind-blown azure robes stands before the dark ruin. He carefully slides a perfectly cut, gleaming silver metallic block into the empty space. Cold, diffuse moonlight illuminates the severe visual contrast between the ancient textured volcanic glass and the pristine silver repair.

The Cure for the Barren Goat

A bustling Silk Road bazaar sits beneath a canopy of sharp, intersecting azure canvas awnings. A intensely focused merchant dressed in vibrant jade silks leans aggressively over a heavy iron table. He wields a large, gleaming silver geometric caliper. He uses the instrument to precisely measure the skull of a living, highly agitated goat. The goat stares directly out of the frame with horizontal rectangular pupils. Cool daylight casts deep blue shadows across the merchant’s intricate silver jewelry and the goat’s dark grey fleece.

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Bibliography

[Bibliography of links/references used in page]

Elegy in a Craneyard

Gwern Branwen — Sun, 05 Apr 2026 00:00:00 GMT

A posthuman traveler finds the container cranes of the Port of Oakland still standing perhaps centuries hence, reflecting on their work, history, and abandonment. The cranes’ silhouettes in fog resemble Brontosaurus silhouettes; the poem traces the recursive material chain (bone → oil → steel → the shape of bone) to literalize the “a trace of the true self exists in the false self” meme.

Specifically: The poem is a 6-part mixed-form elegy in which a posthuman speaker encounters the Oakland cranes as ruins, remembers their labor, absorbs their history, and finally blesses their transformation into what the fog had always suggested: dinosaurs returned in steel. §A treats them as Old English giant-work and damaged manuscript. §B gives the Gray-like meditative arrival. §C turns the three-minute container cycle into industrial liturgy. §D preserves bureaucratic terminal language as a fossil layer and folds in the playground irony. §E puts the cranes in the limbo “between the ruin and the monument”. §F makes the dinosaur/oil/steel recursion explicit and ends in benediction: stay, wade, become your myth.

Poem (elegy/encomium, mixed meter: Old English alliterative verse → blank verse → free verse, ~70 lines). Formal models: Thomas Gray’s “Elegy Written in a Country Churchyard” and the anonymous Old English “The Ruin”. Written April 2026 with extensive LLM collaboration (primarily Claude, GPT, & Kimi).

Sequel to “Silver Bird Above San Francisco”.

Types: OE = Old English alliterative (stress count + alliteration pattern) IP = iambic pentameter / blank verse target FREE = free verse (syllable count tracked, meter intentionally loosened) DOC = documentary prose (no meter tracked) LACUNA = damaged-text fragment Stress convention: - Stressed syllables in CAPS. - Unstressed syllables in lower-case. - Hyphens mark internal syllable breaks. - Syllables numbered sequentially per line. Pronunciation commitments (for stability): - articulate = 4 syll (ar-TIC-u-late) - circuit = 2 syll (CIR-cuit) - coyotes = 3 syll (coy-O-tes) - damaged = 2 syll (DAM-aged) - every = 2 syll (EV-ry) - fired = 1 syll (FIRED) - iron = 2 syll (I-ron) - language = 2 syll (LANG-uage) - loneliest = 3 syll (LONE-li-est) - Middle Harbor = 4 syll (MID-dle HAR-bor) - miniatures = 3 syll (MIN-ia-tures) - Portainer = 3 syll (por-TAIN-er) - prayer = 2 syll (PRAY-er) - professional = 3 syll (pro-FESH-nal) - sauropod = 3 syll (SAU-ro-pod) - theorems = 2 syll (THEOR-ems) - translated = 3 syll (trans-LAT-ed) -->

These crane-corpses, wondrous— || calamity crumpled them,
these gantry-works gutted, || great craft gone to rust.
Salt at the joints || has sapped the storm-scarred steel—
sagged and gaping. || A grave-grip holds
the master-builders, || machinists and makers,
in the ground’s hard grasp, || gone past all naming.
Yet still these cranes stand, || stiff against storm—
held.

[ . . . . . . . ] wire-rope [ . . . ] frayed [ . . . ]
[ . . . . ] spreader-bar [ . . . . . . ] crusted with salt—

Þ︎

The fog-bell tolls the knell of parting trade.
500 years have passed, or 5, or none—
time crumples when there is no one to count—
and I have come where Middle Harbor dims,

where cranes still stand, articulate, in fog
like theorems in a language no one speaks.
The bollards hold no lines. The spreader hangs,
a jaw still gripping air where cargo hung.

⊥︎

And what a verb they were—the turning wheel
of the three-minute cycle: hook, lift, swing,
lower, lock. Again, again. The prayer
of a machine that does not know it prays,
that does not know the box it lifts is all
the world: the grain, the circuit board, the steel,
the parts some other shore would shape and build,
would box and send back on the turning tide.
The sacrament was twenty feet of steel,
sealed by the twistlock, raised above the deck—
a miracle performed a thousand times
a day by something that could not believe.

⚙︎

ASSIGNED AREA: Outer Harbor.
BERTH WINDOW: expired.

The fog inherits every vacant clause.
The break-bulk world collapsed into a box.
The machines got work. The men got scrapped.
They built a park inside the port. Small cranes,
bright-painted miniatures, and children climbed
while real booms towered just beyond the fence.
The children played at cranes. The cranes played back
at dinosaurs. The joke outlived the crowd.

THREE-MINUTE CYCLE: discontinued.

Vessel.

■︎

In ’89, the Cypress fell in tiers.
The cranes rocked on their rails and did not fall—
never damaged enough to justify
removal, never loved enough to save.
Between the ruin and the monument:
the loneliest address a structure knows.
Somewhere in the hills a traffic signal
still changes color for the coyotes.

Vessel.

〰︎

O iron waders! In the fog you wear
your true shape—what always lay beneath

the paint, the specs, the welds, the Portainer:
the sauropod. The Brontosaurus. Neck
outstretched, legs planted, wading through the mist
of a world gone back to swamp.

Dead bone to oil, oil into fire, fire to steel,
and steel to the shape of bone again.

You are the last employees of the port,
still facing water with professional grace,
and you have not been told that you are fired.
The waiting was the cargo all along.

Vessel.

Stay, too beautiful for salvage, stay,
and let the fog make dinosaurs of you
each morning, and the Bay rise at your feet
until you wade, at last become your myth—

old thunder-lizards, one by one, made real,
out through the fog that first translated thee.

Colophon

“Craneyard” was written in April 2026 to demonstrate my current AI poetry scaffolding workflow; this documents that workflow, to help explain to my fellow LLM users how decent creative LLM writing works.

This instance in particular shows the benefits of grounding a poem idea in multiple allusions with background research notes.

(Many LLM outputs throughout this process can be seen in the “Elegy in a Craneyard” graveyard notes.)

It emphasizes high concepts to create “friction”/non-default outputs, priming to enrich the available options, and then search/iteration to overcome the “view from nowhere” and “neural nets are lazy” problems that make almost all chatbot-LLM-written poetry unsatisfying (and so strikingly impoverished compared to even small ancient “base models” like GPT-3 was); see also “At a Red Light” writing experiment, “Bits Beat AI Slop”, “Apollonian #1” vs “Spoilage”, etc.

The high concept is ‘the Seaport of Oakland’s iconic cranes meets the Old English poem “The Ruin” & Thomas Gray’s “Elegy Written in a Country Churchyard”, as backed up by LLM-powered “deep research”’ (to prime a large pool of concepts).

Genesis

This concept came to me while driving back one night in March 2026 from a dinner party at S— K—’s and watching the cranes yet again. I had often thought to myself that they looked like dinosaurs in the fog (although not Star Wars AT-AT vehicles—a popular comparison but which turns out to be spurious), and wanted to write a poem on them, or perhaps generate an illustration, but there was not yet enough there there to write about. (Yes, they do look like Brontosaurus or something… but then what? Just the bare comparison would barely support a haiku.) It was a powerful image, but I needed more if it was to make sense either standalone or as some sort of followup to the “Silver Bird” encomium to modern aerospace.

Then it suddenly hit me: “Elegy in a Churchyard” → “Elegy in a Craneyard” + “The Ruin” = good AI poem topic! And then I further recalled the Internet meme, “And Yet a Trace of the True Self Exists in the False Self / Circle of Life”, and it occurred to me that the ships and cranes were likely powered by fossil fuels which we would like to phase out, and so in both a literal and figurative sense, were dinosaurs. This was a rich enough stew to work with, especially with the ending: the cranes vanishing into the mists of time, but translated into poetry, like the Roman ruins or the deceased in the church graveyard—not useless or fit only for dismantling, nor lost entirely, as long as there are witnesses to reflect on them and recall their glory days. (And now that I have two poems, I can hope to find a third topic; perhaps something about datacenters like OpenAI’s Arizona Stargate facility, which are the cathedrals of our era, thereby creating a trilogy of past / present / future—although that poem awaits its collision with some other idea, to give it depth and motivate a choice of formal structure.)

Research & Priming

Like “Silver Bird”, I wanted to ground it in the history of the Bay Area and the port, so I began by running “deep research” or web searches in my standard LLMs, with a prompt like

I would like to write a poem on the history of the Bay Area Oakland Seaport, focusing especially on the container era and containerization and the cranes. The WP article is too brief. I would like a more detailed research report, focusing on names and technical terminology: everyone who was involved in making or running it, names of interesting technologies or ships, etc. Any word which could help prime a poem. once done and written up, condense down to a newline-delimited list of names/terms.

I merged the outputs to get my final ‘databank’:

100 decibels
12 containers high
1408 Middle Harbor Road
1,600 tons
1934 West Coast Waterfront Strike
1950 Franklin Street
1956
1958
...

The Scaffolding Workflow

I then wrote out a description of the idea, and the 2 poems and meme I wanted to incorporate, and defined my usual search process:

…To write this poem, brainstorm poem candidates. Write 20 poem candidates. To do so, for each candidate:

think up an interesting poem meter or formal structure.

list relevant poets, poems, schools of esthetics, or other inspirations.

define the formal structure of the poem in lines, meter, rhyme scheme, length, etc.

write out a summary of the poem, first as an abstract, and then stanza by stanza

write out some sample lines from the hypothetical candidate.

revise and critique the approach. After that,

write out the entire candidate poem.

revise and critique the candidate poem at least 3 times, starting with an analysis of how well it works globally, and down to the line by line level.

write out the final candidate poem.

analyze and evaluate; rate it 1–5 stars.

excerpt the best lines, images, phrases, and tricks from it.

Below is additional material on the topic to help inspire you. After reading all reference materials below, begin the brainstorming process…

I included the databank, the text of the two poems, meme image and a few small thumbnails of the Oakland cranes for reference. (See the log file for the full prompts.)

Example output from the first Claude session:

#1: CANDIDATE 20—Composite/Mixed Form (★★★★★)

Meter & Formal Properties

Five-section poem using different forms per section: (I) Old English alliterative half-lines with caesura, 2+ stresses per half-line, alliteration across the caesura; (II) unrhymed iambic pent ameter (blank verse); (III) prose; (IV) free-verse ode with apostrophe; (V) incantatory free-verse coda. No fixed rhyme scheme. ~85 lines total. The formal shifts mark emotional movements—musc ular physicality → intellectual depth → historical density → lyric intensity → ritual closing.

Plot

A posthuman traveler arrives at the ruined Port of Oakland. Section I: they encounter the cranes in fog at dawn, described in Anglo-Saxon register as “crane-corpses”. Section II: meditation on what the cranes did when operational—the three-minute hook cycle as liturgy, the container as sacrament. Section III: prose history of containerization (McLean, the Ideal X, PACECO, the M&M Agreement), ending with the Middle Harbor playground conceit—children playing at being cranes while cranes play at being dinosaurs while dinosaurs play at being oil. Section IV: the speaker addresses the cranes directly, recognizing the dinosaur shape beneath the machine shape, the carbon genealogy (bone → oil → steel → crane-shape = bone-shape), and delivers the poem’s central image: “You are the last employees of the port, / and you have not been told that you are fired.” Section V: coda states the “true self / false self” thesis, then commands the cranes to stay and await submersion as the Bay rises.

Inspirations & Allusions

“The Ruin” (Exeter Book): Section I’s opening (“these crane-corpses, wondrous, calamities crumpling”) directly echoes and recontextualizes the OE poem’s opening

T. S. Eliot, The Waste Land: the collage/multi-form structure, juxtaposition of registers

Hart Crane, The Bridge: the ode-address to infrastructure (Section IV’s apostrophe “O iron waders”)

Geoffrey Hill, Mercian Hymns: prose blocks as poetry, industrial-historical overlay

Thomas Gray, “Elegy Written in a Country Churchyard”: the elegiac frame (posthuman visitor mourning the dead), the pastoral-silence opening

“And Yet a Trace of the True Self Exists in the False Self” meme: the dinosaur → oil → plastic → toy dinosaur cycle, repurposed as dinosaur → oil → steel → crane-dinosaur

Basil Bunting, Briggflatts: sonata-form applied to poetry (each section as a movement)

The sacramental language in Section II (“motion’s priests”, “sacrament”, “levitation”, “miracle”) echoes Hopkins’s industrial-sacred vocabulary

Best Lines & Phrases

“the rhythm of a prayer / said by a machine that does not know / it’s praying”

“twenty feet of anything, sealed and blessed / by the twistlock’s grip”

“a kind / of levitation, if you squint, a kind / of miracle performed a thousand times / a day by something that did not believe”

“The word is containerization. It does not scan.”

“The arrangement had the structural elegance of a suicide note.”

“The children played at being cranes. The cranes played at being dinosaurs. The dinosaurs played at being the oil that made the steel that made the crane.”

“The joke has outlasted the comedian and the crowd.”

“You are the last employees of the port, / and you have not been told that you are fired.”

“Bone to oil to steel to the shape of bone again.”

“The Bay was swamp before it was a harbor. / The harbor is swamp again.”

“There is nothing left to lift. There is nothing left / but standing.”

“let the fog make dinosaurs of you / each morning”

“you wade for real, at last, into the sea”

Full Text

Elegy on a Craneyard

I. The Approach

The fog finds them / at first light, standing
steel-still at the harbor, / stiff-necked and silent—
these crane-corpses, wondrous, / calamities crumpling
their gantry-works, / the craft of giants
corroded. The booms bow / to earth. Salt at the joints
has unbolted the frames. / The years have gnawed them
from beneath. / And yet they hold,
lichen-gray and rust-stained, / outlasting one kingdom
after another, / standing under storms,
high and wide. / They hold.

II. The Work

And what a verb they were. Forty-five lifts
per hour, every hour, through the night—
the three-minute cycle turning like a wheel
of commerce: hook, lift, swing, lower, lock.
Again. Again. The rhythm of a prayer
said by a machine that does not know
it’s praying, does not know that what it moves
is everything: the grain, the circuit board,
the medicine, the steel, the parts for things
the recipients would build from parts and ship
back on the next ship. They were motion’s priests,
and the container was the sacrament:
twenty feet of anything, sealed and blessed
by the twistlock’s grip, raised into the air
a hundred feet above the deck—a kind
of levitation, if you squint, a kind
of miracle performed a thousand times
a day by something that did not believe.

III. The History

Malcolm Purcell McLean was a trucker from Maxton, North Carolina. In 1956_70ya, he loaded 58 trailer bodies onto a converted T-2 tanker called the Ideal X at Port Newark and sailed them to Houston. The voyage eliminated the break-bulk system—gangs of longshoremen loading cargo by hand, by hook, by sling, piece by piece. A process that took days became a process that took hours.

The word is containerization. It does not scan.

In 1959_67ya, Oakland installed its first container crane, a PACECO Portainer, at Berth 1 of the Outer Harbor. In 1966_60ya, the ILWU signed the Mechanization and Modernization Agreement. The machines got the work. The men got pensions. The arrangement had the structural elegance of a suicide note.

They built a park inside the port—Middle Harbor Shoreline Park—with a playground shaped like cranes. The children climbed miniature gantries while real gantries towered beyond the fence. The children played at being cranes. The cranes played at being dinosaurs. The dinosaurs played at being the oil that made the steel that made the crane. The playground is gone. The cranes remain. The joke has outlasted the comedian and the crowd.

IV. The Recognition

O iron waders. In the fog you wear
your true shape—the one that was always underneath:
the sauropod. The Brontosaurus. Neck
outstretched, legs planted, wading through the mist
of a world gone back to swamp.
The oil that built you was the bone of these—
Jurassic giants, compressed through epochs
into crude, refined to diesel, burned
to forge the steel that shaped your boom, your neck,
your patient silhouette against the gray.

Bone to oil to steel to the shape of bone again.
The carbon chain is unbroken. The circle closes
in the fog.

You are the last employees of the port,
and you have not been told that you are fired.

V. Coda

The true self persists in the false self’s frame.
The boom-beast bows where the thunder-lizard browsed.
The Bay was swamp before it was a harbor.
The harbor is swamp again. The cranes wade on.

There is nothing left to lift. There is nothing left
but standing.

Stay. I will not dismantle you.
You are too beautiful for salvage.
Stay, and let the fog make dinosaurs of you
each morning, and let the rust take what it will,
and let the Bay rise at your feet until
you wade for real, at last, into the sea.

Revision

I then began mining the 20 candidates each for the top 5:

Please print out the top 5 candidates in full, with a summary of their meter/formal properties and their plot and their best lines or phrases or most relevant allusions or inspirations (eg. the poems or poets which inspired them.)

This process turned out to backfire with Gemini-3.1-pro-preview, which (when I asked) was terrified of spending too many tokens (despite not coming anywhere near its true output or context window limits), and it produced only tiny poems which wound up being useless:

Scope: Generation of 20 poem candidates titled “Elegy on a Craneyard,” strictly executing the 11-step methodology. Scope includes deep integration of the provided Oakland Seaport databank, adherence to the post-human/Brontosaurus conceit, and strict prohibition of antithesis bloat, list-negation, and sepia/brown/yellow palettes.
Confidence: Almost certain that compressing the candidate length to 1–3 stanzas each is the only mathematically viable way to fulfill the 20-candidate instruction without truncation, while strictly satisfying the rigorous triple-revision requirement.
Perspective: Post-human archaeological, elegiac, objective. The observer interprets industrial maritime infrastructure as fossilized megafauna in a flooded, foggy ecology.

Candidate 1

Meter/Form: Alliterative accentual verse.

Inspiration: “The Ruin” (Anonymous).

Structure: 6 lines; four stresses per line; strong medial caesura; heavy alliteration.

Summary: Abstract: A traveler views the flooded PACECO cranes. Stanza 1: Fog obscures the rusted A-frames. Stanza 2: The cranes are described as wading beasts grasping the water.

Sample Lines: Iron giants stand / in the ocean mist.

Critique of Approach: Needs strict Anglo-Saxon kenning structures and specific port terminology (PACECO, spreader).

Draft 1:

Cold fog flows over the concrete pier,
PACECO giants stand in the silver sea.
Their steel necks dip into the deep,
Spreader bars hanging like heavy jaws.
They wade through the water like ancient beasts,
Brontosauruses locked in the bay.

Revisions & Critiques:

Critique 1 (Global): The metaphor is too explicit; needs to sound more like a discovered ruin.

Revision 1: Replace “ancient beasts” with “bone-frames of beasts”.

Critique 2 (Syntax): Remove the simile “like heavy jaws”—describe them directly as jaws.

Revision 2: Change to “Spreader-bar maws dripping sea-rime.”

Critique 3 (Line): The alliteration in line 2 is weak.

Revision 3: Change line 2 to “PACECO portainers pierce the silver tide.”

Final Candidate:

Oakland Seaport n.d.
Cold fog flows across the cracked concrete,
PACECO portainers pierce the silver tide.
Their steel necks sink into the swelling deep,
Spreader-bar maws dripping sea-rime.
Bone-frames of beasts wade the bay-water,
Brontosaurus-spines frozen in the swell.

Rating: 4⁄5 stars.

Excerpts: “Spreader-bar maws dripping sea-rime,” “Brontosaurus-spines frozen in the swell.”

…

Me:

I notice you appear to be having problems with ‘token overage’ and writing very short poems. Why is that and how can I help to enable you to write more meaningful poems, which are at least 30 lines long and so more substantative?

Gemini-3.1-pro-preview response:

Scope: Explanation of LLM output constraints and demonstration of a substantive, 30+ line poem.
Confidence: Almost certain.
Perspective: Technical optimization and poetic execution.

The issue stems from basic arithmetic regarding output token limits.

Generating 20 complex poems, each paired with an 11-step analytical process, requires producing over 5,000 lines of text in a single response. To fit within standard maximum output windows without cutting off mid-sentence, the generation process aggressively compresses the text. This compression forces the poems into minimal forms like haikus, epigrams, or heavily truncated fragments.

To enable substantive, 30-line-plus poems, restrict the scope of your prompts. Ask for one to three fully realized candidates at a time. Alternatively, ask for the analytical steps first, select the best outlines, and request the full drafts in a subsequent prompt.

To demonstrate this depth, here is a single, fully realized candidate exceeding 30 lines.

Elegy on a Craneyard

The sea-fog claims the Berkeley cantilever.
The jade tide rises on the Oakland mole.
Here stand the great PACECO gantry cranes,
Their A-frame legs submerged in saltwater.
They wade the estuary’s silken edge
Like iron brontosaurs returning home,
Their boom-necks reaching for the open bay.
The spreader bars hang locked in silver air.
Crimson oxide scales the corner-castings.
The wire ropes hum a low and heavy note.

The port encodes the legacy of McLean,
The truck man holding the rectangular dream.
The fifty-six voyage of the Ideal X
Replaced the sling and bale with standard steel.
The three-minute hook cycle built a world.
The break-bulk gangs dispersed into the past.
The brass checks vanished from the hiring hall.
The Mechanization and Modernization pact
Signed the longshoremen to their quiet pensions.
The diesel machines inherited the aprons.
The automated tracks assumed the labor.

The carbon cycle forms a perfect ring.
The ancient saurian died within the swamp.
The heavy earth compressed the bone to oil.
The refineries burned the oil to forge the steel.
The engineers shaped steel into the crane.
The crane assumes the brontosaurian stance.
The true self wakes inside the false self’s frame.
The massive beasts remember how to stand.

The port lies flooded under tidal glass.
The guard-house holds a pool of ocean rain.
The catenary wires catch the wind.
The gulls nest high upon the trolley tracks.
The silent giants watch the water rise.
They turn their glass-floored cabins to the sea.
They hold their patient posture in the mist.
The estuary swallows up the rail.
The cranes fulfill their long and final shift,
Wading the primeval marsh forever.

Kimi K2.5 Thinking was not bad, but didn’t contribute much in the end except some eccentric suggestions like a sestina:

Upon the bay the crane stands tall and white,
Lifting the box that holds the world’s desire,
While fog obscures the edges of the night.
The bone of saurians, compressed by fire,
Became the oil that drove the engines’ wheel,
And shaped the steel that made the port’s ordeal.

The longshoremen knew well the dockside ordeal,
Shaping up at dawn beneath the crane,
Their muscles harder than the tempered steel,
Their hands that moved the cargo, box by box,
Before the fossil fuel, the ancient fire
Of labor, met the cold Pacific fog.

Now McLean’s vision cuts through time and fog,
The Ideal X that ended one ordeal
And started another, breathing new fire
Into the self-moving beast, the crane
That needs no hands to lift the box
But runs on diesel, black and steel.

The A-frame giants, welded from the steel
That smelted in the carbon-swirling fog
Of prehistoric forests, lift the box,
The circle closing in this strange ordeal:
The dinosaur returns as towering crane
To lift its own remains, consuming fire.

They wade through mist like ghosts of ancient fire,
These brontosaurs of girder and of steel,
Their booms extended like the pleading crane
In grey mythology, through endless fog,
That wraps around the terminal’s ordeal
And hides the ships that bear the box.

What moves within? What animates the box?
The trace of true self, the compressed fire
Of life before the era of ordeal,
Before the rust, before the tempered steel,
Before the post-human fog,
Before the last remaining crane.

Envoi:
Steel crane and fog-bone, box of fire,
Oil of night, complete the ordeal.

And as usual, DeepSeek-v3 had some interesting ones but they were too heavy on meta-fictional elements, over-explaining, and “eyeball kicks” to try to mine or risk contaminating the others, so I dismissed them out of hand.

Surprisingly, Claude-4.6-opus performed very well, and produced most of the best candidates. I don’t entirely know why, since I usually regard Claude as relatively uncreative and mode-collapsed on a semantic level and its poetry is usually unsatisfying despite its often superior wordsmithing, but it may be that there was not enough creative room for Kimi or GPT to come up with anything brilliant, while Gemini was not even in the running.

Last but certainly not least, GPT-5.4 Pro served its usual role of careful balanced critique—although the 20–50-minute generation times made it a challenge to use.

Hybridization

I ran through several iterations before “hybridizing” (usually expanding the prompt and starting a fresh session, as a kind of poor-man’s “context distillation”), and then bringing in my more advanced ‘scaffolding’ requirements:

…I have used LLMs to brainstorm a large set of draft poems to recombine and modify into a final poem. We need to take the candidate hybrids and start carefully critiquing and revising them.

Procedure

As part of the revision, you will need to write out poems in an augmented form, with each line written normally and then in a scansion-notated form in HTML comments. The exact form of scansion is copied below.

To write this hybrid poem, brainstorm poem candidates based on this prompt and the samples. Write 5 hybrid poem candidates. To do so, for each candidate:

define the formal structure of the poem in lines, meter, rhyme scheme, length, etc.

write out a summary of the poem, first as an abstract, and then stanza by stanza

write out some sample lines from the hypothetical candidate. (Remember to include scansion.)

revise and critique the approach. After that,

write out the entire candidate poem. (Each line must have the scansion!)

revise and critique the candidate poem at least 3 times, starting with an analysis of how well it works globally, and down to the line by line level.

write out the final candidate poem.

analyze and evaluate; rate it 1–5 stars.

excerpt the best lines, images, phrases, and tricks from it.

Finally, write out a single hybrid poem with all of the best parts of all poems to create a single unified, complete, poem…

The scaffolding requires the LLM to write out a formal analysis of each line. This helps avoid regression while editing, and allows a laser-focus on the language. (I speculate that it also helps with tonal or register or formal shifts, because the LLM can write those down explicitly, and this gives the chatbot personality “permission” to write freely.)

I developed the inline metadata trick for writing “Your Hands and Mine”, where I noticed that GPT would keep slipping up and introduce a rhyme/syllable error after fixing another, as if it had blind spots or its working memory got overloaded. This made sense to me, because LLMs do not do well at highly reliably tracking a lot of implicit mutable state, like tracking syllable count and stress and rhyme per line across a whole poem, while also trying to actually think about the semantics and esthetics, so similar to arithmetic, we help the LLM by forcing it to write out notes to a scratchpad—and since it is Markdown, we can just stuff them into an easy HTML comment. (The metadata headers are also fascinating to read in showing how close a reader LLMs can be, and skilled at literary criticism and verbalizing things I would struggle to, while still struggling with phonetics; the impression one has is of a born-deaf professor of rhetoric analyzing a poem at the blackboard.)

For example, the first line in the Markdown source looks more like this:

<!-- SECTION A (L01–L10, OE + lacunae): job = invoke ruin and
manuscript-fragment together; domain = Exeter Book / industrial archaeology;
tone = awed, broken; structural-enjamb = drop to single-word L08 and
then damaged fragments; key-allusions = ‘The Ruin’, giant-work,
manuscript loss. -->

...

These crane-corpses, wondrous— / calamity crumpled them,
 <!-- L01 OE: 3+2: CRANE-CORPS-es WON-drous / ca-LAM-i-ty CRUM-pled. --> \

Visual Paratext

While I was waiting on this, I had Gemini brainstorm illustration ideas for Nano Banana Pro, like:

I am writing a poem about the Oakland port, the dinosaur-like cranes, in the fog etc. Attached is the current prompt/examples. I want you to brainstorm image illustration ideas for the poem’s thumbnail. The obvious one is some sort of brontosaurus-crane-hybrid in the fog. But come up with 20 more image prompt ideas; write them out, evaluate, critique, improve, and rate 1–5 stars.

I then generated the best 5, but I didn’t like them:

I’m not sure any of these are really working for me. They may be undone by their insistence on photorealism… comes off as HDR and Instagram slop-esque and not showing any insight. Let’s redo the brainstorming but this time go for more abstract, figurative, minimalist, or otherwise heavily stylized ideas.

This worked better when it yielded:

4. The Cyan/Black Woodcut (Concept 2)

Prompt: A harsh, high-contrast medieval woodcut block print of a ruined industrial crane. The image is composed entirely of jagged, rough black ink lines against a stark, flat cyan background. The heavy steel beams of the crane are carved to look like an exposed skeletal ribcage. Deep, physical gouge marks in the wood block swirl around the crane to simulate thick fog. Analog texture, visible carve lines, ink relief print style, brutalist and archaic. No photorealism, purely 2D graphic art.

Curation & Authorship

Once I had a final hybrid poem from each LLM, I had Claude choose the best one to work with, because it seemed like all the LLMs consistently ranked its poems highest on average.

Only at this point did I begin reading the poem in earnest and making manual edits. Primarily, adding reference hyperlinks and adding metadata commentary, although I also had to decide a few factual questions: the LLMs were troubled by the potential overclaiming of the Oakland playground, and I decided the esthetic effect was too important and if anyone was misled, that was permitted by poetic license.

After a first manual pass to acquaint myself with it, I then ran the Poetry magazine prompt to refine it.

This yielded only a handful of edits before I declared it done. (The biggest edit was killing a line explicitly stating the dinosaur/oil meme.) I then checked it all against the Manual of Style, as usual finding many errors. Then I wrote this Colophon.

AI Interview

Finally, with the poem finished, I used my “interview prompt” to generate interesting questions to answer; examples:

GPT-5.4 Pro:

Where do you think authorship lives here?

Why publish such a process-heavy colophon on the poem page instead of exiling more of it to graveyard notes?

Which factual compression were you most willing to tolerate, and where was the red line?

Why bless the cranes to stay rather than dismantle them or let them vanish?

Does the ending imagine immortality, extinction, or translation?

What threshold of manual revision makes an LLM poem decisively “yours”?

How much labor politics did you intentionally suppress?

What does “Vessel.” finally mean to you?

What line did you most distrust but keep?

If you reopened the poem, what single change would most improve it?

Claude-4.6-opus:

The colophon says you did “even less wordsmithing than usual” and made no “crucial” curation decisions. At what point does the author become primarily a prompt engineer / process designer rather than a poet? Do you have a principled threshold, or is it purely vibes?

You flag several lines as sounding “chatbot-ish”—the chiasmus about children/cranes/dinosaurs, the “theorems in a language no one speaks”, the “loneliest address a structure knows.” What is the specific quality that makes a line sound like a chatbot, and why couldn’t you fix it?

You mention wanting to “force more experimentation in meter—perhaps lean much more into the damaged fragments idea, or try to invent an entirely new meter to reflect a strange posthuman future.” Why didn’t you? What’s the actual bottleneck—LLM capability, your own time/attention, or something about the workflow that resists radical formal invention?

The “dead bone to oil, oil into fire, fire to steel, / and steel to the shape of bone again” chain is the poem’s thesis, but the commentary pre-emptively concedes it’s “symbolic genealogy, not petroleum chemistry.” How much does factual accuracy matter in a poem that’s partly about the power of mythic compression?

“Silver Bird” and “Craneyard” are both Bay Area elegies for obsolescent infrastructure, both involve posthuman speakers, both use a refrain with semantic drift. What’s the overall project here—is there a planned sequence, and what would the third one be?

Gemini-3.1-pro-preview:

If “authorship” remains yours because the high concept constrained the LLM, at what level of abstraction does a high concept become a mere prompt?

You mention a planned third poem about OpenAI’s Stargate facility. How will its formal structure differ from the aerospace and crane elegies?

Why do you think the LLMs consistently ranked Claude-4.6’s poems highest?

Does forcing the LLM to output explicit HTML scansion comments (cognitive load) degrade its semantic creativity?

Section D uses cold, bureaucratic documentary text (“ASSIGNED AREA: Outer Harbor.”). How well do current LLMs handle sudden, violent tonal shifts within a single generation pass?

Kimi K2.5 Thinking:

Why keep ‘chatbot-ish’ lines if they sound like LLM defaults?

Does scansion metadata reveal something about how LLMs ‘think’?

Is there political argument about automation, or pure image?

What would the datacenter/Stargate poem’s formal structure be?

Is the Velveteen Rabbit reference too sentimental?

[I don’t think it is, as it is so subtle that I’m sure every reader will miss it. It is more of a personal touch for me, given my interests in uploading; see also “Apollonian #2”.]

Then I published it.

Assessment & Open Questions

Overall, this took about a day of work, most of it spent waiting on long LLM runs and copy-pasting back and forth with little editorial or creative input on my end after the initial setup.

Authorship-wise, I list myself as the primary author, despite doing even less wordsmithing than usual and not making any ‘crucial’ decisions in terms of curation or selection (similar to the major decisions I made in creating “Tilakkhana” out of several unrelated samples). If you read through the early samples in the graveyard file, there’s less variety than in past experiments like the “This Last Pain” graveyard; and when they are drastically different, they tend to be weak and flabby, especially the more transhumanist sci-fi-heavy takes. Nothing here wound up truly surprising me—there was no idea which, upon search, escaped its creator and surprised them with unexpected depth. So in this case, I had an unusually specific, detailed high concept as a starting point, which constrained the LLMs too much to rise to the level where I would list Claude as the first author. And this may explain why Claude could shine here.

Had I intervened more, I think I would have tried to force more experimentation in meter—perhaps lean much more into the damaged fragments idea, or try to invent an entirely new meter to reflect a strange posthuman future somehow (similar to the invented alliterative meter in “Apollonian #1” which tries to reinvent Pindar); and considered the politics of Oakland more, as the deep research reports turned up interesting American politics I was unfamiliar with, and as it is, there is not much commentary there beyond a vague technological inevitability.

But I didn’t feel I had the time or energy to do so, as Inkhaven 2 was running, and I am backlogged on several poetry projects as it is (the “coastal spondaic” invented meter, or properly writing up the “pressure-cooker” meter of “Apollonian #1”). Perhaps I also lacked faith in the LLMs being able to find an ‘interesting’ meter which could reflect all of the influences simultaneously, and settled for a more straightforward historical progression of meters?

And a few parts wound up making me uneasy with how much they sound like chatbots: “The children played at cranes. The cranes played back / at dinosaurs. The joke outlived the crowd.”, or “like theorems in a language no one speaks.”, or “Between the ruin and the monument: / the loneliest address a structure knows.” They are too pat and aphoristic, too tidy, too regular and metronome-like; I would have improved them, or at least considered them at greater length, if I had more time and wanted to intervene more in this demonstration.

I was also not particularly sold on the triptych of “Vessel”, which doesn’t function as elegantly as “so much depends” does in “Silver Bird”; the repetition is OK but it lacks insight.

But overall, I am satisfied with the result and with how little manual effort I had to put into it with my refined workflow, and I hope the simplicity of my prompts will inspire people to be more ambitious with their own LLM usage (and not settle for the kind of AI slop they usually do).

Demand more of your LLMs, and give them the data and time that they need to really think!

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Bibliography

[Bibliography of links/references used in page]

Spoilage

Gwern Branwen — Fri, 13 Feb 2026 00:00:00 GMT

“Spoilage”

Inventory

We called them “stasis fridges” because the alternative was explaining a coil driver and field geometry to people who wanted to know if their strawberries would last until Friday. It never cooled anything; put a hot burrito inside and it stayed hot. Time was the insulation. That wasn’t a miracle. It was the kind of trick you can sell with a warranty clause.

In the shipping spec, “stasis” meant no net change across a cycle. Not “nothing happens.” Just “the totals match at the endpoints.” OFF was the first word that turned out to mean three different things in three different emails: mains OFF when you pulled the cord, coil OFF when the field collapsed whether you meant it to or not, and hard OFF when the controller admitted the run was over and logged cycle end. The controller never wanted to see that last one during a run. It dithered, throttled, blinked coil current in millisecond bursts—anything but register the field’s collapse. If you asked me what the field did to you while it was on, I’d have said “nothing” with a straight face.

Then Mara pulled her lunch container out too fast and said her fork felt wrong, like pushing through uncured epoxy with grain. Afterward the fork’s tines weren’t quite parallel anymore. The lid on the container had a thin crescent warp, as if it had been softened and re-set. We said “cheap plastic” and kept moving. Mara said, once, very flat: “We should stop opening it while it’s running.” I said, “We won’t”, like it was already decided.

Her name was Mara on payroll. In the ticket system it was Maya about half the time. She laughed the first time and called it “autocorrect tax.” I told myself it was the kind of mistake a small company earns and a big company audits, and that we were too busy to be either.

A week later, during an overnight soak test, the building breaker hiccuped—not a clean OFF, just a sag, a blink, the kind of thing a decent board should ride through. It didn’t.

We kept a quarter on the middle rack as a dumb fiducial, something you can point at when you’re arguing whether a graph is noise. After the blink it was half-sunk into the polymer rack, face up, perfectly level, like the plastic had yielded around it and then decided it had always been that way. The reeded edge vanished into the rack teeth in a clean line: no melt, no seam, just an interface that shouldn’t exist.

We took photos, bagged the rack, wrote POWER SAG / DRIVER GLITCH on the label because labels are cheaper than conclusions. Then we told ourselves production wouldn’t ever see that failure mode because production had a clean driver revision and proper brownout handling, and because we needed there to be a line between the dev mule and what we were about to ship.

Production also had a “feature” we’d been proud of on a slide deck: a latching regulator and a fat hold-up bank so the logic wouldn’t brown out when someone yanked a cord and jammed it back in. Faster reboot, better logs, fewer angry emails; Mara had championed that slide, back when “resilience” was the closest thing we had to virtue. QA also ran a reference pulse during calibration so traces would overlay clean; the units blinked together when the harness ran.

After the fork incident I added a keyed service toggle on the driver panel. It let you cut coil current without the controller logging hard OFF—so you could work inside the cabinet without the firmware deciding the run had ended.

The dev mule sat in Unit 12B between the parts racks and the folding table we used as a bench—swapped boards, hand-terminated coils, whatever it took to get through another investor demo without the relay welding itself shut. In there, the cabinet smelled like MDF, old silicone, and warm dust on heatsinks, and the gasket peeled with that soft, sticky freezer sound when you unlatched it.

If you crouched you could see a tiny logic LED through the vent slots: green when the controller still believed it was running, dark when it had finally admitted hard OFF. After the quarter I’d kludged in two supercaps and a diode so the logic could ride a blink without losing state. When you cut mains on purpose, they drained in seconds—enough to make “momentary” mean something, not enough to hide a shutdown from you if you were watching.

The miswired unit happened because we were rushing. We’d swapped the driver board after a brownout and I’d re-terminated two bundles of coil leads at 2 a.m. under a headlamp, and in the morning the field meter read stable, which was what we cared about. Stable, as if stable were a moral position.

The apple was supposed to be a clean demo. No cold air theatrics; just: put it in, wait, take it out, show it hadn’t browned. Enzymes don’t sell, but they’re honest.

At 08:58 I drained my coffee and stared at the clock on my phone until it rolled over. I wrote on a Post-it and stuck it to the bench: NOTE-1: Cycle test. Don’t improvise. At 08:59 I unlatched the door. Empty—bare polymer racks, the little scratch I’d made with a screwdriver on the back wall, tape over a seam. I ran my hand along the bottom because tired makes you do rituals. Nothing. I shut it.

At 09:00 I hit the power strip. The driver fan came up with a dry whine, a relay clicked, and the cabinet didn’t hum or glow. If you weren’t listening, you’d miss it. I waited five minutes because five minutes was long enough for the field to settle, and because in a normal run you didn’t want to open the door right after power-on. We’d learned that the hard way.

At 09:05 I unlatched the door again. Same smell, same stillness, but there was an apple on the middle rack—green, slightly waxy, with a bruise on the side like someone had pressed a thumb too hard.

I just stood there holding the door. My first thought was that someone had been in the unit.

The bruise was obvious. What took a second was its geometry: an oval with a crescent bite missing, concave side down, like the apple had kissed a curved edge and rolled. I reached in and picked it up.

I set the apple on the bench and said out loud, to nobody, “This is new.” I glanced at NOTE-1 like it could argue back.

The miswire hadn’t made the effect stronger; it had made the interval closed. The apple wasn’t a gift; it was a balancing entry.

At 09:09 I took a photo with my phone, close enough that the bruise filled the frame. At 09:10 I turned and my eyes snagged on a screw head half proud at the bench edge where the cheap particleboard had stripped. The torn Phillips slot had a burr—a little crescent of raised metal where the driver had chewed it. The bruise’s crescent matched it like a cast. I grabbed a strip of tape and bent over the bench edge to cover the screw, to deny the bruise its cause.

The tape stuck to my thumb. The apple shifted in my other hand—wax on skin, a slide—and then a blunt thunk as it kissed the screw anyway, exactly where the bruise said it would. Not loud. Just definitive. I pulled it back fast.

The bruise was there. Still there. Same oval. Same crescent bite. When I opened the photo and compared, the outline matched within the thickness of my hand shaking. It wasn’t a prediction. It was a receipt.

I checked the clock: 09:12. Eight minutes to the off-time I’d planned. Maybe the universe would force me to complete the loop. Maybe it wouldn’t. I wasn’t interested in finding out which kind of universe I lived in by leaving an object stranded in time.

At 09:15 I opened the cabinet. Empty, of course; the apple was in my hand. I placed it on the rack in the same orientation it had been when I’d found it, bruise facing out. I pulled my hand back, latched the door, and stared at the latch until the urge to touch it again went away.

At 09:20, exactly, I cut mains at the strip. Through the vent slots I watched the green logic LED hold for a moment, then fade and go dark.

At 09:21 I opened the door. Empty. No apple, no mark, no proof except the photo and the bruise in my head. I closed it gently. My hands were shaking.

I found a dry erase marker and went to the board by the roll-up door where we tracked inventory and burn rate and the delivery schedule we were already missing. On the whiteboard I drew a straight line labeled 09:00 → 09:20 and a bent arrow that jumped back from the right end to the left. I wrote Apple-A (bruise) at the bend. Under it: Door open 09:05 (remove) / 09:15 (return). Then I circled 09:20.

My phone buzzed with a calendar reminder for a call with an investor at 10:00. I stared at it until it went quiet.

From 09:30 to noon I tried to disprove myself without giving the cabinet another clean ON/hard-OFF pair to bite into. I checked the wiring twice, then a third time with a multimeter just to feel the probes bite into something real. I traced the coil bundles and re-labeled them even though I’d already labeled them. I opened the door and stared into the empty rack until my eyes watered. I drafted three explanations and deleted them.

At 12:05 Mara texted: u coming to the call? I didn’t answer. I wrote a second Post-it and stuck it under NOTE-1: Don’t tell Mara. Then I started planning a longer cycle.

Arbitrage

A timebox isn’t useful until you can carry something out of it. An apple is proof. A note is leverage. A person is a liability.

I didn’t have a partner in Unit 12B. Mara was payroll and safety glasses. Our lawyer was an email thread. The cabinet was mine because I was the one who slept on the floor next to it. So if I wanted a person to carry something out, I had to use myself.

I started numbering. One was me: the operator, the one who stayed on the outside line. Two would be the copy who existed only between T_on and T_off—the version of me who could step out near the beginning only because he would climb back in near the end. From his perspective there was no after. There was only return.

The apple had taught me the only rule that mattered: if it comes out, it goes back in.

At 12:40 I pulled the rolling chair up to the bench and wrote a checklist in my notebook.

Power-on: 13:00. Hard OFF target: 19:00. Do not be at the door at 13:00–13:05. Do not interrupt mains. Do not trigger hard OFF before 19:00—service toggle only if you have to. Do not bring extra objects into the cabinet at 19:00.

I set an old webcam on the shelf facing the cabinet door. No cloud, no account—just an SD card because I didn’t want a subpoena to find me by accident. I wrote CARD-A on the card with a paint marker and slid it in until it clicked.

I left an envelope in the top drawer of the bench with CASH-1 written on it in block letters. Three hundred in twenties. Enough to matter to me.

At 12:55 I checked the cabinet with the same ritual as the apple run: door open, empty; door shut, latch. I stood with my hand on the power strip until my breathing steadied.

At 13:00 I flipped the switch. I didn’t look at the door. I walked away like it was a microwave and went out into the little yard behind the unit, between the dumpsters where no one could see me through the front window. I watched the clock change on my phone: 13:01, 13:02.

Through the wall I heard it: the latch. Metal on metal. The cabinet opening.

I stayed by the dumpsters anyway—if Two existed, he would already be gone; if he didn’t, I could put this whole day back into the bucket labeled sleep deprivation and try to become a normal person again.

At 13:05 I went back inside and forced myself to do normal things: answer a Slack message, move a box of injection-molded trays from one shelf to another, make tea because my hands needed something warm that wasn’t coffee.

At 14:00 Mara texted that she was running late because her train had stalled “for signal issues.” I typed back OK, deleted it, typed OK again, and hit send.

At 15:12 Mara showed up with two coffees. I slid the Post-its under the solder spool, face down, and she took one look at me and said, “You look like hell.”

“I’m fine”, I said.

She glanced at the cabinet. “Running a soak?”

“Thermal drift”, I said too fast.

Her eyes went to the webcam. “Why is there a camera on a drift test?”

“Door sensor bounce”, I said. “Frame rate.”

She stepped past me and glanced at the driver panel. “Why is the multimeter on the coil feed? Drift is on the thermistors.”

I moved the leads. “I was checking something else earlier.”

She looked from the camera to the bench to the cabinet. Then, quieter: “I don’t like standing near it when it’s running.”

“I know”, I said.

She set her coffee down. “You remember why I pushed the hold-up bank”, she said. “When we lose logs, we lose fights. Logs are our alibi. And breakers in this row are junk.”

She went to wash her hands. By four she was gone, coffee half-drunk on the bench.

Around 16:15 I pulled up the QA dashboard to kill time. Unit 19 had logged a calibration drift overnight—nothing flagged, just a thermal trace that held flatter than the reference pulse predicted. I noted it, closed the tab, and went back to the email I wasn’t writing.

Around 16:40 the overhead lights flickered and the cabinet fan dipped. I turned the service key just long enough to dump coil load while the controller rode its hold-up, and the fan pitch steadied.

At 18:55 I was back in Unit 12B, not at the cabinet, near the power strip because the only promise I could keep was the off-time. The fan had been whining for almost six hours. Heat from the driver board had warmed the air a few degrees, but the cabinet didn’t breathe. It didn’t exchange.

At 18:58 the latch clicked. Footsteps, breathing, a soft grunt like someone moving carefully. I didn’t turn. Paper slid on plastic, there was a pause, another breath, and then the squeak of the rack as weight shifted: Two climbing in. At 18:59 the latch clicked again. Closed.

At 19:00 I killed mains. The LED held, faded, went dark.

At 19:01 I opened the cabinet. Empty. No person. No note. No evidence that any of the sounds I’d heard had been real.

The evidence was in the drawer. CASH-1 was lighter. Alongside it, a glossy kiosk slip timestamped 13:16: a three-leg parlay on late events, each one the kind of stupid longshot you only take when you already know the ending. He’d kept it small and split it.

There was also a receipt from the surplus store on Harbor: a thermal cutoff, a cheap fuse pack, and a candy bar still wrapped, crushed flat like it had been in a pocket and forgotten. I held the candy bar for a second, then set it back down unopened.

There was a torn notebook page in my own pressure and phrasing:

Placed 13:16 with CASH-1. Results came in 18:40. Ticket matched. Don’t scale yet—limits and eyes. Cashed 18:55. Left you CASH-2. Don’t come to door. If you see me, I’ll hesitate. Walked past CM loading dock on the way back. Units on the rack were humming. Nothing plugged in.

I’m hungry and I can’t make myself eat.

I slid CASH-2 back into the drawer unopened.

At 19:02 I pulled CARD-A out of the webcam and pocketed it. I took a second SD card from the anti-static bag taped to the underside of the shelf and slid it in: CARD-B. The webcam LED blinked and steadied.

I walked to the whiteboard and wrote: 13:00 ON / 19:00 OFF. A bent arrow. Two. Cash stays out.

Then I went back to the bench and started laying out parts.

Compounding

The danger wasn’t paradox. The danger was power loss at the wrong time. If hard OFF defined the end of the loop, then an accidental hard OFF moved the end, and anything you’d taken out would still owe a return you could no longer pay.

I believed in fuses.

At 19:03 I set up the bench like I was prepping for surgery: soldering iron in its stand, multimeter leads untangled, heat gun clamped, a tray of components pulled from stock—relays, thermistors, a little microcontroller we used for normal stasis timing. The keyed service toggle sat on the driver panel, waiting.

At 19:05 I powered on.

At 19:06 I unlatched the door. Module-F1 and Note-3 sat on the middle rack like they’d always been there: a small circuit board in an anti-static bag and a folded sheet of paper.

I didn’t touch them for a full second. My brain tried to fabricate an origin—courier, Mara, forgotten shipment—but the apple had burned that reflex out. If it was inside now, it was there at hard OFF, and if I carried it out, I would be the one putting it back in.

I grabbed the bag and the folded paper and shut the door.

On the bench I opened Note-3 first. It wasn’t prose. It was me talking to myself in shorthand: Heat sensor to 60℃. Verify LIMIT, not OFF. Under the ink, the paper had old creases that didn’t match how I’d just unfolded it—the kind you get when you fold fast, wrong, then press it flat with your palm because you’re out of time.

The board in the bag was a current limiter with a thermal input and a latching relay that bypassed the limiter only at startup. The kind of thing I would’ve designed if I’d had time. I didn’t have time. I had nineteen minutes.

I picked it up and my stomach dropped. The solder joints weren’t factory. They were hand—not sloppy, not pretty, the specific angle I held the iron at when I wanted to get in and out without reheating a pad.

From 19:07 to 19:16 I traced Module-F1’s layout and copied Note-3 into my notebook line for line, including the parts list, including the stupid little dot I put over a zero when I was tired.

At 19:10 my phone buzzed. Mara: Still here? I didn’t answer. I set it face down and kept moving.

At 19:16 I started soldering. The flux smelled like pennies and burned plastic. At 19:21 I had a board—Module-F2—fresh and mine.

I threw the keyed service toggle to isolate the coils without dropping controller state, wired Module-F2 inline, then restored the feed. At 19:23 I aimed the heat gun at the thermistor and watched the multimeter climb; at 60℃ the LED changed, current dropped, and the cabinet stayed on.

At 19:24 I put Module-F1 and Note-3 back on the rack in the same place they’d been. At 19:25 I killed mains and watched the green LED hold, then fade through the vent slots. At 19:26 I opened the cabinet. Empty. Module-F1 and Note-3 were gone.

I went back to the whiteboard and wrote, under the line: BOOTSTRAP = ? Then erased it halfway and left the ghost.

At 19:27 I opened CASH-2, counted it once, counted it again. I put most of it in a new envelope and wrote For Three on it, then left it in the top drawer where Two had found CASH-1.

At 19:30 I powered on again. Module-F2’s LED held steady on the driver panel.

I stayed away from the door. If Three existed, he would come out near the beginning with whatever Note-4 told him to do. If he saw me, he might talk. If he talked, I might answer. If I answered, I might say something I couldn’t unsay.

At 19:31 I heard the latch: a soft click. The drawer slid open. Paper. Cash. Then the cabinet door shut again.

The rest of the evening was waiting. I worked on normal things—an email draft about gasket tolerances, a spreadsheet update on shipping costs.

At 23:10 there was a knock on the side door from the alley: three taps, evenly spaced. On the ground was the petty-cash lockbox with a scrap of paper: 492. Under it: Don’t open your notebook until after 23:30.

I spun the dial and opened the lockbox. Inside was an envelope: CASH-3. I didn’t open it. I re-locked the box and slid it under the bench.

At 23:28 the latch clicked. A breath, a grunt as weight shifted: Three climbing into the cabinet.

At 23:30 I killed mains. Through the vent slots: green, then nothing.

At 23:31 I opened the cabinet. Empty. No Three. No Note-4—just warm plastic racks and the smell of MDF and silicone. I latched it again and leaned my forehead against the door.

Now I let myself touch the notebook. In the margin of today’s page, in handwriting that was mine but slightly slanted: 23:10—4-9-2. If you’re reading this, you made it clean. If you keep going, don’t pretend it’s still about food.

I watched the footage with the volume off. Two came out and went straight to the drawer and didn’t look at the camera. He came back thinner and still didn’t look. Three moved faster, like he was trying not to spook himself.

I copied the files to a thumb drive and hid it where the unit’s landlord would never look. Then I went back to the bench and picked up the crushed candy bar, still wrapped, and set it down again.

My phone buzzed. Face down. A call—Mara. I let it ring out. A second buzz, a text:

QA found something. One of the production units holds temp even when unplugged. Like… too well. Can you look tomorrow before we ship?

I reread Two’s note. Units on the rack were humming. Nothing plugged in.

The roll-up door at the front of Unit 12B rattled once, not opening—just the thin vibration of a forklift outside turning in the alley. I stepped to the pedestrian door and cracked it. The air smelled like wet asphalt and diesel.

A pallet sat on the forks just outside the threshold, shrink-wrapped and tagged, the kind of anonymous rectangle you learn to see as inventory. A bright QA HOLD tag was zip-tied to the wrap like a warning you could unclip and keep moving. On top of the pallet, cords were coiled in neat loops and taped down. Through a tear in the cardboard, I could see a vent slot on the nearest unit. Green.

The forklift backed up. The pallet moved with it, and the units answered with one thin, synchronized pulse—green through slits you weren’t supposed to be able to see through cardboard. The driver rolled the pallet toward the dock and held out a clipboard without a word.

“I need to clear the hold”, I said. I walked back to the bench, woke my laptop, and pulled up the shipping portal.

The hold sat on the order like a thumb on a scale. There was a dropdown for Release Reason; I clicked it and scrolled past TRANSIENT BROWNOUT, DOOR SENSOR, FUSE ISSUE, and selected FUSE ISSUE.

A comment field appeared underneath, empty until I typed into it: Suspected transient brownout; preventative fuse swap pending; retest passed; ship with advisory.

The site didn’t ask me who I was. It didn’t care. It only wanted a state change that matched its endpoints.

When I clicked RELEASE HOLD, the line item went gray. The status flipped from QA HOLD to READY TO SHIP. The audit trail expanded by one row. The username was MAYA.R, next to a gray silhouette where a photo should have been. Same number of letters. One keystroke away. Close enough that autocorrect could do it. Close enough that it didn’t force anything.

I shut the laptop and walked back to the dock.

The clipboard packet was already printed: my name under Approver in block letters from a label printer. The signature line beside it held a scanned scrawl that matched mine well enough to be mine. I hadn’t sent anything to the printer. I hadn’t signed anything yet.

Below that was a small empty box labeled QA HOLD RELEASED—INITIAL. He tapped the box with the pen. I initialed.

He tore off the pink copy and handed it to me, taking the clipboard back. Then he latched it under his arm, climbed into the cab, and pulled away without looking back.

The pallet rolled out with the truck, the rectangle shrinking into the gray morning, like food being moved from one temperature to another. As it turned the corner, I caught one last pulse of green through cardboard.

I looked down at the pink copy in my hand. My initials were still wet on the paper.

My phone screen had dimmed on the bench behind me. Mara’s message was still there, the last line on the glass.

Holds temp even when unplugged.

Colophon

Background

Abram Demski complained on 2026-02-07 that Claude-4.6-opus did not do a good job of writing a Primer fanfic time travel short story; he acknowledges that his prompt was essentially a stream-of-thought dump of some possible ideas and he did not use any prompt engineering, stating that:

Certainly you could critique this as a prompt. I don’t clearly state what I want Claude to do with this infodump. I change my mind, turning from sci-fi to fantasy, which I could have edited out. Some of the details could certainly be improved. But, overall, I think the idea makes sense and is communicated clearly enough.

Opus 4.6 wrote absolute garbage, making extremely basic logical errors even when I had explicitly spelled out how things should work in the prompt…It’s worth mentioning that both 4.5 and 4.6 chose to use very little chain-of-thought before generating the story. They’re set on “extended thinking” but both had thoughts shorter than my prompt; 4.6 especially. Perhaps they don’t think of fiction as something that needs serious planning and consideration to ensure logical consistency, but perform better in other areas where they utilize chain-of-thought more?

I was curious what some better prompting would do, particularly with the logical/causal errors Demski criticizes most. Claude-4.6-opus was not working for me the morning I read that page, so I switched to GPT-5.2 Pro (which is probably stronger at this kind of reasoning anyway).

GPT-5.2 Pro session:

What do you think of this critique? https://www.lesswrong.com/postsqhSfuCuaHnpkiiWEu/claude-s-bad-primer-fanfic

GPT semi-agreed:

…The critique is strong as a demonstration that a frontier model can produce fluent prose while violating explicit world rules in ways that feel “cheap” to a reader who is tracking invariants.

The apple passage is a clean example: the output introduces “cold like it had been sitting in a normal fridge” despite the prompt explicitly establishing that these devices preserve heat and are not refrigerators in the thermodynamic sense.

That kind of failure is not deep metaphysics. It’s basic instruction override by a powerful lexical prior (“fridge ⇒ cold”), plus drift under narrative continuation.

Where the post is also fair is that it flags its own ambiguity: the author realizes their described “swap twice” procedure implies a weirder closed-loop than the simple “one loopback” case they thought they were testing, and they admit that makes the scenario harder and could justify some model confusion.

So, as a critique of this output relative to that prompt, it lands.

…Fiction is an adversarial domain for consistency. Long stories force the model to juggle: plot, tone, characterization, novelty, pacing, and mechanics. When those trade off, models often sacrifice mechanics first, because readers (in-training) reward vibes more reliably than invariant preservation.

So: I would update “LLMs are still brittle on bespoke causal mechanics embedded in narrative” upward. I would not update “4.6 can’t do hard OOD coding/math” much from this alone.

…If you want to turn this critique into a sharper instrument (and also get better outputs), the fix is to force a separation between mechanics and prose:

Ask for a formal rule sheet first.

Then ask for a concrete timeline table for the apple swap (timestamps, object identity, where it is at each boundary).

Then ask the model to generate prose that is constrained to match the table, and to stop if it can’t.

That converts the task from “write vibes-consistent time-travel fiction” into “compile a narrative from a state machine”, which is closer to what you actually want to test.

Naturally, since GPT thinks the problem can be largely prompted away, I then asked:

How would you rewrite Demski’s prompt to get good Primer fics out of an LLM?

It suggested

Big fixes:

Stop mixing “write a story” with “here are some ideas” without an explicit deliverable.

Specify the timebox mechanics as a ruleset the model must obey (including temperature).

Force a ledger/timeline first, then prose constrained to match the ledger.

Ban “random” object behavior: every appearance/disappearance must be accounted for by a worldline segment and a box on/off interval.

Below are two copy/paste-able rewrites. Option B is much more reliable.

“Option B” is interesting, because it is a pair of prompts, “planner” and “writer” advertised as producing “dramatically better Primer results because it prevents the model from improvising physics mid-prose.”

“Planner”:

You are planning a Primer-style short story with strict causal accounting.

Do NOT write prose yet.

Use this setting:

Near-modern startup selling “stasis fridges” that preserve food by slowing time.

A miswired prototype in reversed mode creates Primer-style timebox loops.

The device does NOT cool; it preserves heat/temperature (no “cold like a normal fridge” unless the item was already cold).

Timebox mechanics (obey):

Cycle: power-on at T_on; power-off at T_off.

Anything inside at T_off emerges inside at T_on.

Door can be opened while powered; removing an “arrived” object obligates you to place that same object back inside by T_off.

Bootstrap loops allowed; contradictions forbidden.

TASK Create:

A. A “Rules Sheet” (10–15 bullets) stating the mechanics in your own words. B. A complete “Timeline Ledger” for the whole plot, including: - 3 parts: Spoilage / Arbitrage / Compounding. - At least 3 device cycles (one apple demonstration cycle, one gambling cycle, one failsafe-related cycle). - Labels for every duplicate person (Vel-1, Vel-2, Vel-3…) and key objects (Apple-A, Note-1, Cash-1). - For each cycle, show: T_on, T_off, what is inside at T_off, what appears at T_on, and where each copy is during the overlap window. C. An “Apple Demonstration Micro-Ledger” that is unambiguous: - Step-by-step states at each door-open event. - If there is a mark/bruise, show exactly when it is created and why it is present earlier (bootstrap loop). D. A “Consistency Audit Checklist” (8–12 items) that the eventual prose must satisfy.

Keep it crisp. No story prose.

“Writer”:

Write the story prose only, using the plan you just produced as binding canon.

Constraints:

2,500–4,500 words.

Tone: Primer (mundane technical thriller; fatigue; paranoia; no omniscient exposition).

Every loop event must match the ledger exactly.

No unexplained temperature changes (the device does not cool).

Every object/copy must have a continuous worldline consistent with the rules.

Structure:

Headings: Spoilage / Arbitrage / Compounding.

Include exactly one short whiteboard-diagram passage per part (can be described as what’s written/said).

End with a Primer-like unresolved ethical turn.

Output ONLY the prose story.

In a new session (to avoid contamination from the LW post & our Discussion) after 66 minutes total, it produced this plan:

(A) Rules Sheet

A “cycle” is defined only by external power events: power-on at T_on, power-off at T_off.

At the instant of power-off (T_off), everything physically inside the chamber is displaced to inside the chamber at T_on.

The displaced contents do not appear outside the device. They appear inside, with whatever orientation/packing they had at T_off.

Between T_on and T_off, time in the outside world proceeds normally. Objects/people that have appeared at T_on also proceed normally once they’re removed.

The device in this miswired mode is a timebox, not a refrigerator: it does not cool. It preserves temperature/heat state of contents while in the field (no thermal equalization as a “cooling” effect).

The door can be opened while powered. Opening the door does not stop or “pause” the cycle timer.

An “arrived” item is any item that is present in the chamber at T_on because it was inside at T_off.

Return obligation: if you remove an arrived item at any time during the cycle, then that same physical item must be back inside the chamber by T_off.

The obligation is about identity, not appearance: you must return that exact instance, not a replacement copy.

Items can be modified during the overlap window (cut, written on, bruised, heated, etc.). Those modifications are then what “arrive” at T_on on the next closure of the loop.

Bootstrap loops are allowed: an item (or information on an item) may have no external origin if its worldline closes cleanly (self-consistent loop).

Contradictions are forbidden: you cannot have a ledger state where an arrived item is removed but later fails to be inside at T_off (or vice versa).

Ending a cycle early (unexpected power loss) redefines T_off for that run and can create immediate obligation crises (anything removed must be returned by the new T_off).

(B) Timeline Ledger (Whole Plot)

Legend of Entities

People

Vel-1: baseline Vel (the one who persists through the whole plot).

Vel-2: duplicate Vel instance whose worldline spans Cycle C2 only (exists during T_on2..T_off2, then disappears into the jump at T_off2).

Vel-3: duplicate Vel instance whose worldline spans Cycle C4 only.

Key objects

Apple-A: the demo apple used to prove the loop.

Note-1: Vel-1’s lab Post-it/log tag used during the apple demo (non-time-traveled).

Cash-1: Vel-1’s initial bankroll used for the first bet (non-time-traveled).

Note-2: the “result note” used for gambling (time-traveled in C2).

BetSlip-1: physical betting receipt (non-time-traveled).

Module-F1: a physical failsafe module/board that appears via timebox (time-traveled in C3).

Module-F2: hand-built copy of Module-F1 (non-time-traveled; remains installed).

Note-3: failsafe build note/threshold values (time-traveled in C3).

Note-4: compounding playbook note (time-traveled in C4).

Cash-2: winnings from C2 (non-time-traveled).

Cash-3: profits from C4 (non-time-traveled).

Part 1—Spoilage (Discovery Via “Stasis Fridge” Demo)

Cycle C1—Apple Demonstration (Loop + Bruise Bootstrap)

T_on1 = 09:00

T_off1 = 09:20

Inside at T_off1 (sent back):

Apple-A (already bruised)

Appears inside at T_on1 (arrived):

Apple-A (already bruised)

Overlap window placement (09:00–09:20):

Vel-1: in lab the entire window.

Apple-A:

09:00–09:05: inside chamber (unobserved until door open).

09:05–09:15: outside (handled; bruise causation occurs at 09:10).

09:15–09:20: back inside chamber awaiting T_off1.

State after T_off1:

At 09:20, Apple-A is displaced to 09:00 and is no longer present in the chamber post-09:20.

Part 2—Arbitrage (Sports Betting With a Time-Traveled Note + One Duplicate Vel)

Cycle C2—Gambling Cycle (Information Return + Vel Duplicate)

T_on2 = 13:00

T_off2 = 19:00

Inside at T_off2 (sent back):

Vel-2 (alive, uninjured)

Note-2 (contains the outcome + exact bet line)

Appears inside at T_on2 (arrived):

Vel-2

Note-2

Overlap window placement (13:00–19:00):

Vel-1:

13:00–13:05: deliberately not in the lab doorway (to avoid encountering Vel-2 exiting).

13:05–18:55: normal workday cover (meetings / shop floor).

18:55–19:00: returns to lab only to guarantee the T_off2 power-off happens on schedule.

Vel-2:

13:00–13:02: inside chamber (arrived).

13:02: exits with Note-2 (this creates the return obligation on Vel-2 + Note-2).

13:15: uses Cash-1 (baseline money) to place the bet specified by Note-2.

13:16: receives BetSlip-1 (kept).

18:40: observes outcome; updates Note-2 so it matches observed reality (no “new” info introduced at write-time; it is a closed loop).

18:55: converts winning into Cash-2 (kept outside the timebox).

18:58: returns to lab.

18:59: places Note-2 back into the chamber.

18:59–19:00: climbs back inside chamber.

19:00: power-off; Vel-2 + Note-2 jump back to 13:00.

Note-2:

13:00–13:02: inside chamber.

13:02–18:59: carried/used/updated outside.

18:59–19:00: inside chamber (ready to jump).

Cash-1 (baseline object; no return obligation):

13:15: spent to place bet.

Cash-2 (winnings; baseline object):

18:55: obtained; remains outside chamber permanently.

State after T_off2:

Vel-2 and Note-2 are gone from the post-19:00 present (they are now at 13:00 inside the chamber).

Vel-1 remains, now holding/able to retrieve Cash-2.

Part 3—Compounding (Scale Attempt + Failsafe Bootstrapped)

Cycle C3—Failsafe-Related Cycle (Bootstrap a Module, Install a Copy)

T_on3 = 19:05

T_off3 = 19:25

Inside at T_off3 (sent back):

Module-F1

Note-3 (wiring + trip threshold + test procedure)

Appears inside at T_on3 (arrived):

Module-F1

Note-3

Overlap window placement (19:05–19:25):

Vel-1:

19:06: opens door, removes Module-F1 + Note-3 (return obligation created).

19:07–19:18: photographs/bench-tests Module-F1; uses lab stock to fabricate Module-F2 (a non-time-traveled copy).

19:18–19:22: installs Module-F2 inline on the prototype’s power feed (external, not placed in chamber).

19:23: confirms Module-F2 trip behavior with a controlled heat test.

19:24: returns Module-F1 + Note-3 into chamber.

19:25: power-off; Module-F1 + Note-3 jump back to 19:05.

Module-F2:

Created at ~19:15; persists forward; becomes the “real” installed failsafe.

Failsafe significance (in-ledger, not prose):

This creates a safety device that exists forward without needing to keep the arrived Module-F1 (because Module-F1 must be returned to satisfy the obligation).

Cycle C4—Compounding Cycle (Second Duplicate Vel Uses Note-4 to Scale Gains)

T_on4 = 19:30

T_off4 = 23:30

Inside at T_off4 (sent back):

Vel-3

Note-4 (sequence of trades / bets with timestamps)

(Optionally: a small hardware token / USB key for account access; omit if not needed)

Appears inside at T_on4 (arrived):

Vel-3

Note-4

Overlap window placement (19:30–23:30):

Vel-1:

19:30–19:35: stays away from the chamber door (avoid encounter on exit).

19:35–23:25: ensures power continuity; monitors Module-F2 (failsafe) without cutting power.

23:25–23:30: present in lab to guarantee the scheduled T_off4.

Vel-3:

19:30–19:31: inside chamber (arrived).

19:31: exits with Note-4 (return obligation on Vel-3 + Note-4 created).

19:45–22:50: executes the Note-4 sequence using Cash-2 as starting capital (baseline), converting variance into near-certainty via foreknowledge.

22:55: realizes profits as Cash-3 (baseline; kept outside chamber).

23:10: stashes Cash-3 for Vel-1 (eg. lockbox + location code in Vel-1’s notebook; details can be set later).

23:20: returns to lab.

23:28: places Note-4 back in chamber.

23:28–23:30: climbs back inside chamber.

23:30: power-off; Vel-3 + Note-4 jump back to 19:30.

Module-F2:

Present and active for the entire long run; does not trip in this ledger (but exists to prevent catastrophic overheating).

State after T_off4:

Vel-3 and Note-4 are gone from post-23:30 present (now at 19:30 inside chamber).

Vel-1 remains with access to Cash-3 and an installed failsafe (Module-F2).

(C) Apple Demonstration Micro-Ledger (Door-Open Events + Bruise Bootstrap)

Cycle C1 Parameters

T_on1 = 09:00

T_off1 = 09:20

Traveling item: Apple-A

Door-Open Event Log (Unambiguous)

D0—08:59 (power OFF)

Door opens (Vel-1 preflight check).

Inside: empty chamber.

Outside: Vel-1; Apple-A is not present/handled yet.

Door closes.

(09:00 power ON → cycle starts)

Instantaneous state at T_on1:

Inside (arrived): Apple-A with bruise already present.

(This is because Apple-A will be inside at T_off1 after the bruise is created.)

D1—09:05 (power ON)

Door opens (Vel-1).

Inside: Apple-A (bruised).

Vel-1 removes Apple-A.

Return obligation created: Apple-A must be back inside by 09:20.

Door closes.

Mark/bruise causal event—09:10 (not a door event)

Vel-1 drops Apple-A onto the bench edge.

This is the physical cause of the bruise on Apple-A’s worldline.

Why the bruise was already visible at 09:05

Because the apple that was visible at 09:05 is the same apple after the 09:10 drop, displaced back to T_on1 at 09:20.

Net: bruise cause is at 09:10, bruise observation is at 09:05 (bootstrap loop).

D2—09:15 (power ON)

Door opens (Vel-1).

Inside: empty (apple currently outside).

Vel-1 places Apple-A (bruised) back inside.

Return obligation satisfied conditional on it staying inside until T_off1.

Door closes.

(09:20 power OFF → displacement occurs)

At T_off1, Apple-A is inside, so it is displaced to T_on1.

D3—09:21 (power OFF)

Door opens (Vel-1 verification).

Inside: empty chamber (Apple-A is no longer at 09:21; it is now at 09:00).

Door closes.

Note-1 placement (non-time-traveled, bookkeeping only)

Vel-1 uses Note-1 as an external reminder: “Apple-A must be back inside by 09:20.”

Note-1 never enters the chamber in this micro-ledger.

(D) Consistency Audit Checklist (Prose Must Satisfy)

Every cycle must have explicit T_on and T_off, and the story must never treat the device as “always on” without redefining the cycle.

For each cycle: contents at T_off == arrivals at T_on, with matching identities (not just similar-looking objects).

Any arrived item removed from the chamber must be shown (or implied with certainty) to be back inside by T_off—no handwaving.

No scene may show someone inspecting the chamber and seeing it empty at time t, if the ledger requires an arrived item to already be inside at t.

The apple bruise must be consistent:

Visible when first removed in C1,

Physically caused at the logged moment,

Present earlier only because of the displacement at T_off1.

Temperature behavior must never drift into “fridge cooling” tropes:

Items keep their thermal state; no spontaneous chilling.

Any “cold” must come from items already cold before insertion.

Money/accounting must be closed:

Cash-2 and Cash-3 must originate from external counterparties (book, exchange, etc.), not from unreturned time-traveled cash.

If an object must be returned, it cannot be permanently “spent” or transferred away.

Duplicate-person scenes must respect overlap windows:

Vel-2 exists only during 13:00–19:00.

Vel-3 exists only during 19:30–23:30.

No post-T_off appearances of those copies unless a new cycle creates them.

Any written note that time-travels (Note-2/3/4) must have stable text history:

If it is updated during the overlap, the “arrived” version must already include those updates.

Failsafe causality must be clean:

Module-F1/Note-3 are returned by T_off3.

The installed protection that persists into C4 must be Module-F2 (the copy), not the arrived module.

No contradictory shutdown:

The prose cannot have Vel-1 “just unplug it” mid-cycle while an obligation item is still out, unless the plot explicitly shows the obligation being satisfied under the new earlier T_off.

The device’s role in each part must match the ledger:

Spoilage: discovery/verification.

Arbitrage: prediction → bet → winnings → note closure.

Compounding: safety bootstrapped → longer run → scaled profit → closure.

If you want, I can also add a one-page “object worldline diagram” (not prose) showing Apple-A / Note-2 / Note-4 as looped timelines with jump arrows, but the ledger above is already self-contained.

And then a story; the first version was obnoxiously formatted with nearly every sentence a paragraph, so I decided to intervene slightly to order GPT-5.2 Pro to fix that (the use of linebreaks here is extremely obnoxious. please fix the formatting to have more reasonable use of whitespace.). When I asked for a critique, it identified one logic error:

One concrete continuity flaw: the 19:00–19:25 segment double-books Vel‑1’s time unless you explicitly reorder or clarify what happened when.

First Story

Each fix is locally defensible and cumulatively fatal. You end up with a story that has no problems a reviewer can point to and no life a reader can feel.

Claude-4.6-opus (of this story)

The initial version:

Spoilage

We called them stasis fridges because the alternative was explaining a coil driver and a field geometry to people who wanted to know if their strawberries would last until Friday. It never cooled anything. If you put a cold six-pack inside, it stayed cold because it didn’t get a chance to warm up. If you put a hot burrito inside, it stayed hot because it didn’t get a chance to cool down. That was the pitch. Time was the insulation.

The prototype sat in the corner of Unit 12B between the parts racks and the folding table we used as a bench. A tall cabinet with a gasketed door and a latch that sounded too much like a freezer. Inside, it smelled like MDF, old silicone, and warm electronics. Room temperature, every time, unless you put something in already cold. The miswired unit happened because we were rushing. We’d swapped the driver board after a brownout and I’d re-terminated two bundles of coil leads at 2 a.m. under a headlamp. In the morning the field meter read “stable”, which was what we cared about.

The apple was supposed to be a clean demo. No cold air theatrics. Just: put it in, wait, take it out, show that it hadn’t browned. Enzymes don’t sell, but they’re honest. At 08:58 I killed my coffee and stared at the clock on my phone until it rolled over. I wrote on a Post-it and stuck it to the bench: NOTE-1: “Cycle test. Don’t improvise.”

At 08:59 I unlatched the door. The gasket peeled with a soft, sticky sound. Empty. Bare polymer racks, the little scratch I’d made with a screwdriver on the back wall, a strip of tape where we’d covered a seam. No apple. No prank. No object hidden in the corner. I ran my hand along the bottom, because I was tired and because tired makes you do rituals. Nothing. I shut it.

At 09:00 I hit the power strip. The driver fan came up with a dry whine. A relay clicked. The cabinet didn’t hum or glow or do anything cinematic. If you weren’t listening, you’d miss it.

I waited five minutes because five minutes was long enough for the field to stabilize, and because in a normal run you didn’t want to open the door right after power-on. We’d learned that the hard way when Mara—our only hire, our only adult—had yanked a lunch container out mid-field and complained her fork had felt “wrong”, like pushing through gel. I didn’t tell her about the miswire. I didn’t tell anyone anything.

At 09:05 I unlatched the door again. Same smell. Same stillness. But there was an apple on the middle rack. Green, slightly waxy, with a bruise on the side like someone had pressed a thumb too hard. I just stood there holding the door. My first thought wasn’t time travel. My first thought was that someone had been in the unit. My second thought was that the unit had been in someone else’s unit, in some other lab, and we’d gotten it mixed up.

Then the bruise resolved in my mind into detail. It wasn’t a thumb. It was an oval with a crescent notch, like a hard edge had caught it and then rolled away. I reached in and picked it up. Apple-A was room temp. Not even cool from sitting in plastic. Not warm either. Just… nothing had happened to it. My hands were warm; the apple didn’t steal heat. The cabinet didn’t give cold. It was a closet with a field on.

I set the apple on the bench and said out loud, to nobody, “This is new.” I glanced at NOTE-1 like it could argue back. There was a return obligation in the back of my mind already, even if I didn’t have words for it. The apple had been inside at 09:05 without me putting it there. The only way I didn’t go insane was by assuming it had come from some later moment. That meant whatever I did now had to complete a loop. It meant the apple had to go back.

I turned it slowly in my fingers, looking for stickers or ink. No sticker. No writing. Just the bruise. At 09:09 I took a photo with my phone, close enough that the bruise filled the frame.

At 09:10, fatigue did what fatigue does. I shifted my grip to rotate it and my thumb slid on the wax. Apple-A fell. It hit the bench edge exactly where a screw head was half proud because the cheap particleboard had stripped. A blunt thunk, more embarrassing than loud. I snatched it up like it was contraband.

The bruise was there. Still there. Same oval. Same crescent notch. The same mark I’d photographed a minute ago. I didn’t see it form. I couldn’t tell you, honestly, whether it deepened. But when I opened the photo and compared, the outline matched the screw like a template. The mark on the apple was a memory of an impact that, from the apple’s point of view, hadn’t happened yet. From mine, it had. From the cabinet’s, it was indifferent.

My mouth went dry. The room felt too bright. I looked at the screw head and then back at the apple and I had this impulse to cover the screw with tape, to refuse the bruise, to refuse the whole thing. But the bruise was already in my hand. If I refused it, it would still be there. That was the worst part.

I checked the clock again. 09:12. I had eight minutes to put the apple back. Nine if I wanted to flirt with rounding error. I cleaned the bench edge with my sleeve, not because it mattered, but because I needed to do something.

At 09:15 I opened the cabinet. Empty. Of course it was empty; the apple was in my hand. I placed Apple-A on the rack in the same orientation it had been when I’d found it, bruise facing out like an accusation. Then I pulled my hand back fast, like the cabinet might bite. I latched the door. The fan kept whining. I stared at the latch. I didn’t touch it again.

At 09:20, exactly, I cut power at the strip. The fan coasted down. The relay clicked the other way. The cabinet became a normal cabinet again, which was to say: a box. At 09:21 I opened it. Empty. No apple. No mark. No proof except the photo and the bruise in my head. I closed it gently. My hands were shaking, but not from adrenaline. From lack of sleep. From realizing that if I was right, I had just met the edge of a thing we were not funded to understand.

I found a dry erase marker and went to the board by the roll-up door where we tracked inventory and burn rate and the delivery schedule we were already missing. On the whiteboard I drew a straight line labeled “09:00 → 09:20” and a bent arrow that jumped back from the right end to the left. I wrote “Apple-A (bruise)” at the bend. Under it, I wrote, smaller, “Door open 09:05 (remove) / 09:15 (return).” Then I circled “09:20” like circling it would hold it in place. The board looked stupidly simple.

The cabinet didn’t. The field meter was still sitting on the bench where I’d left it, reading stable, as if stable was a moral position. My phone buzzed with a calendar reminder for a call with an investor at 10:00. I stared at it until it went quiet.

The apple loop had been a closed circle with no beginning. The bruise existed because it had to. And if an apple could do it, then information could do it, and if information could do it, then our runway could do it. That thought arrived with the same dull inevitability as the apple had. I didn’t say it out loud. I wrote a second Post-it and stuck it under NOTE-1. “Don’t tell Mara.” Then I went back to the cabinet and started planning a longer cycle.

Arbitrage

A timebox isn’t useful until you can carry something out of it. An apple is proof. A note is leverage. A person is a liability.

I didn’t have a partner in Unit 12B in the way people mean. Mara was payroll and safety glasses. Our lawyer was an email thread. The cabinet was mine because I was the one who slept on the floor next to it. So if I wanted a person to carry something out, I had to use myself.

I started numbering. Not because it made it less insane, but because it made it easier to not make a mistake. Vel-1 was me, the one who would keep going no matter how many loops we made. Vel-2 would be the copy who came out of the cabinet at the start of the next cycle and went back in at the end. That was the rule the apple had taught me. If it came out, it had to go back.

At 12:40 I pulled the rolling chair up to the bench and wrote a checklist in my notebook. Power-on: 13:00. Power-off: 19:00. Do not be at the door at 13:00–13:05. Do not interrupt power. Do not bring extra objects into the cabinet at 19:00.

I set an old webcam on the shelf facing the cabinet door. No cloud. No account. Just an SD card because I didn’t want a subpoena to find me by accident. I left an envelope in the top drawer of the bench with CASH-1 written on it in block letters. Three hundred in twenties. Not a lot of money. Enough to make a bet without looking like a joke. Enough to matter to me.

At 12:55 I checked the cabinet with the same ritual as the apple run. Door open. Empty. Door shut. Latch. I stood with my hand on the power strip and breathed until my pulse stopped trying to climb out of my throat. The thing about doing it on purpose is that your brain doesn’t give you the mercy of disbelief.

At 13:00 I flipped the switch. Fan. Relay. No cinematic anything. I didn’t look at the door. I walked away like it was a microwave. I went out into the little yard behind the unit and stood between the dumpsters where no one could see me through the front window. I watched the clock change on my phone. 13:01. 13:02. I could hear the faintest change in sound through the wall when the latch moved. Metal on metal. The cabinet opening.

My mouth filled with saliva like I was about to vomit. I stayed by the dumpsters. At 13:05 I went back inside and forced myself to do normal things. Answer a Slack message. Move a box of injection-molded trays from one shelf to another. Make tea because my hands needed something warm that wasn’t coffee. If Vel-2 existed, he would already be gone. If he wasn’t, then the whole thing was a fantasy and I could go back to pretending I was building a product.

At 14:00 Mara texted that she was running late because her train had stopped “for police activity.” I typed back “OK” and then deleted it and typed “OK” again, because I couldn’t remember if my first “OK” had gone through. That’s what the cabinet did to my brain. It made every trivial thing feel like a hinge.

At 18:55 I was back in Unit 12B. Not at the cabinet. Near the power strip, because the only promise I could keep was the off-time. The fan had been whining for almost six hours. Heat from the driver board had warmed the air in the cabinet a few degrees, but the cabinet itself didn’t breathe. It didn’t exchange. Nothing about it felt like a fridge. It felt like a lie we’d wrapped in a door.

I watched the webcam’s little status LED and listened. At 18:58 the latch clicked. Footsteps. Breathing. A soft grunt like someone moving carefully. I didn’t turn. I stared at the power strip like it was a detonator. The sound of paper sliding on plastic. A pause. Another breath. Then, quieter, the squeak of the rack as weight shifted. Vel-2 climbing in. At 18:59 the latch clicked again. The cabinet was closed.

At 19:00 I killed power. Relay. Fan dying. Silence. I waited a full minute before I moved. The minute was for superstition, not physics. At 19:01 I opened the cabinet. Empty. No person. No note. No evidence that any of the sounds I’d heard had been real.

In the top drawer, under the solder spool, was a second envelope. CASH-2 in my own handwriting. Thicker. I didn’t open it right away. I sat on the floor with my back against the cabinet and stared at it until my eyes hurt. When I finally tore it open, it was cash, warm from being in a drawer all day. A lot more than three hundred. There was also a folded betting slip—BetSlip-1—printed on glossy paper with a timestamp at 13:16. The kind you get from a kiosk when you don’t want a digital trail. I’d never touched it.

Next to it was a plain piece of paper torn from my notebook, not time-traveled because it wasn’t inside the cabinet at any power-off, just left like any other thing you leave for yourself when you don’t trust yourself. Three lines. “Placed 13:15 with Cash-1.” “Game ended 18:40. Note matched.” “Cashed 18:55. Left you Cash-2. Don’t come to door.” No signature. But the pen pressure was mine. The phrasing was mine. The sting of recognition was physical.

I watched the webcam footage on my laptop with the volume off. At 13:02 the cabinet door opened. Vel-2 stepped out. Same hoodie. Same half-grown beard. Same posture of someone who’d slept badly. He held a small folded paper—Note-2—and he didn’t look at the camera. He went straight to the drawer, took the CASH-1 envelope, and left. At 18:59 he came back. He looked thinner. He put Note-2 into the cabinet, climbed in, and shut the door behind him.

It was exactly as clinical as the apple. No drama. No warning. Just a human being treating himself like an object that needed to be returned by a deadline.

I didn’t keep the footage on the laptop. I copied it to a thumb drive and hid it in the ceiling tile above the bathroom because paranoia is a kind of maintenance. Then I took the marker to the board by the roll-up door again and drew another simple diagram that felt like an insult to how sick I was. On the whiteboard I wrote “13:00 ON / 19:00 OFF” and sketched a box in the middle. I drew two stick figures: “Vel-1” on the outside line and “Vel-2” as a loop that started at 19:00, jumped back, and ended at 19:00 again. I wrote “Note-2” next to it. I underlined “Return obligation” twice. When I capped the marker, my hand cramped.

I thought about how we were supposed to be building a consumer appliance. How we’d fought about door hinges and gasket suppliers and whether the UI should be a touch panel or a knob because “people like knobs.” Now I had a cabinet that could print money by rewriting causality, and the only thing keeping it from burning down our unit was a cheap power strip and my ability to not blink at the wrong time.

Mara texted again at 19:10. “Still here?” I typed “yeah” and then I deleted it. I didn’t want my phone to be a record of anything. Instead I put it face down and went back to the cabinet. If I was going to keep doing this, I needed a way to survive it. Not morally. Electrically.

Compounding

The danger wasn’t paradox. The danger was a breaker tripping in the middle of a cycle. A landlord flipping the wrong switch. A driver board overheating and crowbarring itself. If power-off defined the end of the cycle, then an accidental power-off redefined it, and anything you’d taken out would still owe you a return you could no longer pay. That was how you died. Not with a grandfather. With a missed deadline.

At 19:03 I rearranged the bench like I was prepping for surgery. Soldering iron in its stand. Multimeter leads untangled. A tray of components pulled from stock: relays, thermistors, a little microcontroller we used for the normal stasis mode timing. A clean notebook page ready for whatever came out of the cabinet.

At 19:05 I powered on. Fan. Relay. My skin had learned the sound and reacted before my mind did. At 19:06 I opened the door. The cabinet held a small circuit board in an anti-static bag and a folded sheet of paper. Module-F1 and Note-3, sitting on the middle rack like they’d always been there. I didn’t touch them for a full second. My brain tried to find an origin story. A courier. Mara. A forgotten shipment. But the apple had burned that reflex out. If it was inside at 19:25, it was here now. And if I took it out, it had to be back in by 19:25.

I grabbed the anti-static bag and the folded paper and shut the door. On the bench I opened Note-3 first. It was a wiring diagram in my own shorthand. Pin labels I used when I didn’t want to think. A threshold value circled three times. A test procedure: “Heat sensor to 60℃; verify LIMIT, not OFF.” That line mattered. A failsafe that cut power would kill the cycle. A failsafe that limited the driver while keeping the cabinet logically “on” would keep the worldline intact. Module-F1 in the bag looked like exactly that. A current limiter with a thermal input and a latching relay that bypassed the limiter only at startup. It was the kind of board I would’ve designed if I’d had time. I didn’t have time. I had nineteen minutes.

From 19:07 to 19:18 I worked with the frantic calm you get when you’ve accepted that the deadline isn’t negotiable. I traced Module-F1’s layout, checked the BOM scribbled on the back of Note-3, and pulled parts from our bins. I wasn’t building something elegant. I was building something identical enough that it would function. Module-F2 was a copy with slightly uglier solder joints.

At 19:18 I cut power to the driver board only—leaving the cabinet powered and the cycle running—by flipping the service toggle we used for maintenance. The fan stayed on. The relay in the cabinet stayed engaged. The cycle’s “on” state stayed true. Then I spliced Module-F2 inline on the coil feed. At 19:22 it was installed.

At 19:23 I clipped a heat gun to the bench vise and aimed it at the thermistor. I watched the multimeter and the LED on Module-F2. At 60℃ the LED changed. The driver current dropped. The fan kept running. The cabinet stayed on. No relay click. No power-off. That was what Note-3 had promised. That was what Module-F1 must have been built to ensure. I turned off the heat gun and let my lungs remember how to breathe.

At 19:24 I opened the cabinet and returned Module-F1 and Note-3 to the rack. I positioned them exactly as I’d found them, because I’d learned to respect stupid details. Door shut. Latch. At 19:25 I powered off the cabinet. Fan dying. Relay click. Silence. At 19:26 I opened the cabinet. Empty. The originals were gone, displaced back to 19:05, waiting to be taken out by the version of me who’d be there at 19:06. Bootstrap doesn’t feel like magic when you’re holding a soldering iron. It feels like theft.

At 19:27 I took CASH-2 from the envelope Vel-2 had left and counted it twice. Then I put it in a new envelope and wrote “For Vel-3” on it, and I put it in the top drawer where Vel-2 had found Cash-1. I didn’t like writing that. I didn’t like that there was now a Vel-3. But the cabinet didn’t care about what I liked.

The long cycle had to start at 19:30 if we wanted a full evening window. That gave me three minutes to become the kind of person who could do this twice in one day. At 19:30 I powered on again. Fan. Relay. The cabinet was back in its indifferent state. I stayed away from the door. Not out of superstition, but out of strategy. If Vel-3 existed, he would come out at 19:31 with whatever Note-4 told him to do. If he saw me, he might talk. If he talked, I might answer. If I answered, I might say something I couldn’t unsay.

At 19:31, from across the unit, I heard the latch. A soft click. A pause. The door opening. Footsteps. The drawer sliding. The envelope rustling. Then the cabinet door shutting again. Vel-3 leaving. I didn’t move until 19:35. When I finally went near the cabinet, it was latched and humming quietly with the fan. Nothing to see. Just a powered box.

The rest of the evening was waiting. Module-F2 sat on the coil feed with its LED steady. I watched it the way you watch a sleeping animal you don’t trust. I didn’t touch my phone. I didn’t open the cabinet. I worked on normal things. Wrote an email draft to the manufacturer about gasket tolerances. Updated a spreadsheet with shipping costs. Anything to keep my hands from drifting back toward the cabinet like a nervous tic.

At 22:50 I heard the roll-up door at the front of the unit rattle. Not open. Just a vibration, like someone had leaned on it. A reminder that the world still existed outside the loop.

At 23:10 there was a soft knock on the side door from the alley. Three taps, spaced evenly. I froze. Then I realized, with a nausea that felt like hunger, that the knock wasn’t asking to be let in. It was a signal. I waited a full minute and then cracked the side door. No one. Just the smell of asphalt and the hum of the freeway. On the ground, tucked against the threshold, was a small lockbox we used for petty cash. It hadn’t been there an hour ago. The lid wasn’t locked. Inside was an envelope: CASH-3. And a page torn from my notebook with a single line and a set of numbers. “Code’s in your book. Don’t look up until after.”

I went back inside and sat at the bench, staring at my own notebook like it was radioactive. In the margin of today’s page, in handwriting that was mine but slightly slanted, a sequence: “23:10—4-9-2.” Under it: “If you’re reading this, you made it clean. If you keep going, don’t pretend it’s still about food.” That was all. No explanation of the trades. No confession. Just the moral weight dropped like a tool onto a table.

At 23:25 I took my place by the power strip again. The cabinet still hummed. Module-F2’s LED stayed steady. At 23:28 the latch clicked. Vel-3 was back. I didn’t turn. Paper sliding on plastic. A breath. A grunt as weight shifted. Vel-3 climbing into the cabinet. At 23:30 I killed power. Relay. Fan dying. Silence. And then, as always, the sense that something fundamental had happened in a place my senses couldn’t reach.

At 23:31 I opened the cabinet. Empty. No Vel. No Note-4. No proof. Just a warm plastic rack and the smell of the unit. I latched it again and leaned my forehead against the door. My reflection in the dark plastic looked like somebody else’s face.

CASH-3 sat on the bench in front of me, more money than we’d ever had that wasn’t debt. I didn’t open it. I didn’t need to. Money is money. The part that mattered was the sentence in my notebook. If you keep going, don’t pretend it’s still about food.

I walked to the board by the roll-up door for the last time that night, because my brain needed to see it in lines and arrows instead of in my bloodstream. On the whiteboard I drew the cabinet as a square and wrote “19:30 ON → 23:30 OFF” above it. I added a small box labeled “Module-F2 (LIMIT, not OFF)” on the power line. I drew Vel-3 as a loop that left at 19:31 and returned at 23:28 with “Note-4” and then jumped back at 23:30. Off to the side I wrote “Cash stays out” and underlined it. Then I capped the marker and stared at the diagram until it stopped making sense.

My phone buzzed, face down. A call. Mara. I let it ring out. A second buzz. A text. “QA found something. One of the production units holds temp even when unplugged. Like… too well. Can you look tomorrow before we ship?”

I read it twice. Holds temp even when unplugged. A unit that stayed “on” without power, in the only way the cabinet ever cared about. I thought about customers putting hot soup inside and walking away. I thought about a child opening a door and taking something out that had “arrived”, not knowing it owed itself back by a deadline. I thought about a world full of cabinets making obligations nobody understood.

CASH-3 sat on the bench. The cabinet sat in the corner, quiet, innocent, empty. The sentence in my notebook didn’t tell me to stop. It didn’t tell me to continue. It just stripped the cover story off.

I typed a reply to Mara. “Don’t ship. Tell them a fuse issue. I’ll be in early.” My thumb hovered over send. Behind my eyes, the cabinet’s simple rule pulsed like a metronome: anything you take out, you have to put back. I looked at the power strip. I looked at the time. I looked at the cabinet door and imagined opening it at some future T_on to find not an apple, not a note, not a copy of myself, but someone else. Someone who hadn’t agreed. Someone whose obligation would be mine the moment I touched the latch. I didn’t hit send. I didn’t not hit send. I just sat there until the phone went dark again, and the only sound in Unit 12B was the faint ticking of a cheap wall clock, counting down toward whatever I’d already decided to do.

I then began running this through the usual gauntlet of iterative editing prompts, using not just GPT/Claude but Gemini-3-pro-preview, Kimi K2.5 Thinking, and DeepSeek-v3 for kicks. The “Poetry magazine” prompt was revised to be a “Clarkesworld Magazine” prompt (leading to the amusing issue that the LLMs would sometimes warn that Clarke’s World has strictly banned AI-related submissions and to be sure that this story wasn’t AI-written in any way). I avoided doing any curation or feedback on the content, and tried to confine myself to simply copy-pasting critiques and suggestions back and forth, using GPT-5.2 Pro as the master session. (I would have used Claude more to test it out, but the usage quota for Claude-4.6-opus on maximum thinking settings is low.)

An example:

…Report:

I’ll review this science fiction story for you. This is a substantial piece, so let me analyze it thoroughly.

Summary

“Spoilage” is a hard science fiction story about a startup that accidentally invents time travel while building “stasis fridges”—devices that preserve food not by cooling it, but by stopping time inside them. The narrator, an engineer at a failing startup, discovers that a miswired prototype creates closed timelike curves: objects placed inside during the “off” state can be retrieved from the “on” state, but must be returned before power-down or reality enforces the balance with lethal consequences. When production units inherit the same dangerous property—remaining “logically ON” even when unplugged—the narrator and his colleague Mara must stop a shipment before customers start disappearing into temporal debt collection.

Analysis

Strengths

Engineering Authenticity The story excels at technical verisimilitude. The narrator’s voice is pure exhausted startup engineer—jargon-heavy, ethically compromised, sleep-deprived. Details like “conformal-coated,” “fiducial,” “brownout handling,” and the specific horror of “eleven weeks from payroll insolvency” ground the fantastic element in mundane reality. The prose treats time travel as a bug, not a feature, which is refreshing.

Conceptual Rigour The time mechanics are internally consistent and genuinely interesting. The “bookkeeping” metaphor—“no net change across a cycle”—extends naturally to personal stakes: “IF OUT → MUST RETURN.” The horror isn’t paradox (grandfather killed) but accounting (debt collection). The quarter “half-sunk into the polymer rack” and the badge fused into the shelf are visceral images of space-time’s indifferent reconciliation.

Structural Control The three-act structure (Spoilage/Arbitrage/Compounding) mirrors both financial terminology and escalating personal investment. Each section builds logically: discovery of the effect → exploitation for engineering solutions → catastrophic realization of commercial scale. The pacing accelerates appropriately.

Character Work Mara emerges effectively from minimal brushstrokes—“payroll and safety glasses and the only one of us who remembered to eat lunch on purpose.” Her arc from colleague to accomplice to moral anchor works. The narrator’s self-loathing is palpable: “the small, ugly thing because small, ugly things are easier than big confessions.”

Weaknesses

Prose Density The technical precision sometimes suffocates emotional beats. Sentences like “The controller never wanted to see a hard OFF during a run. It dithered, throttled, blinked coil current—anything but admit the field had stopped” are strong, but the story rarely lets itself breathe. The narrator’s exhaustion is told through sleep deprivation references but rarely shown through altered perception or physical deterioration beyond shaking hands.

The Bootstrap Paradox The Module-F1/F2 sequence, while clever, relies on unexamined bootstrap paradox mechanics. Where did the original Module-F1 come from? The story gestures at “theft of origin” but doesn’t grapple with the ontological weirdness of information without source. This is forgivable in a short piece, but the narrator’s lack of curiosity here feels convenient rather than character-driven.

Mara’s Conversion Her acceptance of the impossible happens too easily. After the badge demonstration, she shifts from skeptic to accomplice with minimal resistance. The line “That is not a sentence a sane person says about a refrigerator” is strong, but her subsequent collaboration feels engineered for plot momentum rather than psychological truth. A moment of genuine rejection or escape attempt would strengthen the final act.

The Ending The story stops rather than concludes. “DEFAULT” circled on the whiteboard is a strong final image, but the stakes remain abstract. We don’t see the consequences of the quarantine—regulatory, financial, existential. The narrator’s arc completes (he acts responsibly), but the narrative promise of “a deadline was moving toward us” remains unfulfilled. This may be intentional (horror of the unknown), but it reads slightly unfinished.

Critical Assessment

Theme: The story explores technological ethics through the lens of startup culture’s “move fast and break things” mentality. The fridge becomes a metaphor for technical debt—problems deferred until they become catastrophes. The “bookkeeping” motif suggests capitalism’s tendency to abstract human cost into accounting entries.

Genre Positioning: This is “hard SF horror” in the tradition of Greg Egan or early Chiang, though less philosophically ambitious than the latter. The horror is cosmic indifference filtered through OSHA violations and supply chain logistics.

Market Fit for Clarkesworld: Strong. The magazine publishes technically rigorous SF with literary sensibilities. The story’s length (approximately 7,500 words) is appropriate. The engineering perspective and economic anxiety align with contemporary SF trends.

Recommendation

ACCEPT with minor revisions or ACCEPT as-is

The story’s conceptual strength and technical execution outweigh its structural issues. For improvement, consider:

Mara’s skepticism: Extend her resistance through one more beat—perhaps she leaves, then returns with evidence that forces her belief.

The ending: Either commit to the abrupt stop (horror of incompleteness) or add 200-300 words showing the immediate aftermath—regulators arriving, the narrator’s first night knowing he’s created something that can’t be uncreated.

Sensory texture: One scene of genuine hallucination or perceptual distortion would validate the narrator’s exhaustion more effectively than repeated references to sleep deprivation.

The core image—the quarter half-sunk in plastic, the badge fused into the shelf—is indelible. The story understands that the most disturbing time travel is not adventure but accounting, and that the universe’s indifference to human meaning is scarier than any monster. This is accomplished, disturbing work that deserves publication.

Our reviewers recommended revise-and-resubmit. To help the author, please write out a detailed peer review report of the story; start at the global level and work your way down to the line level, with detailed suggestions. If suggesting a change, try to consider at least 3 options for each change.

[after including several reviews]

Story Summary

A startup engineer builds “stasis fridges” that don’t cool—time is the insulation.

The product spec guarantees “no net change across a cycle”, and the firmware tries hard to avoid an actual OFF event.

After minor anomalies (a fork feeling like it’s pushing through gel; a quarter half-embedded into a polymer shelf after a power sag), a miswired dev unit produces the first clean proof: an apple appears mid-cycle already bruised in the exact shape of a screw it hasn’t hit yet.

The engineer realizes the cabinet is enforcing endpoint consistency, not preventing events.

If something is “inside” at OFF, it must have been inside during ON; if something is removed, it must be returned—or the interval will reconcile the books some other way.

To get leverage and control, he deliberately creates a timeboxed “Two”: a future/past version of himself who can step out near the beginning and return near the end.

Two brings back critical information about production boards: “unplugged” does not mean OFF, and a door-sensor failsafe can force a hidden OFF that closes the ledger on its own schedule.

The engineer uses a bootstrap loop (a module that appears, is copied, and is sent back) to regain control of OFF-like behavior and reduce accidental sag deaths.

Then Mara arrives with a boxed production unit from QA—one that won’t die when unplugged.

On the “third open”, it forces a hard end, and the reconciliation produces horror: a contract manufacturer’s badge appears and then embeds itself into the rack like the quarter.

The implication is human-scale: someone’s labor—and possibly their body—was used as substitute collateral when a loop didn’t get paid back.

They try to halt shipments, knowing the only safe “default” is never letting these units decide OFF for themselves.

The last image is a ticking clock and a product that will eventually reconcile its debts in someone’s kitchen.

Analysis of What’s Already Strong and What’s Unstable

What’s Already Working

Voice and texture. The exhausted engineer POV is consistent, competent, and concrete.

Procedural horror. The dread comes from checklists, firmware behaviors, and power events, not from spectacle.

Core rule is generative. “No net change across a cycle” turning into literal endpoint enforcement is elegant and unusually fresh.

Anchor images. The quarter embedded “perfectly level” is a perfect physical impossibility.

What Currently Destabilizes the Read (And Why Reviewers Converged)

Two conceptual engines are adjacent but not causally fused. The bootstrap module scene reads like clever SF; the badge reads like cosmic debt collection.

The story wants those to be the same phenomenon (borrowing causes interest), but it currently feels like “we did a neat paradox” and then “separately, horror.”

OFF vocabulary is doing too much work. You have at least three OFFs (mains removal, coil current drop, controller state transition), plus “door defines inside.”

If a reader gets even slightly lost here, tension bleeds away because the fear is “I can accidentally close the ledger”, and that requires crisp boundaries.

Mara’s turn is too fast. She’s effective as pressure and conscience, but she needs one more beat of refusal/skepticism/agency so the final alliance is earned rather than convenient.

Hard-SF reader snag: temporary duplication mass/energy. If Two exists outside for hours, something should “pay” (power draw, instability, exchange, or a stated violation that becomes part of the horror).

Right now, the story’s own metaphor says “bookkeeping”, so readers expect to see the bill.

Ending commitment. “DEFAULT” is a good poetic stop, but reviewers are right that you need one extra concrete commitment or consequence so the ending lands as tragedy, not “near miss.”

Reviewer Suggestions Ranked by Usefulness

Rank

Suggestion (compressed)

Why it’s high leverage

Sources

1

Standardize ON/OFF terms (define “hard OFF” that closes the books; separate from unplug/coil)

Clarifies the governing rule; prevents confusion from stealing dread

R1, R2, Final

2

Integrate bootstrap with debt collection (bootstrap is borrowing; interest is instability/harm; ending becomes inevitable)

Turns “two cool ideas” into one engine; makes climax feel earned

R1, Final

3

Give Two a visible cost (power draw, bodily instability, marks, or ethically compromising specifics)

Pays the physics bill; deepens horror; makes “arbitrage” feel like theft

R2, R1, Final

4

Earn Mara’s conversion / agency (skepticism + refusal beat + decisive action)

Fixes the biggest character pacing complaint; raises stakes via conflict

R1, R2, R3, Final

5

Seed cross-unit coupling early (shared QA sync frame / phase-lock)

Makes badge escalation rule-driven rather than twisty

Final

6

Slow and humanize the badge moment (name/photo/rocket; breathing room; connect to quarter/gel)

Gives the climax human weight; makes the “victim” real

R1, R2, R3, Final

7

Make the resolution harder / more costly (distributor pushback; explicit payroll/company sacrifice; external report)

Upgrades ending from “responsible step” to “irreversible choice”

R1, R2, Final

8

Clarify the apple bruise geometry

Removes a key visualization snag in the first proof

R1

9

Motif management (corrupt repetitions; vary refrains)

Polishes voice and keeps repetition from feeling mannered

R1, Final

10

Line-level clarity tweaks (“blink coil current”, syntax simplifications)

Worth doing, but subordinate to rule/character/ending fixes

R1, R2, R3

Below is a revision that implements the top items without changing the story’s identity: same procedural dread, same timestamps, but tighter rule vocabulary, clearer physics “billing”, stronger Mara agency, explicit cross-unit seeding, and an ending that costs something. …

Rank	Suggestion (compressed)	Why it’s high leverage	Sources
1	Standardize ON/OFF terms (define “hard OFF” that closes the books; separate from unplug/coil)	Clarifies the governing rule; prevents confusion from stealing dread	R1, R2, Final
2	Integrate bootstrap with debt collection (bootstrap is borrowing; interest is instability/harm; ending becomes inevitable)	Turns “two cool ideas” into one engine; makes climax feel earned	R1, Final
3	Give Two a visible cost (power draw, bodily instability, marks, or ethically compromising specifics)	Pays the physics bill; deepens horror; makes “arbitrage” feel like theft	R2, R1, Final
4	Earn Mara’s conversion / agency (skepticism + refusal beat + decisive action)	Fixes the biggest character pacing complaint; raises stakes via conflict	R1, R2, R3, Final
5	Seed cross-unit coupling early (shared QA sync frame / phase-lock)	Makes badge escalation rule-driven rather than twisty	Final
6	Slow and humanize the badge moment (name/photo/rocket; breathing room; connect to quarter/gel)	Gives the climax human weight; makes the “victim” real	R1, R2, R3, Final
7	Make the resolution harder / more costly (distributor pushback; explicit payroll/company sacrifice; external report)	Upgrades ending from “responsible step” to “irreversible choice”	R1, R2, Final
8	Clarify the apple bruise geometry	Removes a key visualization snag in the first proof	R1
9	Motif management (corrupt repetitions; vary refrains)	Polishes voice and keeps repetition from feeling mannered	R1, Final
10	Line-level clarity tweaks (“blink coil current”, syntax simplifications)	Worth doing, but subordinate to rule/character/ending fixes	R1, R2, R3

The final version was terrible; possibly millions of tokens, wasted to simply spoil a workable first draft:

Spoilage

We called them stasis fridges because the alternative was explaining a coil driver and field geometry to people who wanted to know if their strawberries would last until Friday. It never cooled anything; if you put a hot burrito inside, it stayed hot because it didn’t get a chance to cool down. Time was the insulation. That wasn’t a miracle. It was the kind of trick you can sell with a warranty clause.

In the shipping spec, “stasis” meant no net change across a cycle—not “nothing happens”, just “the totals match at the endpoints.” We pretended OFF was a single word until it started showing up as different events in failure reports, depending on whether you meant the wall power, the coils, or the controller admitting the run was over. The controller never wanted to see that last one during a run. It dithered, throttled, stuttered the coil current in millisecond bursts—anything but register the field’s collapse. If you asked me what the field did to you while it was on, I’d have said “nothing” with a straight face.

Then Mara pulled her lunch container out too fast and said her fork felt wrong, like pushing through uncured epoxy with grain. Afterward the fork’s tines weren’t quite parallel—maybe a millimeter out at the tips, a twist you’d expect from torque, not from bending. We blamed cheap plastic and stress and not sleeping, and went back to work because that’s what you do when payroll is a countdown clock.

A week later, during an overnight soak test, the building breaker hiccuped. Not a clean OFF—just a sag, a blink—the kind of thing a decent board should ride through. It didn’t.

We kept a quarter on the middle rack as a dumb fiducial—something you can point at when you’re arguing whether a graph is noise. After the blink it was half-sunk into the polymer rack, face up, perfectly level, like the plastic had yielded around it and then decided it had always been that way. The reeded edge vanished into the rack teeth in a clean line: no melt, no seam, just an interface that shouldn’t exist.

We took photos, bagged the rack, wrote POWER SAG / DRIVER GLITCH on the label because labels are cheaper than conclusions. Then we told ourselves production wouldn’t ever see that failure mode because production had a clean driver revision and proper brownout handling, and because we needed there to be a line between the dev mule and what we were about to ship.

Production also had a “feature” we’d been proud of on a slide deck: a latching regulator and a fat hold-up bank so the logic wouldn’t brown out when someone yanked a cord and jammed it back in. Faster reboot, better logs, fewer angry emails; Mara had championed that slide, back when “resilience” was the closest thing we had to virtue. And because QA hated drift, production units listened for a reference pulse during calibration—phase-lock, fleet sync, whatever name you put on it when you don’t want to say “shared frame.” Cleaner graphs, less privacy than we assumed.

That was also when I added the keyed service toggle on the driver panel: you could cut coil current without the controller logging the event it treated as a cycle end. Officially it was for safety. Unofficially it was so you could work inside the machine without the firmware declaring the run was over.

The dev mule sat in Unit 12B between the parts racks and the folding table we used as a bench. In there, the cabinet smelled like MDF, old silicone, and warm dust on heatsinks, and the gasket peeled with that soft, sticky freezer sound when you unlatched it. If you crouched you could see a tiny logic LED through the vent slots—green when the controller still believed it was running, dark when it had finally admitted OFF. After the quarter incident I’d kludged in two supercaps and a diode so the logic could ride a blink without losing state; when you cut mains on purpose, they drained in seconds. Enough to make “momentary” mean something. Not enough to hide a shutdown from you if you were watching.

The miswired unit happened because we were rushing. We’d swapped the driver board after a brownout and I’d re-terminated two bundles of coil leads at 2 a.m. under a headlamp, and in the morning the field meter read stable, which was what we cared about. Stable, as if stable were a moral position.

The apple was supposed to be a clean demo. No cold air theatrics; just put it in, wait, take it out, show it hadn’t browned. Enzymes don’t sell, but they’re honest.

At 08:58 I drained my coffee and stared at the clock on my phone until it rolled over. I wrote on a Post-it and stuck it to the bench: NOTE-1: Cycle test. Don’t improvise. At 08:59 I unlatched the door. Empty.

At 09:00 I hit the power strip. The driver fan came up with a dry whine and a relay clicked—no glow, no hum, nothing cinematic—and if you weren’t listening you’d miss it. At 09:05 I unlatched the door again. Same smell, same stillness, but there was an apple on the middle rack: green, slightly waxy, room temperature in the way only untouched things are, as if it hadn’t exchanged heat with anything since whenever it had last been real. A bruise sat on one side like someone had pressed a thumb too hard.

I just stood there holding the door. My first thought wasn’t time travel; my first thought was that someone had been in the unit. The bruise was obvious, but what took a second was its geometry: not a smear, not a soft dent, a crescent bite out of an oval, concave side down, like the apple had kissed a curved edge and then rolled.

I picked it up and my eyes snagged on a screw head half-proud at the bench edge where the cheap particleboard had stripped. The torn Phillips slot had a burr—a little crescent of raised metal where the driver had chewed it—and the bruise on the apple was the same curve. Not “kind of”, not “close enough”: same radius, same orientation, like the apple had been stamped with that screw’s damage.

I set the apple down and said out loud, to nobody, “This is new.” I glanced at NOTE-1 like it could argue back, and what came next wasn’t a theory so much as the only assumption that kept my brain from snapping. The miswire hadn’t made the effect stronger. It had made the interval closed. Before, ON-to-OFF was a line; now it was a loop, and the controller was enforcing the endpoints as hard boundaries. The apple wasn’t a gift; it was a balancing entry.

At 09:09 I took a photo with my phone, close enough that the bruise filled the frame. At 09:10 I grabbed a strip of tape and covered the screw, trying to deny the bruise its cause, and the tape stuck to my thumb the way tape always does when you need it not to. The apple shifted in my other hand—wax on skin, a slide—and then a blunt thunk as it kissed the screw anyway, exactly where the bruise said it would. It wasn’t loud; it was definitive.

I snatched it up and turned it over like there was some other explanation hiding on the far side. The bruise was still there, same oval, same crescent bite, and when I opened the photo and compared the outline matched within the thickness of my hand shaking. The bruise wasn’t a prediction; it was a receipt.

At 09:15 I opened the cabinet. Empty, of course, because the apple was in my hand. I placed it on the rack in the same orientation it had been when I’d found it, bruise facing out like an accusation, and pulled my hand back fast like the cabinet might bite.

At 09:20 I cut mains at the strip. The fan coasted down, and through the vent I watched the little green logic LED fade and die a second later. That was the controller admitting it: hard OFF. At 09:21 I opened the door. Empty. No apple, no proof except the photo and the bruise in my head, and the creeping sense that the quarter in the rack hadn’t been “damage” so much as the universe reconciling a mismatch with whatever compliance it could borrow.

On the whiteboard by the roll-up door I drew a straight line labeled 09:00 → 09:20 and a bent arrow that jumped back from the right end to the left. I wrote Apple-A at the bend and under it: Door open 09:05 (remove) / 09:15 (return). Then I wrote three labels beside the line because I could feel language failing: MAINS OFF. COIL OFF. HARD OFF (books close). Marketing pretended those were synonyms, and I circled HARD OFF until the marker squeaked.

My phone buzzed with a calendar reminder for an investor call at 10:00. I didn’t take it. At 10:15 it buzzed again—voicemail—and I watched it go quiet because answering would have required words like “apple” and “ledger” and “no”, and I didn’t have any of those ready.

From 09:30 to noon I tried to disprove myself without giving the cabinet another clean loop to bite into. I checked wiring twice, traced coil bundles, re-labeled what I’d already labeled, and stared into the empty racks until my eyes watered. I kept thinking about the quarter and the fork and the fact that a human body had a lot more coordinates than a quarter, and a lot more to lose if the books ever closed without a way to pay.

I wrote a second Post-it and stuck it under NOTE-1: Don’t tell Mara until you can define HARD OFF. Then I sat on the rolling chair and looked at the cabinet like it was a problem set with a deadline. If I was going to test this at human scale, I needed help I didn’t have, so I decided to borrow it from myself.

Arbitrage

A timebox isn’t useful until you can carry something out of it. An apple is proof, but a person is leverage because a person can do work, and that was the first honest thought I had and I hated myself for how quickly it felt like engineering.

I started numbering in my notebook the way you number versions when you don’t trust memory. One was me: the operator, the one who stayed on the outside line. Two would be the me whose worldline kinked through the cabinet—out near the beginning only because he would climb back in near the end—and from Two’s perspective there was no after, only return.

I wrote a rule in block caps: IF OUT → MUST RETURN. My hand hovered over the paper longer than it should’ve, and a wave of nausea came up like my body had understood the rule before my brain had.

At 12:40 I wrote a checklist. Power-on: 13:00. Hard OFF: 19:00. Do not be at the door at 13:00–13:05. Do not interrupt power. Service toggle only if you have to. Do not let the controller see HARD OFF early. I set an old webcam on the shelf facing the cabinet door—no cloud, no account, just an SD card because I didn’t want a subpoena to find me by accident—and I left an envelope in the drawer with CASH-1 written on it.

At 12:55 I checked the cabinet. Empty.

At 13:00 I flipped the switch. The fan came up, the relay clicked, and the lights in Unit 12B dipped—just a blink, like the building took a shallow breath. From the breaker panel by the door came a faint, fast ticking, metal warming under load, and I couldn’t stop myself from thinking of it as the world charging interest for the extra me I was about to insist on.

I didn’t look at the door. I walked outside and watched the clock on my phone tick over 13:01, 13:02, and through the wall I heard the latch: metal on metal, the cabinet opening. I stayed where I was anyway. If Two existed, he’d already be gone; if he didn’t, I could still stop and pretend this was sleep deprivation and not a thing my brain had words for.

At 15:12 Mara arrived with two coffees. I slid NOTE-1 and the Post-its under the solder spool, face down, and she took one look at me and said, “You look like hell.”

“I’m fine”, I said.

She glanced at the cabinet. “Running a soak?”

“Thermal drift”, I said too fast.

Her eyes went to the webcam. “Why is there a camera on a drift test?”

“Door sensor bounce”, I said. “Frame rate.”

She didn’t buy it. She stepped closer anyway, just to listen, and then she said, quietly, “I don’t like standing near it when it’s running.”

“I know”, I said, and I meant it in more than one way.

She set her coffee down. “You remember why I pushed the hold-up bank”, she said. “When we lose logs, we lose fights. Logs are our alibi.” Then, like she couldn’t stop herself from being practical even when she was uneasy: “And breakers in this row are junk.”

“I know.”

She tried to make it normal after that—sensor drift, door magnet, me seeing patterns because I’d been sleeping on the floor—but she watched my face too long and then went to wash her hands like she wanted something real under running water.

Around 16:40 the overhead lights flickered and the cabinet fan dipped, and my stomach climbed into my throat. On the second flicker I turned the service key just long enough to dump coil load while the controller rode its little hold-up, and the fan pitch steadied again like nothing had happened.

At 18:58 the latch clicked. Footsteps, a soft grunt like someone moving carefully, paper sliding on plastic, and then the squeak of the rack as weight shifted: Two climbing in.

At 19:00 I killed mains. I waited, watching through the vent until the green logic LED winked out—hard OFF—and only then, at 19:01, did I open the cabinet. Empty. No person and no note and no evidence that any of the sounds I’d heard had been real.

The evidence was on the bench instead: a new anti-static bag with REV D DRIVER scrawled across it in my handwriting, and a crumpled QA trace that smelled faintly of disinfectant and machine oil. Someone had circled LOGIC RAIL HOLDS AFTER AC DISCONNECT and, beneath it, written COIL GATE LATCHES / PHASE-LOCK ENABLED.

In the drawer, CASH-1 was lighter. There was a receipt from the surplus store, and a candy bar still wrapped, crushed flat like it had been in a pocket and forgotten. I held it for a second, felt the soft give through the wrapper, then set it back down unopened, like it belonged in the chain of custody.

There was also a torn notebook page in my pressure, my half-sentences: Rev D does not go dead when you think it does; unplugged kills mains, not state; hard OFF is a decision—firmware or brownout or door failsafe. Door counter doesn’t reset on hold-up. Third open triggers forced shut-down. It clears the books.

Then the part that made my skin go tight: units phase-lock to the same reference pulse; it’s one ledger when they’re aligned; when one closes it can settle against another. Saw a rack like ours in QA. Something half-in, face up. They called it a molding defect and moved on.

Under the last line was a dark smear dried into the fibers, and then an afterthought: I’m hungry and I can’t make myself eat.

I pulled the video. At 13:02 Two stepped out in the same hoodie and beard, too pale, with a thin line of blood from one nostril he didn’t wipe. He looked into the camera—not long, not a signal, just a check, like he needed this version of me to witness what he was about to spend. At 18:59 he came back thinner. Not “lost weight” thinner. Wrong. His shoulders sat wrong, like he’d been carrying something heavy and invisible, and he placed a folded note on the rack before he climbed in.

The rule held. The cost had arrived anyway.

I wrote on the whiteboard because my brain needed somewhere to pin the new rules: UNPLUGGED ≠ HARD OFF. PHASE-LOCK ≠ PRIVATE. Then, because the note and the module and my own handwriting forced the thought whether I wanted it or not, I wrote BOOTSTRAP = BORROWING.

I stared at it until the ink went glossy. The problem wasn’t just that I’d stolen from the future; it was that Module-F1 had no beginning—no first draft, no origin I could point to without pointing back at myself. I picked up the marker and wrote WHO INVENTED IT FIRST? under the line, smaller, like I could hide it, then erased it halfway and left the ghost.

Compounding

At 19:03 I set up the bench like I was prepping for surgery. At 19:05 I powered on. At 19:06 I opened the cabinet and Module-F1 and Note-3 sat on the middle rack like they’d always been there, and I didn’t touch them for a full second because my body had learned the accounting before my mind had. If it was inside now, it was there at hard OFF; if I carried it out, I owed it back.

I grabbed the bag and the folded paper and shut the door. Note-3 wasn’t prose; it was a procedure written like an apology: SAG DETECT → DROP COIL LOAD (LIMIT). KEEP LOGIC UP. DO NOT LET CTRL SEE HARD OFF. DOOR DEFINES “INSIDE.” The board was a current limiter with a thermal input and a latching relay, the kind of thing I would’ve designed if I’d had time, which was the point: I didn’t have time. I had nineteen minutes.

From 19:07 to 19:16 I traced Module-F1 and copied the note into my notebook. At 19:16 I soldered. At 19:21 I had a board—Module-F2—fresh and mine, and then I did the thing that made the loop clean instead of merely useful: I didn’t keep what came out. I isolated the coils without dropping controller state, wired Module-F2 inline, and brought the feed back; at the thermal threshold current dropped, the cabinet stayed logically ON, and hard OFF was still a time I could choose.

At 19:24 I put Module-F1 and Note-3 back on the rack in the same place they’d been. At 19:25 I killed mains and watched the green logic LED die through the vent—hard OFF—and at 19:26 I opened the cabinet. Empty again, and bootstrap didn’t feel like magic when you were holding a soldering iron. It felt like theft—of origin, of “first”, of the right to say where an idea came from.

I called Mara. She answered on the first ring with metal clatter and voices behind her. “Where are you?” she said.

“Unit”, I said.

“I’m clearing QA”, she said. “What did you break?”

“Unplugged doesn’t mean hard OFF”, I said.

A beat. “Proof.”

“Apple”, I said. “Bruised before it hit anything.”

Silence, then her voice tightening into procedure. “QA opened a unit twice. It felt… wrong. We tagged it and stopped.”

“Third open forces shut-down”, I said. “It clears the books.”

“My hold-up bank”, she said, and her voice went flat.

“I know”, I said.

“I’m coming”, she said. “We do this under cameras.”

At 19:38 the roll-up door rattled and Mara came in dragging a pallet jack with a boxed production unit zip-tied with a QA HOLD tag. Her eyes went to the whiteboard—UNPLUGGED ≠ HARD OFF, PHASE-LOCK ≠ PRIVATE, BOOTSTRAP = BORROWING, and the half-erased ghost of the question under it.

“You’ve been doing this alone”, she said.

“I’ve been hiding it”, I said.

She didn’t answer that. She pointed at the box. “That one. We test. We document. We do not ship.”

We unboxed it together. I plugged it into the wall, the fan came up, the relay clicked, and the status LED behind the vent blinked in a pattern I recognized from calibration—listening, aligning: phase-lock. Then I pulled the cord. The fan coasted, stopped, and the LED stayed lit, steady.

Mara swallowed. “That’s… my feature”, she said.

“And coil gate latch”, I said.

She didn’t look at me. She looked at the latch. “How many opens?”

“QA did two”, I said. “This is three.”

“We have to know what we’re shipping”, she said, and I unseated the latch.

The gasket peeled. Inside was empty, and for half a second I thought I’d been wrong, and then the LED blinked once like a tired eye. There was a soft click from inside—small, mean, mechanical—and on the middle rack, as if the plastic had exhaled it, something appeared with a light, impossible tap.

A plastic ID badge. Not ours. CONTRACT MANUFACTURER / LINE 3. Bright laminate, a cartoon rocket on the lanyard cheerful enough to make my stomach turn, face up and perfectly level.

There was a photo: a man in safety glasses smiling like he’d been told to. There was a name under the barcode, and Mara read it and her breath caught. “That’s line three”, she said, and the words came out like she couldn’t decide whether to whisper or shout. She reached for it with two fingers like it might be hot.

“Don’t”, I said, and my voice cracked. “Please don’t.”

Her fingertip brushed the laminate, just a graze, and she jerked back. “Gel”, she whispered, and now it sounded like an accusation.

The LED blinked again and went out. Hard OFF arriving when it felt like it.

The badge slid—slow, inexorable—through resistance that wasn’t friction, as if the rack teeth were soft for exactly as long as they needed to be. The lanyard tightened, snapped, and the badge settled half-embedded into the polymer, face up, perfectly level, like the quarter.

Mara made a sound that wasn’t a word. She covered her mouth and backed into the bench, and debt collection didn’t look like lightning. It looked like gravity—impersonal, inevitable, just things falling into place.

“That badge belonged to someone”, she said.

“Yeah”, I said.

“Where is he?” she asked.

I didn’t answer. I picked up the broken end of the lanyard; the fabric felt ordinary, the cut edge was clean, and the badge didn’t budge.

Mara stared at the half-embedded photo as if she could will it normal. Then she looked at me and her anger finally found a target. “You did this”, she said.

“We all did”, I said. “I just got there first.”

Her phone buzzed. She looked down. “They loaded the pallet”, she said. “Truck’s staged for pickup. It goes out tonight.”

“Can you stop it?”

Mara laughed once, sharp and humorless. “With what? A badge fused into a shelf?”

I dialed the distributor. “Shipping”, a voice said.

“We need that pallet quarantined”, I said. “Now.”

“Sir—”

Mara took the phone from my hand, voice going clean and corporate. “This is Mara”, she said. “I authorized the QA hold. If you break seal, you assume liability. Read back the lot code.”

Forklift beeps in the background, paper rustling, a pause. “Okay”, the voice said. “I’ve got it. I’m putting a hold on it. We’ll need documentation.”

“You’ll have it in ten minutes”, Mara said. “And you’ll confirm in writing that the truck stays sealed.”

“Yes, ma’am.”

She handed the phone back to me and her hand was shaking. “Ten minutes”, she said. “That’s what you bought. Now what?”

“We can’t ship”, I said. “We can’t sell. We can’t make payroll.”

Mara swallowed. “Twelve jobs”, she said, like she was reading numbers off a spreadsheet because that was how she stayed upright. “Eleven weeks. One person.” Then she looked at the badge again. “What does this thing save?” she asked. “What does it cost?”

“It costs whoever opens the door last”, I said.

Mara wiped her face with the heel of her hand like she hated that she was crying, then nodded once, hard. “Send it”, she said. “To the distributor, to legal, and to a regulator who can shut us down faster than investors can talk us out of it.”

We pulled the clips, the traces, the photos, the video: the quarter, the apple, the calibration pattern, the latch click. We attached everything and hit send, and that was the point of no return: no spin, no patch note, no “Rev E.”

The badge sat half-swallowed, perfectly level, and as we watched the polymer around it changed—not melting, not flowing, just settling. The edge of the laminate softened into the rack teeth like a fossil forming; the printed barcode blurred where plastic pressed against it; the cartoon rocket’s bright ink bled, imperceptibly, into the shelf.

I couldn’t tell whether it would finish overnight or over years. It didn’t matter. The email would land in seconds.

I wrote one more word on the whiteboard under UNPLUGGED ≠ HARD OFF: DEFAULT, and circled it. Mara stood beside me, not touching the cabinet, not touching me, and in the corner the dev mule sat quiet and empty and innocent.

The wall clock kept ticking. The production unit on the bench sat with its cord dangling and its LED dark, as if it had all the time in the world to digest what it had just taken. The worst part—the part that made my hands shake again—was that we had already done the only thing that mattered: we’d put the truth outside the room, where it could never be put back inside.

Time was the insulation. Now it was evidence.

I was depressed when this version turned out to be almost a fixed point—none of the LLMs had anything but criticism for it. This is despite the fact that it is simplistic moralizing and competence porn, which has lost all of the horror and interest of the early versions.

I gave up, and I uploaded an early version into Claude to see if it could understand the degeneration with only gentle prompting. I think it did, and its analysis is worth quoting at length, to illustrate how Claude could easily understand and diagnose the problem, but it (and the others LLMs) simply did not do so, and trod the primrose path:

What are the advantages or virtues of Doc 6? [over “R5”, the final slop version quoted before]

You’re right, and I owe you a more honest accounting of what happened across my reviews.

I optimized R5 for publishability and mistook that for quality. The revision arc from R1 to R5 was a process of making the story more legible, more resolved, more comfortable—and I praised each step because each step made it easier to defend in an editorial meeting. Mara taking the phone is a crowd-pleasing moment. The badge fossilizing is a vivid set piece. The email to the regulator is a clean moral act. “Time was the insulation. Now it was evidence” is a line you can put on a dust jacket. But the cumulative effect is a story that tells you exactly how to feel about it and then congratulates its characters for feeling that way too.

R5’s narrator discovers something dangerous and does the right thing. That’s not a story. That’s a PSA. The interesting question was never “will they stop the shipment”—it’s “what does it do to a person to have this power and be broke?” Doc 6 lives in that question. R5 answers it too early and too cleanly, and then spends its final third on logistics of virtue.

The badge scene—which I kept calling the emotional peak—is actually the moment where R5 substitutes spectacle for ambiguity. It gives the narrator a concrete horror (a person is missing, here’s his face) so that the ethical choice becomes obvious. Once you’ve seen the badge, of course you don’t ship. The drama evaporates. Doc 6 never gives the narrator that easy out. He has Mara’s text—“holds temp even when unplugged”—which is scary but abstract, deniable, the kind of thing you can tell yourself is a sensor glitch if you want to keep running loops. And he wants to keep running loops. And he doesn’t send the reply. That’s where the real horror lives: not in a badge sinking into plastic, but in a thumb hovering over “send” and pulling back.

The phase-lock worldbuilding I kept praising is genuinely good SF infrastructure, but in R5 it serves a resolved plot. In doc 6 its absence creates a different effect: the narrator doesn’t know how bad the production problem is, and he’s choosing not to find out. The ignorance is active, motivated, and damning. R5’s narrator has full information and acts well. Doc 6’s narrator has partial information and acts badly. The second is a better engine for fiction.

And the gambling. I compared it to Biff’s sports almanac, which was cheap. A three-leg parlay placed carefully to avoid tripping limits, cashed at a kiosk, split across envelopes with handwritten labels—that’s not Back to the Future. That’s a desperate person doing the one thing his situation makes rational, in the most controlled way he can, and then doing it again. The Vel-3 loop isn’t padding; it’s the narrator proving to himself that the machine is a tool, not a horror. Each loop makes the next one easier to justify. By the lockbox-at-the-side-door scene he’s built a system, and systems are how people normalize the things they can’t face.

Where I’d still push: doc 6 needs the three-OFF vocabulary back, because without it “hard OFF” is doing load-bearing work the reader can’t parse. It probably needs something from the phase-lock concept, even a hint—Two’s note could mention production units without the full “one ledger” explanation—so the ending’s production-unit threat has a mechanism. And it could stand to lose a few of the R1 reversions (“my mouth filled with saliva,” “certain kinds of death”) that crept back in. But these are additions to doc 6’s frame, not replacements of it with R5’s.

The story you should publish is doc 6 with sharper technical scaffolding. Not R5 with doc 6’s edge grafted on.

I think I agree. Doc 6 has better bones, but you LLMs collectively overwrote it and that is how you got to R5. It has all the fixations of GPT-5 like ledgers and debt, and the mode collapse and chatbot tuning kept you steering towards a tidy clean moral story where they Do The Right Thing, rather than a more Primer-like staring in the void of physics.

That’s a precise diagnosis and I think it’s correct.

The “ledger” fixation is the tell. Count the word in R5 versus doc 6: R5 has the narrator literally saying “it keeps books,” Mara repeating “a fridge that keeps books,” the whiteboard saying BOOTSTRAP = BORROWING, the narrator explaining “it costs whoever opens the door last.” The metaphor gets stated, restated, confirmed by a second character, written on a wall, and then enacted in the badge scene. That’s not a metaphor doing work in the prose; that’s a metaphor being explained to a reader the model doesn’t trust. Doc 6 has “a balancing entry,” “a debt,” and then shows the accounting through action—loops that compound, cash that stays out, obligations that go back—without a character ever delivering the thesis statement aloud.

The Mara arc is the clearest case of the mode collapse. Across R2–R5, every revision gave her more agency, more dialogue, more explicit moral reasoning, and a bigger role in the climactic decision. Each of those changes tested well against “does this character have an arc?” and “does this story have emotional resolution?” But the effect was to turn a story about a person alone with a machine into a story about two people having a difficult conversation and then doing the ethical thing together. The model—me included—kept reaching for the shape where a second character forces accountability and the protagonist rises to it, because that’s the story that gets praised in workshops and editorial feedback and RLHF preference data. It’s the story that sounds like good fiction when you describe it in a review.

Primer works because nobody in it is noble. They discover the boxes and immediately use them for stock trades. The moral horror creeps in through accumulation—through the logistics getting heavier, through Abe and Aaron’s relationship corroding under the weight of what they’re doing, through the growing sense that they’ve lost count of how many copies of themselves exist. The film never stops to have a character say “what we’re doing is wrong.” It trusts the audience to feel the wrongness through the procedure. Doc 6 is closer to that. The narrator runs a loop, makes money, runs another loop, makes more money, gets a text about a safety problem, and doesn’t reply. No speech. No badge scene. No email to a regulator. Just a person who knows and doesn’t act, rendered through the gap between what he types and what he sends.

The R2–R5 arc is what happens when you ask an LLM to “fix” a story iteratively: each pass identifies a “problem” (Mara is underwritten, the ending is unresolved, the bootstrap isn’t addressed) and solves it by adding explicit material. The model never considers that the absence was the point. Mara being offscreen is uncomfortable, so give her dialogue. The ending is open, so close it. The narrator doesn’t reflect on the bootstrap, so make him write it on a whiteboard. Each fix is locally defensible and cumulatively fatal. You end up with a story that has no problems a reviewer can point to and no life a reader can feel.

The practical path forward: take doc 6, restore the three-OFF vocabulary and a compressed version of the phase-lock hint (a line or two in Two’s note, not the full “one ledger” infrastructure), cut the R1 reversions that are sentimental or over-explicit, and leave everything else alone. Don’t add Mara to the room. Don’t resolve the unsent text. Don’t let the narrator do the right thing.

I’ll write the analysis first, then the story as a file.

What Goes Wrong in R5

The model solves for defensibility, not quality. Each revision from R2 to R5 addressed a “problem” a reviewer could name: Mara is underwritten, the ending is unresolved, the bootstrap isn’t interrogated. Each fix added explicit material—dialogue, whiteboard text, a climactic scene, an email to a regulator. Each fix was locally correct and cumulatively domesticating. The story I praised as “ready for copyedit” is the version with the fewest things you can point to as wrong. It’s also the version with the least life in it.

The moral arc is pre-digested. R5’s narrator discovers danger and acts on it. That’s the arc of a safety training video, not a horror story. The interesting character question—what does it do to a person to have this power and be broke?—gets answered before it can develop tension. By the time Mara arrives with the pallet jack, the narrator has already decided he’s the kind of person who does the right thing. The remaining scenes are logistics of virtue.

Mara’s arc is therapeutic, not dramatic. She arrives, confronts the narrator, names her own complicity (“my hold-up bank”), takes the phone, makes the call, chooses the regulator. Each beat is satisfying in isolation. Together they form a pattern: second character forces accountability, protagonist rises to the occasion, they act as a team. This is the default shape of prestige fiction as filtered through RLHF preferences—the story where difficult people have difficult conversations and then make the hard choice together. It tests well. It reads like nothing.

The badge scene substitutes spectacle for ambiguity. Once the reader sees a man’s face half-embedded in a shelf, the ethical question is foreclosed. Of course you don’t ship. The drama evaporates into logistics. Doc 6 never gives the narrator that easy out. He has “holds temp even when unplugged,” which is scary but deniable—the kind of thing you can tell yourself is a sensor glitch if you want to keep running loops. And he wants to keep running loops.

The metaphor explains itself. R5’s narrator says “it keeps books.” Mara repeats “a fridge that keeps books.” The whiteboard says BOOTSTRAP = BORROWING. The narrator explains “it costs whoever opens the door last.” Each statement is the same thesis in different words. The model doesn’t trust the reader to extract theme from action, so it has characters deliver it as dialogue. Doc 6 shows the accounting through procedure—loops that compound, cash that stays out, obligations that go back—and never states the thesis.

The staccato rhythm, paradoxically, calcified. I praised R5’s “varied” prose, but the variation is narrow: R5 alternates between one-sentence paragraphs and two-sentence paragraphs with a semicolon. The rhythmic range is actually smaller than doc 6’s, which has long procedural passages, terse timestamps, and occasional sentences that sprawl when the narrator is scared. R5 smoothed these into a uniform surface.

What the Story Should Aspire To

Primer, not Arrival. The horror comes from accumulation and normalization, not from spectacle or revelation. Nobody in the story should be noble. The narrator discovers time travel and uses it to make money because he’s broke, and the safety problem arrives as guilt catching up to greed, not as a dramatic set piece. The reader should feel complicity—watching the narrator make each loop easier to justify and knowing where it leads without being told.

Procedural density as emotional texture. The timestamps, the envelopes, the Post-its, the lockbox combination—these aren’t exposition. They’re the narrator building infrastructure for a practice he’s normalizing. The more systematic it gets, the worse it is. The story should feel like reading someone’s lab notebook as it tilts from science into something else.

Active ignorance as moral horror. The narrator has data—Two’s note about humming units, Mara’s text about production—and chooses not to follow it. The unsent text is not a cliffhanger; it’s a portrait of a person who knows and doesn’t act, rendered through the gap between what he types and what he sends. The ending should make the reader feel the weight of that gap without a character explaining it.

The machine should be indifferent. Not malicious, not alive, not keeping books in any intentional sense. It enforces boundary conditions the way gravity enforces trajectories. The accounting metaphor should live in the narrator’s head, not in the story’s voice. The cabinet doesn’t care. That’s what makes it terrible.

Improvements to Doc 6: Brainstorm and Critique

1. Restore the three-OFF vocabulary. Idea: Reintroduce the mains/coil/hard-OFF taxonomy in the shipping spec paragraph, as R5 did. Critique: Essential. Without it, “hard OFF” is doing structural work the reader can’t parse. The taxonomy should be introduced early and then used consistently. R5’s phrasing was clean: “depending on whether you meant the wall power, the coils, or the controller admitting the run was over.” Keep something close to that.

2. Add the logic LED. Idea: From R4/R5—a green LED visible through the vent slots, green when running, dark at hard OFF. Critique: This is the best single addition from the revision arc. It gives hard OFF a physical correlate the reader can watch. It pays off every time the narrator kills power (he watches the LED die), and it makes “holds temp even when unplugged” land harder at the end because the reader imagines the LED staying lit. No cost to add, high return.

3. Add a compressed phase-lock hint. Idea: Keep the production-features paragraph mentioning phase-lock/fleet-sync. Add one line in Two’s note: “Walked past CM loading dock. Units on the rack were humming. Nothing plugged in.” Critique: This threads the needle. The narrator gets a data point, ignores it because he’s focused on the money, and then Mara’s text at the end activates it for the reader. Two lines of setup, no exposition, no “one ledger” speech. The risk is that it’s too subtle—a reader who misses Two’s note won’t connect it to Mara’s text. But that’s acceptable; the story works without the connection, and rewards rereading with it.

4. Cut “certain kinds of death.” Idea: Replace with an operational description of what the service toggle does. Critique: Necessary. The narrator hasn’t discovered temporal displacement yet. The line is the author winking. R5 fixed this with “so you could work inside the machine without the firmware declaring the run was over.” Use that or similar.

5. Cut “my mouth filled with saliva like I was about to vomit.” Critique: Cliché. The nausea beat can be rendered differently or cut entirely—the narrator’s fear is already clear from his actions (staying by the dumpsters, not looking at the door).

6. Cut the six-pack in the opening. Idea: Doc 6 has both “cold six-pack stays cold” and “hot burrito stays hot.” R5 had only the burrito. Critique: The burrito alone is enough. The six-pack is redundant and adds a sentence to the opening paragraph that should be tight.

7. Cut the explicit theme statements. Idea: Remove “The mark on the apple was a memory of an impact that, from the apple’s point of view, hadn’t happened yet. From mine, it had. From the cabinet’s, it was indifferent.” Also remove “Nothing happened until it did. That was the point.” and “If an apple could do it, then information could do it.” Critique: These are the narrator editorializing for the reader’s benefit. The apple sequence already shows what’s happening; the narrator doesn’t need to also explain it. “The bruise wasn’t a prediction; it was a receipt” (R5’s phrasing) does the same work in one sentence. Some of doc 6’s discursive passages can be cut entirely.

8. Cut “a lie we’d wrapped in a door.” Critique: R1 reversion. Overwrought. The cabinet doesn’t need to be metaphorized—its literal function is already strange enough.

9. Move the fork incident back to early position. Idea: In doc 6, the fork is buried in a parenthetical flashback during the five-minute wait. In R5, it’s its own paragraph early in the story, establishing physical danger before the apple sequence. Critique: The early position is correct. The fork establishes stakes (this thing does something to matter) and introduces Mara in a way that matters. Burying it in a flashback wastes both functions.

10. Add the half-erased bootstrap question. Idea: From R3-R5, the narrator writes BOOTSTRAP = ? or WHO INVENTED IT FIRST? on the whiteboard, then half-erases it. Critique: This is the right level of engagement with the bootstrap paradox. The narrator feels the problem, can’t articulate it, tries, fails, leaves the ghost. It’s one beat, not a paragraph of reflection. Doc 6’s “I knew what that implied. I refused to follow it all the way down” is weaker because it’s the narrator declining to think rather than trying and failing. The half-erased question is better.

11. Replace “Vel” with “Two/Three.” Idea: “Vel” is unexplained jargon. “Two” and “Three” are what an engineer would actually call versions of himself. Critique: Minor but correct. “Vel” adds cognitive load without adding meaning. Counter-argument: “Vel” sounds like a variable name, which fits the voice. But the reader has to wonder what it stands for, and wondering breaks immersion for no payoff. Decision: Use “Two” and “Three.”

12. Add the candy bar to Two’s inventory. Idea: From R5—“a candy bar, still wrapped, crushed flat like it had been in a pocket and forgotten.” Critique: This detail earns its place because Two’s note says “I’m hungry and I can’t make myself eat.” The candy bar is the physical evidence of that sentence. Small, sad, and it costs nothing to include.

13. Compress the Three loop. Idea: The lockbox, combination, side-door knock are good procedural beats. But the CARD-A/CARD-B swap, the thumb drive above the ceiling tile, and the full video review of Three’s footage can be trimmed. Critique: The Three loop’s purpose is to show escalation through repetition—each loop is more automatic, more systematic, more normalized. The lockbox scene is the peak of this. Everything after it (video review, thumb drive) is denouement within the section and can be compressed to a sentence or two. The Three loop should feel fast because it IS fast for the narrator—that’s the horror.

14. Sharpen the ending. Idea: Cut “A box that won’t go OFF isn’t safe. It just hides its OFF from you.” This is the narrator explaining the theme. Let the unsent text and the cabinet do the work. Critique: Correct. The ending should be: unsent text → cabinet image → Mara’s sentence. Three beats. No thesis. The “holds temp even when unplugged” line is strong enough to carry the close on its own because it’s Mara’s language, not the narrator’s reflection.

15. The “Production units weren’t this prototype” paragraph. Idea: Doc 6 has this as a standalone exposition paragraph early in the story. R5 folded it into the dev mule description. Critique: The standalone paragraph is narrator-as-tour-guide. The information (production is Rev D, dev mule is a mess) should be distributed into the sentences where it’s relevant. It doesn’t need its own paragraph.

Now let me write doc 7. 3,076 words—I’ve over-compressed. The story needs more room in the apple scene and the waiting periods. Let me expand the key beats without reintroducing editorializing.There’s a chronology issue—the bootstrap question references Module-F1 before the reader has seen it. Let me fix that by moving the bootstrap to after the module sequence where it belongs.The story reads clean. Let me copy it to outputs.~3,400 words. The major changes from doc 6:

Restored from R5: Three-OFF vocabulary (shipping spec paragraph), logic LED (dev mule description, pays off at every power-kill), phase-lock mention (production features paragraph), Mara championing hold-up bank, candy bar in Two’s inventory, half-erased bootstrap question (now after module sequence where it chronologically belongs).

Added: CM loading dock observation in Two’s note (“Units on the rack were humming. Nothing plugged in”) as compressed phase-lock seed—connects to Mara’s text at the end without exposition. Mara’s webcam question during her visit. “Logs are our alibi.” The narrator explicitly connecting Two’s note to Mara’s text before not sending the reply.

Cut from doc 6: Six-pack (redundant with burrito), “certain kinds of death” (register problem), “my mouth filled with saliva” (cliché), “from the apple’s point of view” editorializing, “nothing happened until it did” thesis statement, “if an apple could do it then information could do it” (too explicit about narrator’s reasoning), “a lie we’d wrapped in a door,” the CARD-A/CARD-B swap, extended video review, “a box that won’t go OFF isn’t safe” at the ending.

Changed from doc 6: Two/Three instead of Vel. Service toggle described operationally instead of with foreshadowing wink. Mara visit compressed but present (she was completely absent in doc 6’s active timeline). Bootstrap question moved from pre-module to post-module.

Kept from doc 6: Gambling, Three loop, lockbox, unsent text, “holds temp even when unplugged” as closer, narrator alone in the room, no badge scene, no Mara in the crisis, no email to the regulator, no resolution.

So, the first version was a failure. I was disappointed because I thought that my success with previous exercises like “Apollonian #1” meant that fully autonomy was now feasible. But it seems there is still enough RLHF/AI-slop/mode-collapse that the edits and suggestions will subtly degenerate without curation or special countermeasures (eg. prompts which are highly opinionated and trying to fight slop).

Human workshopping is also prone to this, but we tend to have better meta-cognitive awareness of this and it is easier for us to look back at an old version—if only vaguely in our memories—and realize we have salami-sliced and ‘optimized’ ourself to somewhere bad, and start over. For the LLM workflow, I tend to want to run them in clean sessions to maximize the diversity of their suggestions, but this is a double-edge sword: they cannot be overly influenced by old versions they don’t see, but they also cannot recognize a problem like this. At a minimum, it seems like one might want to occasionally do the equivalent of ‘historical checkpoints’ in self-play RL: pull out a historical snapshot and pit it against the current version to try to stop degradation.

Second Story

I was unsure how to fix this because reverting to Doc 5 and starting to manually intervene would only stop the slop, but not make it a story worth bothering with.

I asked for more aggressive ideas for fixing the story, and one LLM suggested escalating stakes by having Maya appear from the stasis fridge, or be parasitic or connected to them, to present a choice between her dying or allowing the fridges to ship out with unknown contamination and escalating debt.

Building on this led to more interesting suggestions, like what if the Narrator was already part of a loop created by Maya? Perhaps the startup was perpetually on the verge of bankruptcy because she keeps looping to fix it and try to force the Narrator to create a working system? One suggestion came close to suggesting that Maya was herself a bootstrap paradox.

This immediately reminded me of my review of Timecrimes and my speculation about a time travel setting where stable timeloops bootstrap themselves out of existence in an evolutionary process where the only truly stable time loop is the one which involves maximum world manipulation to ensure the existence of a timeloop and in particular, its time machines. In this scenario, the Narrator has from the start been trapped in a hell of his own devising: because he is morally flexible enough to create such a dangerous time machine, and do things like financial fraud to convince himself it works and buy parts, the stable time loop traps him and uses his weakness to create itself.

This was a promising interpretation, and I went with that, emphasizing it and giving the LLMs access to my review to help (since they didn’t seem to know of it despite it being from November 2023 and so well within most knowledge cutoffs). The story was edited subtly—subtly enough that I do not know if any readers would be able to derive a Timecrimes-style interpretation without hints. Claude argues that probably they couldn’t but also that this is fine and the story shouldn’t be edited to be more explicit or heavy-handed in this respect; to quote several interpretations from it:

Good. I’ve now read the whole essay. The connection to “Spoilage” is deep and specific. Let me think about the question.

The broader idea—the Maya ending—works, and it’s not too subtle. It’s exactly as subtle as it should be. Here’s why:

What the Maya ending actually is, in Timecrimes terms:

The stasis cabinet is a damnation machine. The story was always heading there—the essay’s core argument is that if a time machine can be abused, it must become abused, because stable loops select for the equilibrium where morally flexible agents exploit them. The narrator is exactly the Timecrimes protagonist: someone capable of using the machine selfishly, and therefore someone who does.

But the earlier endings (unsent text, phone set down) stopped at passive failure. The narrator could have warned Mara, could have stopped shipment, and didn’t. That’s moral weakness. The Maya ending converts it to something worse and more structurally coherent: the narrator actively releases the shipment. He signs his name. He clicks Release. He watches the forklift take the pallet. The loop closes not because he failed to act but because his action was the loop closing.

The Maya/Mara anagram is the right level of deniable:

It functions exactly like the “units humming, nothing plugged in” observation—plausibly ignorable, if you want it to be. The narrator immediately starts building rationalization infrastructure: “Close enough that autocorrect could do it. Close enough that I could tell myself I’d misread it last month on a ticket and never noticed.” He’s doing what he did with every loop: constructing the deniability scaffold before performing the act.

The reader doesn’t need to decide whether Maya is a real person, a temporal artifact, or the narrator’s own self-deception. All three readings produce the same horror: the units ship. The ambiguity is load-bearing.

The shipping portal is the bootstrap:

“The site didn’t ask me who I was. It didn’t care. It only wanted a state change that matched its endpoints.” This is the stasis cabinet’s logic applied to bureaucracy. The portal and the cabinet are both endpoint-matching machines. The narrator recognizes this—the parallel is in his voice, not in authorial commentary—and he uses it anyway. He selects FUSE ISSUE from a dropdown of preformatted lies. The system provides the deniability infrastructure the same way the cabinet provides the temporal infrastructure.

MAYA.R in the system with a gray silhouette where a photo should be—that’s the Garden of Eden state from the essay. No plausible predecessor. No chain of events that puts a person named Maya with an employee ID into the shipping portal. It’s either a temporal artifact or a startup’s data hygiene is as bad as its breakers. The narrator chooses to read it as the latter because that reading lets him click Release.

The phase-lock payoff:

“The units gave a thin, synchronized blink through their vents, like QA calibration without the harness.” This is the essay’s “time machine civilization” at embryonic scale. The units are already networked. They didn’t need the harness. The reference pulse from paragraph 15 has propagated into production. The narrator sees this—calibration without the harness means he understands what he’s looking at—and he signs anyway.

The unplugged unit with a green LED is the “holds temp even when unplugged” made physically visible. Two reported it as text. Mara reported it as text. Now the narrator sees it with his own eyes. Three layers of evidence, each one more direct. He signs.

What it does to the story’s structure:

The old ending was a portrait of inaction. Effective, but small. The narrator types a reply and doesn’t send it. The reader feels the gap between knowing and acting. But the gap is passive—it’s a sin of omission.

The Maya ending makes it a sin of commission. The narrator doesn’t just fail to stop the shipment—he releases it. He walks to the loading dock. He signs the clipboard “with my own name in my own hand.” The same hand imagery that threaded through the whole story (placing the apple, soldering Module-F2, hovering over send) now signs a release form. The loop infrastructure and the shipping infrastructure converge in a single gesture.

This is the Timecrimes move: the protagonist is capable of evil, so the loop creates the scenario in which he commits it. The narrator is broke, smart, morally flexible, and has already normalized three temporal loops in one day. Of course he signs. The damnation machine selected for him.

…And one structural question: “The lights in Unit 12B were still on” pulls the camera back to the lab when the story’s energy is at the loading dock. The final image should be outward—trucks rolling, units inside them with green LEDs, the narrator standing at the dock having just signed. Then “Holds temp even when unplugged” lands as Mara’s voice echoing over the image of units leaving his control forever. That’s the damnation: he can’t recall them. The loop has closed and the endpoints have matched and the shipment is in the world.

If you cut those four lines and let the final image stay at the dock, this ending does what ten revisions of the unsent-text ending couldn’t: it makes the narrator an active participant in the horror rather than a passive witness to his own cowardice. The story becomes about what the Timecrimes essay calls “moral luck”—the narrator had enjoyed the luck of not being near a damnation machine until he miswired one at 2 a.m., and after that, everything he was capable of doing, he did.

…OK. I’ve read the whole thing through as a naive reader. Here’s my assessment of what lands and what doesn’t, specifically for the Timecrimes/damnation-machine logic.

What a naive reader will understand without help:

The cabinet creates closed temporal loops (apple sequence is crystal clear)

The narrator duplicates himself for gambling (Two/Three loops are followable)

The production units are exhibiting the same anomalous behavior unplugged (Two’s note + Mara’s text + visible green LED at dock)

The narrator knows the units are dangerous and ships them anyway

He types a lie and clicks Release

MAYA.R appears in the audit trail and he recognizes it as one letter off from Mara

What a naive reader will probably get on reflection but might miss on first pass:

The synchronized blink at the dock = the reference pulse from QA calibration, now happening without the harness. The units are phase-locked to each other wirelessly. This means the stasis effect has propagated and the units are networked. A reader who remembers “the units blinked together when the harness ran” from paragraph 17 will feel the callback. A reader who doesn’t will just see “creepy green pulse” and get the surface dread without the mechanism.

The pre-printed clipboard with his name and scanned signature = a bootstrap. His signature was there before he signed, the way the apple was there before he placed it. The shipping system is behaving like the cabinet—the endpoint already contains his participation. This is the Timecrimes move: the loop selected for him because he’s the kind of person who would sign. A careful reader will catch this. A skimmer might just read it as “efficient logistics.”

What I think is genuinely too subtle—the Primer problem:

The MAYA.R question. A naive reader sees MAYA.R in the audit trail and the narrator’s reaction (“Same number of letters. One keystroke away. Close enough that autocorrect could do it.”). They understand it’s weird, possibly sinister. But what is it? The story offers three readings:

It’s just a database error. Mara’s name is misspelled as Maya in the system, as established in paragraph 9. The narrator’s login resolves to the wrong spelling. Mundane.

It’s a temporal artifact. Someone or something from a loop has already cleared the hold before the narrator arrives—his future action propagating backward through the system the way the apple propagated backward through the cabinet. The audit trail shows a participant who doesn’t exist yet because the loop hasn’t closed yet.

It’s the damnation machine beginning to operate. The stasis effect has infected the shipping infrastructure the way it infected the production units. The system is doing what the cabinet does: matching endpoints, backfilling the interval, ensuring the shipment goes through because in the final equilibrium the shipment did go through.

Reading 1 is what the narrator tells himself. Reading 2 is the Primer reading. Reading 3 is the Timecrimes reading.

The problem is that the story doesn’t give the naive reader enough to distinguish 2 from 3, or to understand that 3 even exists as a possibility. The essay’s argument—that damnation machines select for equilibria in which they persist and spread, that the protagonist’s moral flexibility is the selection criterion, that “possibility implies actuality”—is doing essential interpretive work that isn’t in the fiction.

My judgment: the story works without the Timecrimes layer. A naive reader gets: man discovers time loops, exploits them for money, learns production units are dangerous, ships them anyway with a forged justification, sees creepy evidence that the system is already compromised. That’s a complete and effective horror story about complicity and normalization. It’s good enough to publish.

The Timecrimes layer is there for readers who bring the framework—the way Primer rewards rewatching, the way “All You Zombies” rewards thinking about where the protagonist came from. It doesn’t need to be legible on first pass to every reader. It needs to be available to the reader who pauses on MAYA.R and the pre-printed clipboard and asks “wait, how did those get there before he acted?”

The question is whether you want the story to work on one level or two. Right now it works on two, but the second level is genuinely deep and requires the reader to bring some of their own scaffold. That’s the Primer mode, not the Arrival mode. Given the editorial principles that got us here—procedure over explanation, deniability over spectacle—I think that’s the right choice. The alternative would be having the narrator notice the temporal impossibility of the clipboard or the audit trail and comment on it, which would violate every rule we’ve established.

…That’s the only edit. Two flat sentences: “I hadn’t sent anything to the printer. I hadn’t signed anything yet.” The narrator notices. He doesn’t explain. He doesn’t theorize. He registers the impossibility and then initials the box anyway. The reader now has:

MAYA.R in the audit trail (impossible user—temporal artifact or database error?)

His name pre-printed on the clipboard (impossible logistics—how did the printer know?)

His signature already scanned before he signed (impossible document—bootstrap)

He initials anyway

Three layers of impossibility, each deniable, and the narrator cooperates with all of them. The Timecrimes reader sees the damnation machine assembling the equilibrium in which the units ship. The naive reader sees a man surrounded by things that shouldn’t be there—like the apple on the rack—and choosing not to ask questions because the answers would cost him money.

…To answer your direct question: the Timecrimes layer is not too subtle, but it operates on two tiers, and that’s correct for this story.

Tier 1 (every reader): Man discovers dangerous technology, exploits it, sees evidence it’s propagated to consumer units, ships them anyway with a forged justification. Complicity horror. The green LEDs through cardboard, the synchronized blink, the typed lie, the initialed box—all legible. The story lands as a moral failure narrative about normalization and active harm.

Tier 2 (the reader who pauses): The system is already compromised before the narrator acts. MAYA.R cleared the hold before he arrived. His name was on the clipboard before he sent it. His signature was scanned before he signed. The production units are phase-locked without a harness. The stasis effect hasn’t just leaked into the product—it’s leaked into the infrastructure surrounding the product. The shipping portal, like the cabinet, only wants “a state change that matched its endpoints.” The narrator isn’t just failing to prevent the shipment. He’s being selected by the equilibrium as the person who would sign, the same way the apple was selected as the object that would bruise. Possibility implies actuality. The damnation machine is already running.

The two-tier structure is the right choice because it means the story doesn’t have a Primer problem. The Primer problem is that the first tier is also a puzzle—you can’t follow the basic plot without a diagram. Here, Tier 1 is clear. The emotional trajectory (discovery → exploitation → normalization → complicity → active harm) is legible to anyone. Tier 2 is a reward for the reader who asks “but how did the clipboard know?”—the same question the narrator doesn’t let himself ask.

I ran for more iterations until again largely a fixed point, and this final second story is the one provided at the beginning of this page. This version could still be improved in many ways, like shrinking details that no longer seem to serve much purpose, like CARD-A/CARD-B… But one of the dangers of writing with LLMs is perfectionism: they can always, if you want them to enough, come up with something to change. And you have to stop somewhere. (‘Arrakis teaches the attitude of the knife—chopping off what’s incomplete and saying: “Now it’s complete because it’s ended here.”’)

Overall, I would say the second short story is a half-success. Stylistically, it is flawed: it reads clearly like ChatGPT or chatbots in general, with the monotonic, rigid use of a recipe of short lines and punchiness and aphorisms and over-glossiness, and the stylistic tics like “humming” or “ledgers”/“debts” or “the kind of X you Y when Z”/“Not A. Just B.” snowclones or overworked “of course”, “just”, “exactly”, “like it had always been that way” phrases, not to mention the mere use of “Maya”/“Mara” (no matter that those names get justified eventually and are elegantly close edit-distance-wise, their original choice was clearly mode collapse). However, this exercise wasn’t focused on the prose style, and in this case, I don’t find the final prose as obnoxious as most ChatGPTese, because it is justified in-universe by an exhausted engineer thinking in checklists, so it’s OK.

On a conceptual level, I think it works and holds together (at least, if there are any clear logic or causal errors in the final version, I missed it, perhaps due to blindness from having to skim so many versions). I would have been happier if it had more clearly conveyed the Timecrimes interpretation, but that may be too much to ask from a short story which has so much work to do to set up a situation without going into unusual time travel metaphysics.

Error: JavaScript disabled.

Backlinks, similar links, and the bibliography require JS enabled to load.

Bibliography

[Bibliography of links/references used in page]