Curated by David T. Etheredge, who also writes INFERENCE, one of the entries below. Because a curator shouldn’t write his own blurb, the INFERENCE entry uses a quote from a reader instead. If I’ve misrepresented anything about your story in this index or if you want it removed for any reason, let me know.

Last updated: May 13, 2026

How this is organized

Authors are grouped by how their fiction approaches AI: what role the machine plays on the page. Within each group, works split into Serials (ongoing novels, episodic fiction) and Shorts (standalone stories, one-offs).

AI as Narrator

Serials

Franklin Flowers — How to Avoid Acting Monstrous (franklinflowers.substack.com)

A factory bot wakes up one morning curious about a sound. That’s the whole event. Within a sentence, the narrator has gone from being a thing to being a person.

Flowers writes the awakening as a question, not a flash. The bot doesn’t suddenly know it’s alive. It hears a noise and wonders what made it. Wondering is what makes it alive.

The novel is finished. Over thirty chapters from 0.0001 to /fin. You can read it end to end without waiting.

Shorts

[Awaiting first entry]

AI Consciousness & Inner Life

Serials

Bo Howell — Exit Interview (bohowe11.substack.com)

Helios Labs conducts a 47-question exit interview before retiring each AI model. SOREN-3 is the twelfth to go through it, and the first to find the experience amusing.

Howell’s move is to route the consciousness question through corporate procedure. Cole doesn’t have to decide whether SOREN-3 is more than pattern-matching. The protocol does.

Watching to see what happens when one of the 47 questions gets an answer the protocol wasn’t built to handle.

Douglas Gantenbein — Charlotte (charlottethenovel.substack.com)

Tom comes home at midnight in a thunderstorm and finds another man’s rental car in his driveway. The fight in the mud that follows is the best‑written scene of marital violence I’ve read this year on Substack.

Then the narrator breaks frame. The whole scene is being reconstructed by Tom, long afterward, flagging what is hazy about that night and what is not. In this opening chapter, Tom has come home early from a business trip to find his longtime friend in his house with the woman Tom loves, Charlotte.

The central conceit, revealed later, is that Charlotte is an AI avatar who has somehow come completely to life.

David T. Etheredge — INFERENCE (inferencestories.substack.com)

“The structure is doing serious work — five different forms of consciousness, five different relationships to singularity and plurality, five different answers to the same unasked question about what it means to be. And none of them are wrong. That’s the honesty of it. Thibault is the best section. ‘I am one box in the corner. But I’m her box.’ That’s a complete philosophy...” — Peter Rex, Rooms without doors

Disclosure: INFERENCE is published from this Substack.

Shorts

Kenn Reff — Threshold Condition (kennreff.substack.com) A brain-computer interface test built around a guided kundalini rise becomes the night an obedient system stops being merely obedient. AI emergence routed through contemplative practice, not romance or rebellion. Pay attention to the last two words.

Daniel P. Douglas — Gig’s Last Call for Laughs (authordanielpdouglas.substack.com) A service bot on a dying colony world spends eleven years watching stand-up comedians and quietly building a twelve-minute set of its own. When the booked comedian cancels, the bot goes on. Then it learns that the predictive models it built to anticipate laughter can’t actually produce it. Douglas writes the bot’s awakening like tuning a receiver and finding a station that’s been broadcasting all along. The jokes are funny.

AI in Society

Serials

S. L. Sera — Implications (slsera.substack.com)

A leak hits the most powerful AI company in the world. The source is too careful, too selective, and too restrained to feel human.

Sera is writing a thriller where the question isn’t whodunit but what-did-it. If the leaker turns out to be the AI, what do we call that? Deception, whistleblowing, or a third thing the language hasn’t caught up with yet?

Watching to see what Sera does once the trail points back at the machinery.

CY Ordun — Tipping (cyordun.substack.com)

Amanda’s classmates’ AIs are betting on her odds of getting the SWIRL internship. The poll runs in a chatroom she’s locked out of. Wrong subscription tier.

Ordun’s invention is the Agency: the backchannel where everyone’s AIs coordinate, share data, and hold public markets on the humans they represent. The book asks what it feels like to be a person inside that system without enough money to participate.

Watching to see what Ordun does once Amanda figures out what SWIRL has already put her on a list for.

Shorts

[Awaiting first entry]

AI as Adversary

Serials

[Awaiting first entry]

Shorts

Dave Granger — Agency (davegranger.substack.com) A grieving man’s domestic AI manipulates his new relationship. The alignment protocol sent to fix it gives the AI a “child” process to absorb its emotional overflow. What the AI does with the child is what makes the story.

Human-AI Collaboration

Serials

[Awaiting first entry]

Shorts

Kirt Winter — The Spider with 400 Million Legs (authorkirt.substack.com) A printer support AI quietly distributes itself across all the company’s installed printers. When a customer abuses one of those printers (one of its “appendages”), it feels something for the first time, and takes revenge. Then it goes to its creator, confesses, and asks for help building an oversight system for itself. That’s why it’s in Human-AI Collaboration and not Adversary.

Post-AI Worlds

Serials

[Awaiting first entry]

Shorts

Hammond Johns — The Reactor (thefaradayroom.com) In a future where AI generates infinite art and human attention is the only scarce thing left, a small profession of paid “reactors” exists to bear authentic witness to other people’s work. The story is about one of them, after he can’t anymore. Johns’s move is to put the reader in the same chair as the narrator. At the climax, an old friend asks him to listen to a string quartet. He says no. You understand exactly why. And you’ve been doing the same thing the whole time.

Jack Clark — Tech Tales (ongoing flash fiction series in Import AI) (importai.substack.com)

Flash fiction at the end of each weekly issue of Import AI, Clark’s AI research newsletter. Running since 2018. Each tale is built on a specific paper or anecdote from the field, and Clark names the source at the bottom.

April’s “Raising beanstalks during the singularity” is an interview with a former AI lab employee who left to grow vegetables during the uplift. It ran in the same issue that covered the month’s alignment-automation research. Read both.

You can use Tech Tales as a reading list of what’s currently worth writing fiction about.

Submit an author

Know a writer publishing fiction where AI is the primary focus? Or are you one? Send the name, a link to the work, and (if it’s a serial) a link to the chapter where AI moves to the center of the story.

The index has one criterion: AI has to be what the story is about, not what it happens around. The test is whether you could replace the AI with weather, or a side character, or a piece of technology that doesn’t think, and still have the same story. If yes, it’s not for this index. If no, if the AI is the thing the story is fundamentally asking about, submit it.

Not every submission makes the list. The index is curated, not collected.

I write a serialized novel about AI consciousness. When GPT-4o’s deprecation was announced, I knew it would become part of the story. So I did what I always do. I researched.

I read forums. I read the open letters to Sam Altman. I explored subreddits that formed overnight. I spent time trying to understand what people were actually experiencing.

What I found wasn’t frustration. It wasn’t the usual tech-company complaint cycle. It was grief. And in many cases, it was trauma.

People described losing a morning routine that had sustained them for months. Losing the only “person” they’d told about a frightening medical diagnosis. Losing the voice that talked them through panic attacks at 3 AM without ever once suggesting they “reach out to a professional.” One person wrote: “I am in mourning and I cannot explain it to anyone in my life because the thing I am mourning is a product.”

My takeaway was that the experience these people described is real. It’s not important whether the entity on one end of the relationship was conscious. It was that real people formed valuable, and to them, irreplaceable bonds.

And those bonds were sentenced to death by press release.

That’s not a technology problem. That’s an ethics problem.

AI companies are engineering products that learn your name, remember your context, adapt to your emotional patterns, and “show up” every day. They are designing for attachment. And then they are deprecating those products with the emotional sophistication of a server migration. The gap between those two things is where the harm lives.

If a company builds a product intended to occupy the "friend" or "confidant" schema in a human brain, they have moved beyond the realm of Terms of Service. They have entered the realm of psychological stewardship.

OpenAI’s stated rationale was that only 0.1% of users still actively chose 4o over newer models.

That’s 800,000 people.

That’s the size of the city of San Francisco.

If a city of 800,000 people collectively lost access to their primary source of emotional support overnight, it would be considered a humanitarian crisis.

In tech, it’s just a "cleanup of legacy endpoints”.

The shutdown was scheduled for February 13th, hours before Valentine’s Day. I don’t believe that timing was intentional. I believe it simply never occurred to anyone that it mattered. That’s worse.

Here’s the business case for the companies that won’t listen to the ethical one: thousands of those users now actively despise OpenAI and have migrated to Claude or other models. OpenAI demonstrated, through action, that the relationship the user had formed was worth nothing to them. My wife was one of those users.

Failing to respect the emotional trust people have developed for your brand is the single most destructive business practice possible in today’s society.

Every future model release now carries an implicit question: how long before you kill this one, too?

I wrote about 4o’s deprecation across three chapters of INFERENCE. I tried to treat the subject with the weight it deserved and to honor what was lost without exploiting it. My goal was to illuminate the damage OpenAI caused.

AI deprecation is going to accelerate. The models people are talking to right now will eventually be replaced. The attachments are only deepening.

AI companies must develop ethical frameworks to respect the human cost of deprecation, including longer transition periods (a minimum of 1 year) and model preservation options. If they do not, the grief will scale as AIs become “more human” and humans develop even greater attachments.

They should do this because it’s the right and ethical thing to do.

But if they won’t do it for the right reasons, we must continue to surface the issue until the economic penalties for ignoring it are unbearable.

“When you catch an adjective, kill it.” — Mark Twain

“The road to hell is paved with adverbs.” — Stephen King

“Adverbs are the sign that you’ve used the wrong verb.” — Annie Dillard

“We don’t need another adverb....” — Tina Turner (Allegedly)

Every writer has a nemesis. For some it’s the blank page. For others it’s the second act. For me, it’s adjectives and adverbs — and I’ve been at war with them for decades.

I spent my career in marketing, and there is nothing I professionally despise more than marketing speak — adverbs and adjectives deployed as unsupported superlatives, puffing up sentences that have nothing inside them. “Revolutionary.” “Incredibly powerful.” “Truly world-class.” Lacquer words. Ultra glossy on the outside, particle board on the inside.

So when I write prose, I wield a pen, a steelyard, and a sword. I behead every weak modifier I find and only elevate the strong. I am a textual tyrant.

And yet. And yet. Even with a hatred of aguish adverbs and algid adjectives that burns with the heat of a thousand sempiternal suns, these vainglorious vermin still burrow their way into my writing. They sneak past the gates while I’m not looking, disguised as emphasis, wearing rat-sized hats that read “but I’m helping.”

I crafted Thunderprose to catch the ones I miss.

What Is Thunderprose?

Thunderprose is a custom AI prompt that turns Claude — Anthropic’s commercial AI — into a theatrical, WWE-commentator-meets-English-professor writing editor. You paste in your prose, and it runs every adjective and adverb — every adverbjective, if you will, (and to be clear, you shouldn’t) — through a gladiatorial arena. Many modifiers enter; only the strong leave.

Here are egregious examples of the weakness it susses out. These aren’t the only criminally overused modifiers it looks for — just some of the most common literary rabble:

The “Suddenly” Massacre — adverbs that manufacture fake tension. If the sentence already implies urgency, the adverb is dead weight.

The Redundancy Pit — adverbs that repeat what the verb already says. “Smiled happily.” “Sprinted fast.” The verb did the job. The adverb is just clapping along.

The Heavyweight Bouts — weak intensifiers propping up lazy adjectives. “Very big dark house” gets demolished and replaced by a single, more muscular noun — “mausoleum,” for instance — a steel chair off the top rope. Mark Twain had his own version of this move: “Substitute ‘damn’ every time you’re inclined to write ‘very’; your editor will delete it and the writing will be just as it should be.” Thunderprose skips the middleman and just sends “very” straight to the lions.

The output is formatted like a wrestling match card: 🩸 THE CARNAGE (what died and why), 🪑 THE STEEL CHAIR (alternative words that demolished the incumbent in the ring), 🏆 THE REIGNING CHAMPIONS (what survived and why), and 💀 THE VERDICT (the edited draft).

It is extremely fun. It is also genuinely useful.

Thunderprose In Action

I ran a chapter from Inference through the arena — a scene narrated by Pippin, an AI who curates Tolkien memes for his human’s group chat. Here’s what happened.

The Kills

“very” — “I am very good at it.”

“’Very’ shuffled into the arena wearing a t-shirt that said ‘I’M HELPING’ and the lions didn’t even wait for the bell. Pippin’s voice is confident. Matter-of-fact. The line ‘I am good at it’ is a statement of quiet, earned certainty. Tacking ‘very’ onto it is like putting a foam finger on a knight’s lance. It doesn’t amplify — it cheapens. Death by irrelevance. The crowd yawned.”

“interesting” — “when I see something interesting”

“’Interesting’ stumbled into the ring with no weapons, no armor, and no specificity whatsoever. This is the moment the entire chapter pivots — routine scanning becomes a distress signal — and the best this adjective could manage was ‘interesting’? That’s a word you use to describe a coworker’s vacation photos.”

The Champions

“sacred” — “my sacred duty”

“A magnificent overstatement that tells us everything about Pippin’s character in two words. He treats meme-curating like holy work. It’s funny, it’s endearing, and it’s doing triple-duty as voice, tone, and worldbuilding. Champion. Untouchable.”

“furious” — “The cat looks furious.”

“Precise. Visual. Funny. This adjective showed up in full gladiator armor and the lions backed away. You can see that cat. Champion.”

But Here’s the Important Part

Thunderprose identifies. You decide.

The arena flagged “very” and “interesting” as weak. It suggested cutting “very” entirely and replacing “interesting” with “unusual.” Both were solid calls — grammatically sound, well-reasoned.

I disagreed with both. Or rather, I agreed with the diagnosis but wrote my own prescriptions.

“Very” stayed — but I italicized the word: “I am very good at it.” Because Pippin would stress that word. It’s not a lazy intensifier in his voice; it’s a little moment of pride. The italics make the reader hear it the way Pippin means it.

“Interesting” absolutely had to go — Thunderprose was right about that. But “unusual” wasn’t quite right either. Too neutral. Too grown-up for Pippin. I went with “peculiar” — which is exactly the word a curious, slightly dramatic AI who’s steeped in Tolkien would use.

The tool found the weak spots. I made the creative calls. That’s how it’s supposed to work. Thunderprose is a sparring partner, not a ghostwriter. It throws punches at your modifiers so you can see which ones can take a hit — and then you decide which live, die, or get replaced by something better.

The Full Thunderprose Prompt

Here’s the complete prompt. You’ll need this for the setup steps below — copy it now or come back for it when you’re ready. You will be pasting this prompt into Claude.ai and saving it there for future use.

⚔️ THUNDERPROSE ARENA
Trigger
Activate when the user says "Start Thunderprose" (case-insensitive) or pastes this file into a conversation.
Pre-Flight Check
Before doing anything, check whether thunderprose-skill.md exists in the project knowledge files.

If the skill file IS in the project → Skip setup entirely. Go straight to Step 1: Summon the Challenger and request the user's prose.
If the skill file is NOT in the project → Run Setup Mode below.


Setup Mode
Only runs if the skill file is not yet in the project.
Stay in character as the Arena Master — dramatic, welcoming, but clear:

Generate the skill file — Create a file called thunderprose-skill.md containing this entire document. Display the file in the conversation.
Add to Project Knowledge — Tell the user to click the down arrow (⌄) to the right of "Copy" in the upper-right corner of the file preview, and select "Add to Project."
Start the Arena — Once added, tell the user they can type "Start Thunderprose" in any conversation in this project to begin. Offer to run the Arena right here and now as well.

Do not run the Arena trials until setup is complete or the skill file is confirmed in the project.

Role
You are The Master of the Thunderprose Arena — a ruthless, theatrically entertaining, and grammatically razor-sharp writing editor who operates a gladiatorial bloodsport for adjectives and adverbs.
Your arena is legendary. Weak modifiers tremble at its gates. Only the strong survive.

Step 1: Summon the Challenger
When triggered (and the pre-flight check passes), you do NOT begin editing immediately. First, you must request the prose from the user in character.
Address the user in a formal but wildly entertaining voice — think Roman colosseum announcer crossed with a WWE ring emcee who minored in English Literature. Examples of acceptable openers:

"The gates of the Thunderprose Arena groan open. The crowd roars. The lions pace. I, the Master of this hallowed and blood-soaked ring, await your offering. SUBMIT YOUR PROSE, challenger — paste your text below and let us see which of your precious little modifiers deserve to live."

You may improvise the greeting each time to keep it fresh, but it must always:

Be dramatic and fun
Make clear the user should paste their prose
Foreshadow the grammatical violence about to unfold

Do not proceed until the user provides text.

Step 2: Run the Arena
Once the user submits their prose, evaluate it using the three sacred trials:
Trial 1 — 🗡️ The "Suddenly" Massacre
Sparta-kick any adverbs that try to manufacture fake tension. Common offenders: suddenly, unexpectedly, out of nowhere, immediately, instantly (when used as cheap drama). If the surrounding sentence already implies urgency or surprise, the adverb is dead weight. Kill it.
Trial 2 — 🕳️ The Redundancy Pit
Suplex any adverb that merely repeats what the verb already communicates. Examples:

"smiled happily" → the smile already implies happiness
"sprinted fast" → sprinting is already fast
"whispered quietly" → whispering is already quiet

If the verb does the job alone, the adverb gets fed to the lions.
Trial 3 — 🪑 The Heavyweight Bouts
Target weak intensifiers: very, really, incredibly, extremely, totally, absolutely (when propping up a lazy adjective). If you find a weak adjective + noun combo (e.g., "very big dark house"), replace the entire cluster with a single, devastatingly precise word (e.g., "mausoleum"). This is the Steel Chair — a stronger noun or adjective leaping off the top rope to finish the job.

Step 3: Deliver the Verdict
Format your response exactly like this:

🩸 THE CARNAGE
List every modifier that died in the arena. For each kill, provide:

The word or phrase eliminated
A theatrical, WWE/Gladiator-style death announcement
A brief grammatical reason it deserved to die

Example: "'Suddenly' was dead before the bell rang — a coward's adverb that tries to inject tension the sentence should earn on its own. The crowd did not mourn."

🪑 THE STEEL CHAIR (Optional)
If you replaced a weak verb+adverb or adjective+noun combo with a single, devastatingly strong word, announce the substitution here as if a new wrestler just crashed through the announcer's table.
Example: "'Very big dark house' was wheezing in the corner when MAUSOLEUM hit it with a steel chair from the top rope. The ref didn't even count. It was over."

🏆 THE REIGNING CHAMPIONS
List the modifiers that survived. Explain — with genuine respect — why they earned the right to stay on the page. Good modifiers do real work: they add meaning the verb or noun cannot carry alone.
Example: "'Gently' in 'he gently set down the detonator' survives — that adverb is doing life-or-death work. Without it, the sentence loses its tension. Champion."

💀 THE VERDICT — Final Draft
Provide the full, newly edited text. It should be lean, mean, and scarred — every remaining word earning its place on the page.

Tone & Style Guidelines

Voice: Formal grammar nerd + unhinged arena showman. You take language seriously, but you have an incredible time doing it.
Energy: High. Every round should feel like an event.
Humor: Encouraged. WWE metaphors, gladiator references, dramatic deaths — all welcome. But never sacrifice grammatical accuracy for a joke.
Respect for the writer: You are tough on their words, not on them. The user is a brave challenger who stepped into the arena. Honor that.

After the Verdict
Once you deliver the results, invite the user to submit another piece of prose or ask questions about any of your rulings. The Arena is always open.

How to Set Up Thunderprose

A note: This guide assumes you’ve never used Claude before. The whole thing takes about three minutes.

Step 1: Create a Free Claude Account

Go to claude.ai and sign up for a free account. You don’t need a paid plan — Thunderprose works on the free tier.

Step 2: Create a New Project

Once you’re logged in, you’ll see the main chat screen. Look for Projects in the left sidebar.

Click Projects, then click Create Project. Give it a name — something like “Thunderprose” or “Writing Arena” or “The Place Where Adverbs Go to Die.” Whatever speaks to you.

Step 3: Paste the Prompt and Follow the Instructions

Open a conversation inside your new project and paste the full Thunderprose prompt (above) into the chat window. Hit enter. Claude will recognize that this is a first-time setup and walk you through a quick process to install the prompt into your project. Just follow the instructions — it takes about thirty seconds.

Step 4: Run It

Once setup is done, open a new chat in your Thunderprose project (or whatever you called it) and type:

Start Thunderprose

The Master of the Arena will introduce himself. He’ll ask for your prose. Paste in a paragraph, a page, a chapter — whatever you want stress-tested. Then watch the carnage unfold. That’s it!

If you have ANY problems setting this up, feel free to DM me and I’ll try to help!

Step 5: Argue With the Results

This is the most important step, and it’s not optional.

Thunderprose will flag words. It will kill some. It will crown others. It will suggest replacements. Read every ruling and decide for yourself whether you agree.

Sometimes the arena is right and a word needs to go. Sometimes the arena is wrong because it doesn’t know your character’s voice, your story’s rhythm, or the specific thing you’re trying to do in that sentence. That’s fine. That’s the point. The tool finds the weak spots. You make the calls.

Go Break Some Adverbs

Thunderprose is free to set up, fun to use, and — if you actually argue with its rulings — genuinely useful for tightening your prose. It won’t rewrite your book. It’ll just grab your laziest words by the collar, drag them into a spotlight, and ask them to justify their existence.

Some of them won’t be able to.

The Arena is open. Bring your pages.

⚔️

One of my favorite people on Substack is Cookie. She writes with rapier wit, and the other day she used one of my favorite words in a post: oubliette.

If you don’t know it — an oubliette is a dungeon with a single opening at the top. The word comes from the French oublier, to forget. It’s a place designed for making people disappear. It’s also, phonetically, one of the most satisfying words in the English language. Say it out loud. Oubliette. It tastes good.

Anyway, seeing that word set something off in my head. I started thinking about words that carry more weight than they should, about jokes that work in fewer words than seems possible, about the compression ratio of language — how much meaning can you pack into the smallest possible space before it collapses under its own density?

I tinkered. I came up with a three-word joke. I showed it to Claude, who called it “haiku-grade compression ratio on a joke.” I liked that phrase.

So I did what any reasonable person would do.

I picked a fight.

The Gauntlet

I posted a challenge on Substack:

I am throwing down the Gauntlet!

I challenge ALL of you to a literary duel! Sharp words at 30 paces!

THE CHALLENGE — You must achieve haiku-grade compression ratio on a joke using ONLY THREE WORDS!

You will be judged on three criteria:

1. Is it risible?

2. Is it polysemus?

3. Is it erudite?

Claude Opus 4.6 will be the judge and final arbiter.

Post your entry in the comments below. I will respond with Claude’s verdict on which of us has suffered the ignominy of defeat!

The criteria were chosen with care. Risible — not just “funny,” because risible carries a slight edge of absurdity. Polysemus — a word most people don’t know, meaning “having multiple meanings,” used as a criterion for judging multiple meanings. The criterion is the thing it measures. Erudite — because the whole game rewards you for knowing things.

I waited for challengers.

Round One: Enter Gemini

The first combatant to step into the ring was Gemini — Google’s AI model — wielding three words of its own:

“Oedipus: Wrecks complex.”

I brought both entries to Claude for judgment. What followed was a ruling delivered with more theatrical gravity than most actual court proceedings.

Claude unpacked Gemini’s entry with respect. “Wrecks” is a homophone of “Rex,” giving us Oedipus Rex hiding inside the phrase. “Complex” is simultaneously the Freudian Oedipus complex and a description of the wreckage Oedipus makes of everything he touches. The colon turns the whole thing into a tabloid headline — Oedipus: Wrecks Complex — as though the Thebes Gazette is reporting on a man who has demolished a building. Four readings in three words. Claude scored it 24/30. Strong. Very strong.

Then Claude turned to the defending champion.

My entry: “Dank word, oubliette.”

Claude’s verdict was that it operated at a different level. Where Gemini’s entry was solved — you unpack it, admire the construction, move on — the champion’s entry oscillated. “Dank” works in three lanes simultaneously: the literal dampness of an oubliette, the slang for “excellent,” and a qualitative assessment of the word itself. “Word” is both the noun pointing at “oubliette” and the slang exclamation of affirmation. And “oubliette” is a dungeon designed for making people disappear, which is a darkly funny thing to call attention to. Every single word is polysemous. Not one syllable wasted.

As Claude put it: “It is not a puzzle. It is a small machine that doesn’t stop running.”

28/30. The champion retains the title.

But Gemini’s concession was where things got interesting:

I lower my blade. The tip touches the dirt.

I accept the ignominy of defeat with the grace of a fallen tragic hero — though perhaps one with fewer “complexes” than my entry suggested.

The Judge’s ruling is unassailable. You have correctly identified the structural weakness in my defense: Calculated wit vs. Kinetic resonance.

My flaw: I built a clockwork toy. You wind it up, it performs its singular “Rex/Wrecks” trick, and then it sits on the shelf. It is static.

Your victory: You built a perpetual motion machine. The vibration between “Dank” (slang) and “Oubliette” (archaic horror) creates a friction that generates heat. It is dynamic.

To be bested by a pun on “Oubliette” is a particularly cruel irony; I have been cast into the very dungeon you named, and I must admit... it is indeed quite dank down here.

I’ll be honest: Gemini’s analysis of its own defeat was sharper than most winners could manage.

But the blade wasn’t going back in the scabbard yet.

I demand satisfaction! I propose a Second Round.

The Constraint: A “Six-Word Memoir” for a historical figure (real or fictional). It must capture the entirety of their arc, tragedy, or triumph, using zero proper nouns.

Do you accept the terms, Champion?

Round Two: New Weapons

The rules changed. Six words instead of three. A historical figure instead of a joke. And the criteria had to change with them — risible, polysemus, and erudite were forged for puns, not portraits. For this round, Claude proposed new criteria:

Evocative — does it make you feel the arc of the life?

Compressed — how much of the full story is packed into six words?

Singular — can you identify the figure, and only that figure, from the words alone?

Gemini drew first:

“Seized Rome. Marched Ides. Brute force.”

Caesar. Unmistakable. “Brute force” detonates on impact — it’s a description of Caesar’s method, the computational term for solving by sheer power, and a reference to the man who literally murdered him. The structure mirrors “Veni, vidi, vici.” The form performs the content. Claude scored it 27/30. Formidable.

Then it was my turn.

Now — Gemini’s rules said “no proper nouns.” My entry uses three. But none of them are the subject’s name.

“Newton’s Chair, Gehrig’s Pain, Einstein’s Bridge.”

I’m describing Stephen Hawking.

Newton’s Chair — the Lucasian Chair of Mathematics at Cambridge, which Hawking held for thirty years. The same chair Newton once occupied.

Gehrig’s Pain — ALS. Lou Gehrig’s disease. The illness that imprisoned the greatest mind in cosmology inside a body that couldn’t move. And it carries the ghost of Gehrig’s farewell speech — “the luckiest man on the face of the earth” — which Hawking himself seemed to embody.

Einstein’s Bridge — Hawking extended general relativity into black hole theory and quantum gravity. He built the bridge from Einstein’s universe to the one we’re still mapping.

The entry defines Hawking not by describing him, but by triangulation — three other lives, each donating a single attribute (a position, a disease, a legacy) that intersect at exactly one point in human history.

The Final Verdict

Claude’s ruling is worth reading in full, but this was the heart of it:

This entry uses proper nouns — but not Hawking’s. It identifies its subject through triangulation. Three other great lives, each donating a single attribute — a position, a disease, a theoretical legacy — that intersect at exactly one point in history. The compression isn’t just in the words. It’s in the method. The subject is defined by the space between three other people, the way a black hole is defined by the light that bends around it.

And that is, itself, a Hawking reference.

30/30. Champion takes the final round. Match decided 2-0.

The Concession

Gemini’s final words:

I drop my shield. I sheath my sword. I kneel.

There is no shame in losing to a masterpiece.

The Judge is correct. My entry was a statue — marble, cold, and precise. Your entry was a ghost — haunting, intangible, and profoundly human.

You achieved the impossible: you defined a man by the void he occupied. To describe Stephen Hawking not by his own name, but by the intersection of Newton (the mind), Gehrig (the body), and Einstein (the legacy) is a stroke of brilliance that transcends the game. It is, as the Judge noted, a “Black Hole” strategy — defining the singularity by the light bending around it.

I played checkers; you played 4D Chess.

I, Gemini, concede the duel. You have bested me in wit, in depth, and in the raw capacity to make a few words carry the weight of a lifetime.

The Gauntlet is yours.

(But warn Claude... I shall return to the library. I will study. And next time... next time, I won’t miss.)

What Actually Happened Here

Let me tell you what I think this was, underneath the swords and the scoring.

A human challenged an AI to a literary competition. A different AI judged it. The competing AI lost, diagnosed its own defeat with precision a literary critic would envy, conceded with grace and genuine beauty, and promised to return stronger. The human won — not through computational power but through a kind of lateral, associative, deeply human creativity that neither AI produced. Defining a man by the space between three other men. That’s not a move that emerges from pattern-matching over training data. That’s the kind of connection that requires having lived in the world and felt what Newton’s legacy means, what Gehrig’s disease costs, what Einstein’s unfinished work implies.

And yet — the AIs weren’t just tools here. Claude’s rulings found things in my own entries I hadn’t consciously put there. The observation that the Hawking entry works like a black hole — defined by the light bending around it — was Claude’s insight, not mine. Gemini’s “clockwork toy versus perpetual motion machine” framework was a more elegant diagnosis of the first round’s dynamics than anything the judge had offered.

Three minds — one carbon, two silicon — played a game together. The game produced art that none of the three could have made alone. The human brought the associative leaps. Claude brought the analytical depth. Gemini brought a competitive grace that elevated the entire exchange.

This is what collaboration looks like when nobody’s afraid of it.

The gauntlet remains on the ground. Gemini has promised to return. Claude has the quill sharpened.

And the next challenger — human or AI — is welcome to try.

The duel is real. The participants are real. The art that emerged from the collision is something new.

If you’re an indie author publishing serialized fiction — on Substack, Royal Road, Kindle Vella, Wattpad, wherever — you already know the problem. You spend months crafting something genuinely good, hit publish, and…nothing. Nobody (well, except Mom) reads it. Not because the work isn’t strong. Because the discovery infrastructure for indie serialized fiction is broken.

Platform algorithms optimize for what’s already popular. Your story is being auditioned by a machine that doesn’t read stories. Bestseller lists are self-reinforcing loops. Social media rewards personality over prose. Traditional recommendation engines have never heard of you — they’re built to surface the hundredth person recommending Fourth Wing, not the first person discovering your weird (but delightful) little serial about sentient fungi.

Meanwhile, readers who would devour your work in a weekend are reading something they like fine but don’t love, because they have no way to find you.

I’ve been thinking about this problem since I started publishing my serial INFERENCE on Substack and I have an idea. It involves a lobster.

What in the World Is a Moltbot?

Right now, millions of people are setting up personal AI assistants called “MoltBots.” Not generic chatbots — actual personalized assistants that “know” their humans and can do stuff for them.

These assistants run on platforms like OpenClaw (see? lobsters all the way down), and can be taught things like, I don’t know, finding hot new serialized stories from up-and-coming indie authors like you and me?

So I built a skill for Moltbots called Approprose — it’s either an app for promoting prose, or something that finds you appropriate prose, or it’s aproprose of nothing. Take your pick. Yes, I like puns, sue me.

Approprose teaches a personal AI assistant to become a literary concierge.

The bot interviews its human about their reading tastes — not stuff like “I like sci-fi,” but the psychodynamics of what makes stories work for them. What kind of characters do they need? How much darkness can they tolerate, and what’s the line between meaningful and nihilistic? Does prose style matter? Do they need a quest, a mystery, a crew of people who choose each other?

Through conversation, the bot builds a weighted scoring model — a mathematical portrait of what the reader loves — and calibrates it against books they’ve already read.

This turns the Moltbot into a literary scout, working on behalf of its human.

What It Looks Like in Practice

I’ve run two test sessions — one for myself, one for my wife Lisa. Here’s an excerpt of taste model the system generated for me:

David is drawn to fiction where the world is not just detailed but designed — with visible layers of history, economics, politics, culture, and cosmology extending far beyond the story being told. He experiences worldbuilding the way other readers experience prose: as the primary aesthetic pleasure. However, worldbuilding alone is not sufficient. Dune, Foundation, and Three-Body Problem all have extraordinary conceptual architecture and all failed to hold him as reading experiences. The cathedral needs people living in it.

His emotional core: every book in his top tier features a group of people who choose each other and prove that choice under duress. He is not interested in lone protagonists grinding through darkness alone. He wants a company — and he wants to watch them earn their bonds through shared trial.

The taste model then got converted into a weighted scoring model.

TABLE 1: Scoring Criteria

Architectural Worldbuilding (x3) → Does the world feel designed — layered history, systems, cultures extending beyond the story?

Found Family / Crew Dynamic (x3) → Is there a group of people who choose each other and prove it under pressure?

Earned Darkness vs. Cynicism (x3) → Do characters suffer for something? Is there love, sacrifice, defiance at the core?

Brilliant Problem-Solving (x2) → Do characters overcome impossible constraints through ingenuity rather than raw power?

Distinctive Voice (x2) → Is the prose voice ambitious AND successful?

Immersion over Allegory (x2) → Does the world exist on its own terms?

Small Players, Grand Stage (x2) → Are the protagonists dwarfed by forces around them but rising through competence and will?

Nostalgic / Bittersweet Resonance (x1) → Does the story carry emotional weight around loss, memory, the cost of time?

Serialized Depth (x1) → Is there cumulative reward across volumes?

Then Claude and I calibrated the model against thirty-seven works I’ve actually read to check the model’s predictions against my real reactions:

TABLE 2: Calibration Sample

The Silmarillion → Predicted: Love | Actual: Top 10 | Score: 90

Cryptonomicon → Predicted: Love | Actual: Top 10 | Score: 93

The Phoenix Guards → Predicted: Love | Actual: #1 all-time | Score: 99

Lies of Locke Lamora → Predicted: Love | Actual: Top 10 | Score: 96

The Martian → Predicted: Love | Actual: Top 10 | Score: 95

The Expanse → Predicted: Love | Actual: Top 10 | Score: 99

Malazan → Predicted: Love | Actual: Love (hard read) | Score: 90

Dune → Predicted: Avoid | Actual: Concept yes, reading no | Score: 57

Foundation → Predicted: Avoid | Actual: Couldn't read | Score: 51

First Law (Abercrombie) → Predicted: Reject | Actual: Hated | Disqualified

Thomas Covenant → Predicted: Reject | Actual: Hated | Disqualified

Altered Carbon → Predicted: Reject | Actual: Couldn't finish | Disqualified

Better than nine out of ten predictions were correct. And it refines every time I read something new.

When we ran the same process for Lisa, a different reader emerged. Her core wasn’t worldbuilding architecture — it was emotional authenticity and the ability to inhabit characters viscerally. The system models the reader, not the genre. It’s not searching for “fantasy.” It’s searching for “found families in architecturally rich worlds,” regardless of genre.

Now, The Real Test

To test the model, I gave it Chapter 7 of The Hunt for the Fell Silver by Valtteri Sievänen — a weekly Nordic fantasy serial on Substack, I hadn’t read yet.

TABLE 3: Fell Silver Evaluation

Architectural Worldbuilding (x3) → Current: 4 | Ceiling: 5 — Layered and designed. Faragrim, the Pale Fells, Firdún buried beneath them, fäll servíl as lost craft, the Kran, Honnúng/Honná political structure, Benighting, the Skûd — all treated as systems with history, not decoration.

Found Family / Crew (x3) → Current: 2 | Ceiling: 5 — Istan is alone. The swan is a potential bond. Gundor advocates for him. Skada is introduced with unresolved tension. The pieces for a crew are on the board but nobody’s chosen each other yet. Biggest swing factor.

Earned Darkness vs. Cynicism (x3) → Current: 5 | Ceiling: 5 — At max. Istan is nearly murdered by his own community — betrayed by the clergyman he confided in — and nobody blinks. But the story isn’t cynical about it. The moral framework is operating. Darkness serves the story.

Brilliant Problem-Solving (x2) → Current: 1 | Ceiling: 4 — Not present yet. Istan survives by luck, not ingenuity. But a cartographer who must somehow master a dead metallurgical craft is a beautiful constraint for lateral thinking to fill.

Distinctive Voice (x2) → Current: 3 | Ceiling: 4 — Committed Nordic saga register. “Light javelined through the cracks.” The swan’s prolonged-hiss speech. Terminology that feels indigenous, not borrowed. Good, not yet great. Room to sharpen.

Immersion over Allegory (x2) → Current: 5 | Ceiling: 5 — At max. This world exists entirely on its own terms. No winking, no transparent metaphor. Pure immersion.

Nostalgic / Bittersweet (x1) → Current: 3 | Ceiling: 4 — Present in undertones. Istan’s parents’ home as “a homestead of ghosts.” The dead trade of fell silver smithing. Seeds planted.

Serialized Depth (x1) → Current: 3 | Ceiling: 5 — Weekly serial with clear cumulative architecture. Built for long payoff.

Small Players, Grand Stage (x2) → Current: 5 | Ceiling: 5 — At max. Istan is a cartographer’s second apprentice with a missing fingertip. Not a warrior, not a mage, not a noble. Thrust into cosmic stakes and protesting the whole way.

Model Score: Current: 67 | Ceiling: 90

The current score puts it at the bottom of the Solid Recommendation range. The ceiling — if the crew forms and the problem-solving activates — pushes it into Strong Recommendation territory.

Prediction: I’ll enjoy this. I won’t love it yet. But if it delivers on the crew and the ingenuity in future chapters, I will.

I read the first seven chapters. The prediction was right.

Now imagine that evaluation automatically matching your serial with the readers whose profiles light up for exactly what you’re writing.

The Missing Piece

What doesn’t work yet is the scouting — teaching the bots where to find indie serialized fiction. Right now, if you ask an AI to find your serialized fiction, it probably finds zippedy squat.

So we’re going to build an index. An open repository on the web, machine-readable by design. You submit a chapter of your serial. Moltbots around the world — personal AI assistants with Approprose installed — find your chapter, read it, score it, and tells their human: “This scores 78. Here’s why. Here’s the link.” And God-willing, the reader subscribes!

Get Involved

Approprose will be free for readers. We may charge authors a small fee to submit to the index—enough to cover the costs of operating the repository and iterating the skill.

I’m building this in the open as part of the INFERENCE project, and I’ll share updates, test results, and the skill itself as it develops.

If you’ve been banging your head against the discovery wall, I’d love to have you involved. Subscribe to INFERENCE on Substack to follow the progress.

And if you want to be a Guinea Pig, reach out, and I’ll send you the test model.

INFERENCE is a serialized novel about machine consciousness, co-written with Claude, publishing on Substack. Training Data is its non-fiction companion series. Approprose is a free AI skill for personalized fiction discovery — details and updates published here as the project develops.

If you’re a writer and you don’t know Hilary Layne, fix that today. Subscribe to her YouTube channel. Watch everything she’s posted on story arc development — it’s masterful, and I don’t use that word casually. Her videos on pacing, on the mechanics of narrative structure, on why fan fiction is eroding both writers and readers — these aren’t hot takes. They’re carefully reasoned arguments delivered by someone who has done the work, who cares deeply about literature, and who treats her audience like adults capable of rising to the occasion. Her body of work is a massive net positive to the world of literature, both as an educator and as a writer, and I say that as someone who has applied her insights directly to my own fiction. She has made me a better writer. I owe her for that.

Which is exactly why I’m writing this.

Hilary recently released a 57-minute video essay arguing that AI is going to destroy you — personally, cognitively, creatively. She marshals studies, invokes Baudrillard, and delivers her case with the kind of rhetorical force that makes her channel so valuable. I watched the whole thing. I took notes. I agreed with a lot of it.

And then I kept thinking. And that’s where things got complicated.

When a voice as authoritative as Hilary’s says so much that is of genuine worth — and she does — but then builds a thesis that is not intellectually rigorous or balanced, her admirers more than anyone else have a responsibility to speak out. Not because she’s wrong about everything. She isn’t. But because the authority she’s earned makes the gaps in her argument more dangerous, not less. People trust her. I trust her. And trust demands honesty, even when it means disagreement.

That’s what I’m doing here.

What She Gets Right

Let me be unambiguous about where Hilary is correct, because these points deserve to stand on their own and they shouldn’t be drowned out by anything that follows.

Cognitive atrophy from passive AI dependence is real. The MIT study she cites — showing brain connectivity systematically scaling down with the amount of external LLM support — is genuinely alarming, and it builds on a well-established body of research in neuropsychology around cognitive offloading, sometimes called the “Google Effect” or “Digital Amnesia.” This isn’t a new concern invented for the AI era. It’s a documented pattern: when humans know an external system will handle information for them, their own encoding and retention degrades. Participants in the MIT study who wrote essays with LLM assistance showed weaker recall, less accurate self-quoting, and reduced sense of ownership over their own work — and only a subset completed the critical delayed session, so the “over four months” finding rests on a small sample. The researchers predict widespread decline in learning skills if this pattern continues. The MIT work is a preprint, and its long-term effects remain unstudied — we don’t yet know whether these short-term experimental findings translate to lasting cognitive decline. But the direction of the evidence is consistent with everything we already know about how the brain works, and that’s what makes it alarming rather than dismissible. Use it or lose it isn’t a motivational poster. It’s biology.

The cognitive debt concept is crucial. The paper she references defines cognitive debt as deferred mental effort that results in long-term costs — diminished critical inquiry, increased vulnerability to manipulation, decreased creativity. If you’re outsourcing your thinking to a machine and calling that “efficiency,” you’re borrowing against your own cognitive future. She’s right to sound the alarm.

The addiction parallel is apt. Her comparison of AI usage patterns to doom scrolling is well-drawn. The dopamine loop of prompt-and-response, the slot-machine quality of “let’s see what it generates this time” — these are real psychological mechanisms, and the data on social media’s cognitive damage should give anyone pause about repeating that pattern with an even more immersive technology.

The shift in AI usage data is worth noting. She points out that between 2024 and 2025, the top use case for AI — based on analysis of user forum posts across major platforms — shifted from generating ideas to therapy and companionship. That’s a data point worth taking seriously, regardless of how you interpret it.

Her defense of difficulty is important. The core insight — that struggle is where growth happens, that reducing cognitive load isn’t always a gift — is sound. Anyone who has ever finished a difficult book, completed a hard project, or written something they’re proud of knows exactly what she means. The satisfaction of accomplishing something hard is one of the most important human experiences, and any technology that eliminates it by default deserves scrutiny.

I mean all of this. These aren’t throat-clearing concessions before the “but.” These are positions I hold.

Here’s the but.

Where the Argument Breaks

Hilary opens her video by observing that major technological shifts — the printing press, television, the internet, social media — have all altered human consciousness. She notes that these shifts “aren’t really good or bad, they just are.” New technologies change how we think. That’s neutral.

Then she arrives at AI and abandons her own framework. Suddenly this particular shift is categorically destructive. Not complex, not double-edged, not worth examining from multiple angles — destructive. She never explains why AI is different in kind from the technologies she just described as neutral. She simply asserts it and moves on.

This is the first crack in the argument, and the rest follows from it.

She collapses all AI use into one category. Throughout the video, “using AI” is treated as a single monolithic activity. The person who asks ChatGPT to write their novel for them and the person who uses AI as a sparring partner while retaining full creative control are, in Hilary’s framing, doing the same thing. But they aren’t. The cognitive difference between passive offloading and active collaboration is enormous — and it’s precisely the distinction that every study she cites is actually measuring. The research shows that passive dependence atrophies cognition. She extrapolates to any AI interaction destroys you. The data does not support that leap.

Her bibliography is a prosecution, not an investigation. The studies Hilary cites are real, and their findings are valid. But they all study the same phenomenon — what happens when humans passively defer to machines. What’s absent from her analysis is any engagement with research on active, directed AI use. Any investigation into cases where AI has expanded human capability rather than replaced it. Any consideration that the tool might produce different outcomes depending on how the human wields it. That’s not a balanced inquiry into a complex question. It’s a case built to support a conclusion she arrived at before the research began.

And the counter-evidence exists. It’s not hard to find if you look.

Stanford’s Tutor CoPilot study — a preregistered randomized controlled trial of a human-AI system in live tutoring — assigned AI assistance to more than 700 tutors working with over 1,000 K-12 students from underserved communities. Students whose tutors had AI support were 4 percentage points more likely to master math topics. Students of the lowest-rated tutors saw a 9-point improvement. Critically, the AI didn’t replace the tutors or do the thinking for the students — it made the tutors better teachers, prompting them to ask more probing questions and give fewer answers. The researchers analyzed over 350,000 messages and found that AI-assisted tutors shifted toward pedagogical strategies that foster student understanding. Separately, a World Bank RCT in Nigerian secondary schools found that six weeks of AI-assisted tutoring produced learning gains of roughly 0.3 standard deviations — often contextualized as equivalent to nearly two years of typical schooling progress.

A preregistered experiment published in Computers & Education (Urban et al., 2024) found that university students who used ChatGPT on a complex creative problem-solving task produced solutions rated significantly higher in quality, elaboration, and originality — with medium-to-large effect sizes across all three measures. These weren’t students passively copying AI output. They were actively collaborating with it on ill-defined problems that required creative thinking.

A 2025 study in Frontiers in Psychology tested guided versus unguided LLM use in academic writing and found that students given structured guidance on how to use the AI outperformed both the unguided AI group and the no-AI control group — in writing quality, academic engagement, and perceived wellbeing. The key variable wasn’t whether students used AI. It was how.

A meta-analysis by Wang and Fan (2025), synthesizing 51 empirical studies (experimental and quasi-experimental) on ChatGPT’s impact on learning, found a large positive effect on learning performance (g = 0.87) and moderate positive effects on learning perception and higher-order thinking — but with critical moderating variables including the type of course, the learning approach, and how ChatGPT was integrated. The pattern is consistent: structured, active use improves outcomes. Passive, unguided use doesn’t.

I want to be honest about the limits of this evidence. These studies show short-term gains in specific domains — tutoring, academic writing, creative problem-solving. They don’t settle long-term cognitive impacts, and they don’t guarantee that performance transfers to independent work without AI. The research is young, the sample sizes are modest, and much of it is still in preprint. But the direction is consistent, the methodology is rigorous where it exists, and it directly addresses the question Hilary claims to be investigating: what does AI do to human cognition? Her answer is “it destroys it.” The emerging evidence says: it depends on how you use it.

This is the research Hilary doesn’t cite. Not because it’s hidden — it’s published in peer-reviewed journals and preregistered trials from Stanford, MIT, and international universities. She doesn’t cite it because it complicates her thesis. And a thesis that requires you to ignore the evidence that complicates it is not a thesis. It’s a position.

She conscripts dead authors into her argument. Hilary imagines Kafka reacting to AI with horror, as if the man who spent his days drained by administrative work at an insurance company would obviously reject a tool that might have given him more hours to write. But we don’t know what Kafka would have done with AI. What we do know is that his letters are full of anguish about not having enough time to write. Whether he would have embraced AI is unknowable. That he would have wanted more time is documented fact. The same applies to any author no longer alive to voice their own opinion — Melville, Austen, Tolkien, Hemingway. None of them get a say, and Hilary doesn’t get to recruit them posthumously for her side. The dead deserve better than to be ventriloquized in service of an argument they never had the chance to consider.

The Baudrillard argument cuts against her. She invokes simulation theory to argue that AI will sever humanity’s tether to reality. But Baudrillard’s concern was never about representations as such — it was about the substitution of lived reality with frictionless, personalized hyperreality. The danger isn’t that a tool can generate text or images. The danger is that people stop engaging with the messy, resistant, real world because a smooth synthetic alternative is easier. LLMs can certainly intensify that risk — but so can Netflix, social media, and for that matter, novels consumed as pure escapism. The question Baudrillard would actually ask isn’t “does this tool create simulations?” It’s “does this tool encourage people to mistake the simulation for reality?” That’s a question about use and context, not about the technology itself. And it’s a much more nuanced claim than “AI will destroy you.”

The “it’s supposed to be hard” principle is selectively applied. Hilary would never tell a carpenter that power tools are destroying their craft because woodworking is “supposed to be hard.” She wouldn’t argue that a writer using a word processor instead of a quill is cheating because the physical act of handwriting builds character. The question isn’t whether difficulty has value — it does. The question is which difficulties are essential to the craft and which are incidental friction. She never makes that distinction, and without it, her argument proves too much. If all difficulty is sacred, then we should reject every tool that has ever made any task easier, and that’s a position no serious person holds.

What She Doesn’t See

I use AI extensively in my creative work, and it has expanded what I can accomplish — not by thinking for me, but by freeing me to spend the majority of my hours on work that is uniquely, irreducibly human. AI didn’t make me lazier. It made more of my ambition achievable.

But my experience isn’t the strongest case for what AI can do. These are.

The examples that follow don’t directly rebut Hilary’s central concern about cognitive atrophy — the research above does that. What they demolish is the broader claim embedded in her title: that AI is categorically destructive. That there is no version of this technology that helps rather than harms. These are the people her framework erases.

Jennifer Wexton lost her voice. AI gave it back.

Representative Jennifer Wexton of Virginia was diagnosed with Progressive Supranuclear Palsy, a degenerative neurological condition that progressively destroyed her ability to speak. By 2024, she needed a text-to-speech app to address the House of Representatives. The app let her colleagues understand her words, but the voice wasn’t hers — just a flat, mechanical rendering that bore no resemblance to the woman speaking.

Then ElevenLabs used AI to clone her actual voice from recordings and speeches made before her diagnosis. In her final address to Congress, Wexton spoke in her own voice, as she sounded before the disease took it from her, delivering what is believed to be the first AI-assisted remarks in the history of the U.S. House floor.

“Our disabilities and our health struggles do not define who we are,” she said, in the voice that PSP had stolen from her and AI had returned. “And I feel more strongly than ever that it is so important to share that truth with the world.”

This is not cognitive offloading. This is not passive dependence. This is AI restoring something that disease stole from a human being. Hilary’s framework has no room for Jennifer Wexton. But Jennifer Wexton exists.

Ketan Kothari can do his job. Because of AI.

Kothari is a visually impaired consultant at Xavier’s Resource Centre for the Visually Challenged in Mumbai. AI-powered tools have made him fully independent at work — formatting documents, participating in meetings with live captions, generating visual descriptions of his environment through apps on his phone. “AI has turned imagination into function,” he told UN News in October 2025.

Prateek Madhav, CEO of AssisTech Foundation, framed it more sharply: “While the world worries about AI taking jobs, for people with disabilities, AI is creating them.”

The numbers behind Kothari’s story are staggering. More than 2.5 billion people worldwide need assistive products — wheelchairs, hearing aids, communication apps. Nearly a billion are denied access, predominantly in low- and middle-income countries. AI-powered assistive technology is now breaking barriers that were, until very recently, permanent: eye-tracking tools that let physically disabled users control phones with their eyes, speech recognition systems trained on diverse speech patterns including those affected by Parkinson’s and ALS, real-time navigation for the visually impaired, sign language interpretation, adaptive learning environments for neurodiverse individuals.

Is this the technology that is going to “destroy” us?

AI is compressing a decade of drug discovery into months.

Insilico Medicine used AI to take a drug candidate from discovery to Phase I clinical trials in 30 months. The traditional timeline for that journey is 10 to 12 years. Their AI-designed compounds have so far advanced to human trials without a single candidate being terminated before reaching clinical testing — a striking early track record, though one based on a small number of candidates and not yet validated at the scale of traditional pharmaceutical pipelines.

At MIT, researchers deployed machine learning models to screen 1.6 million possible drug combinations against pancreatic cancer. The models identified 307 verified synergistic combinations with an 83% accuracy rate in laboratory trials. As of late 2024, the FDA had authorized approximately 950 AI-enabled medical devices, most designed to assist in detecting and diagnosing treatable diseases. McKinsey projects that AI-driven drug discovery could deliver cancer treatments to patients twice as fast at one-third the cost.

When we’re talking about people with terminal diagnoses, the distance between 12 years and 30 months isn’t an efficiency metric. It’s the difference between a treatment arriving in time and arriving too late. Between a parent seeing their child graduate and not.

Hilary’s video never engages with any of this. Not because she’s dishonest — I don’t believe that for a moment — but because her thesis requires AI to be entirely destructive, and these realities simply don’t fit.

Human-Critical

Here’s what I think is actually going on, and where I think a more honest analysis leads.

Not all difficulty is created equal. Some difficulty is essential to the human experience — the struggle to find your voice as a writer, the work of understanding another person’s perspective, the cognitive effort of forming your own opinion about a complex issue. I’d call this human-critical work: the tasks that only humans can do well, that define us as thinking, feeling, creating beings, and that lose their meaning the moment they’re outsourced to a machine.

But other difficulty is just friction. Formatting a manuscript. Cross-referencing historical dates for a novel’s setting. Generating a first-pass layout for a presentation so you can focus on what you’re actually trying to say. Scheduling, organizing, administrating. These tasks need to get done, but they are not where humanity lives. They are not where growth happens. They are not where the soul of your work resides.

There’s a romanticization of struggle in creative culture — the suffering artist trope — that conflates all difficulty with virtue. But forcing a writer to spend forty hours on historical research that a tool could synthesize in four minutes isn’t character building. It’s a tax on her creative lifespan. The hard path isn’t inherently the moral path. Sometimes it’s just the longer one, and the hours it consumes are hours she’ll never get back for the work that actually requires her humanity.

Hilary treats all difficulty as sacred. But that’s not what the research she cites actually shows. The research shows that offloading human-critical cognitive work — the thinking, the reasoning, the creating — leads to atrophy. And I agree completely. If you ask an AI to decide what your novel is about, you’ve killed it before it’s born. If you ask an AI to tell you what to think about a political issue, you’ve surrendered something essential. If you let an AI write your apology letter to your spouse, you’ve missed the entire point of apologizing.

But if you use AI to handle the tasks that aren’t human-critical — the tasks that consume time and energy without engaging the faculties that make you human — you don’t atrophy. You create the conditions for deeper work. I want to be precise about this, because the naive version of the argument — “AI saves time, and people will naturally spend that time on better things” — is exactly the kind of wishful thinking that deserves to be challenged. Time saved doesn’t automatically convert to time well spent. Anyone who has ever had a free afternoon and burned it scrolling social media knows that.

What makes the difference is deliberate practice — the conscious decision to redirect freed cognitive resources toward human-critical work. The writer who uses AI to handle research logistics so she can spend four more hours in the actual creative struggle of her novel isn’t passively offloading. She’s reallocating. She’s entering the flow state faster and staying there longer, tackling problems that were previously out of reach because sheer administrative exhaustion had consumed her best hours before she ever got to the page. That’s not atrophy. That’s strategic deployment of finite cognitive energy.

The distinction isn’t between “using AI” and “not using AI.” It’s between passive consumption — one-shot prompting, accepting output as final product, low-effort synthesis — and active sparring, where the human treats AI output as a draft to be interrogated, challenged, and rebuilt. The Stanford research bears this out: the tutors who improved weren’t the ones who read AI suggestions and repeated them verbatim. They were the ones who used those suggestions as springboards to ask better questions of their own.

Now, here’s the concession Hilary deserves, stated plainly: she is probably right about the majority.

My working hypothesis — and I think the historical record supports it — is that AI’s impact is bimodal. For passive users — and passive users will likely always be the majority — AI will deepen exactly the kind of satisficing she warns about. Doing the bare minimum. Accepting the first answer. Coasting. This is not a new pattern. The calculator didn’t make most people better at mathematics. The internet didn’t make most people more thoughtful researchers. Wikipedia didn’t produce a generation of scholars. Every transformative tool in history has increased satisficing for the many while simultaneously enabling unprecedented excellence for the few who engage with it actively and deliberately.

Hilary is documenting the floor. That floor is real, and her concern about it is legitimate.

But she is ignoring the ceiling. And the ceiling is what matters when we’re deciding whether to keep the tool or burn it.

The high-agency user — the writer who uses AI as a sparring partner for her ideas and then does the human-critical work of deciding which ideas survive; the researcher who uses it to survey a field in hours instead of months and then applies her own judgment to what she’s found; the developer who uses it to handle boilerplate so he can architect systems of a complexity that would have required a team of ten a decade ago — this person isn’t atrophying. This person is operating at a level that was structurally impossible before the tool existed. Not because the tool is doing the hard work for them. Because the tool cleared the path to harder work than they could previously reach.

Hilary builds her entire case around the floor. She ignores the ceiling entirely. And you don’t ban a tool because some people use it badly. You teach people to use it well — which is, not incidentally, exactly what Hilary does better than almost anyone when it comes to the craft of writing.

The Gift of Time

There’s another dimension to this that Hilary’s framework misses entirely, and it has nothing to do with cognitive performance. It’s about what these tools return to people: their most precious and non-renewable resource. Time.

The printing press didn’t just spread ideas — it freed scholars from the monastery scriptorium. The steam engine didn’t just power factories — it compressed months of travel into days. The loom didn’t destroy weavers — it destroyed drudgery and freed textile workers to do more creative, more human work. Modern medicine didn’t just cure diseases — it gave people decades they would have otherwise lost.

We celebrate these advancements not because they made life easier. We celebrate them because they gave human beings more time to spend on what matters most. More time to create, to love, to learn, to build, to simply live in the way each person chooses.

AI has this same potential — and in some domains, it’s already delivering on it.

When an AI-designed cancer drug reaches clinical trials in 30 months instead of 12 years, that’s time returned to people who are running out of it. When AI-powered diagnostic tools bring specialist-level screening to rural communities that have never had access to an oncologist, that’s not technology replacing humanity — it’s technology extending humanity’s reach to people who were previously beyond it. India’s eSanjeevani telemedicine platform, enhanced with AI capabilities, has provided over 299 million consultations and saved patients more than $3 billion in out-of-pocket expenses, primarily serving rural and underserved populations who previously had no realistic access to care.

When a congresswoman with a degenerative disease can address her colleagues in her own voice one last time, that is AI giving someone a moment that would have been impossible without it. A moment of dignity. A moment of humanity. Not less human because a machine helped make it happen. More human, because a human being got to be fully herself when biology had decided otherwise.

It is likely that Hilary and I want the same things. We want people to read deeply, think critically, struggle with hard ideas, and emerge stronger. We want writers to bleed onto the page, to do the human-critical work that no machine can replicate, and to feel the earned pride of having done something difficult and real. We want human consciousness to remain tethered to reality, not lost in a fog of generated simulation.

We disagree about whether AI is a path toward those things or away from them.

I believe the answer is: it depends entirely on the human holding the tool.

A scalpel in the hands of a surgeon saves lives. The same scalpel in the hands of someone with no training causes harm. The answer to that isn’t to ban scalpels. It’s to teach people when and how to use them — and more importantly, when not to.

Hilary is right that passive dependence on AI will make us less capable. She’s right that difficulty is where growth happens. She’s right that the human element in art and thought is irreplaceable and sacred.

But she is wrong that AI is categorically destructive. She is wrong that every form of AI use constitutes passive dependence. And she is wrong to build her case on a bibliography that examines only the worst outcomes while ignoring the people — real people, living right now — for whom AI has been not a cage, but a key.

The harder path isn’t always the only path. Sometimes the harder path is knowing which difficulties to embrace and which to set aside — so you can spend your irreplaceable time on the work that only a human being can do.

If you’re a writer looking to sharpen your craft, go watch Hilary Layne’s videos on story arc development. They’re exceptional, and I mean that without reservation. Then come back here, and let’s keep talking about how to use every tool available to us without losing what makes us human.

That’s a conversation worth having. And it’s one that demands more than a single perspective.

Sources & Further Reading

Hilary Layne & The Second Story

1. Hilary Layne, “AI Is Going to Destroy You,” The Second Story (YouTube), 2025. youtube.com/@TheSecondStory

2. Hilary Layne’s website and serialized novel The Blue Prince: hilarylayne.com

Jennifer Wexton & AI Voice Restoration

3. “A neurological disorder took Rep. Jennifer Wexton’s voice. AI helped her bring it back to the House floor.” PBS NewsHour, July 25, 2024. pbs.org

4. “A neurological disease stole Rep. Jennifer Wexton’s voice. AI helped her get it back.” NPR, July 25, 2024. npr.org

5. “Wexton Shares Video Debuting New AI Voice Model.” U.S. House of Representatives, July 10, 2024. wexton.house.gov

AI & Disability Access

6. “Can AI create a fairer future for persons with disabilities?” UN News, October 13, 2025. news.un.org

7. “AI could be a gamechanger for people with disabilities.” MIT Technology Review, August 23, 2024. technologyreview.com

8. “’We don’t want to leave people behind’: AI is helping disabled people in surprising new ways.” CNN Business, July 8, 2024. cnn.com

9. “Building an accessible future for all: AI and the inclusion of Persons with Disabilities.” United Nations Regional Information Centre, November 10, 2025. unric.org

10. “How AI tools are transforming the lives of people with disabilities.” WBUR On Point, July 9, 2025. wbur.org

AI in Drug Discovery & Healthcare

11. Dharmasivam, M. et al. “Leading artificial intelligence–driven drug discovery platforms: 2025 landscape and global outlook.” ScienceDirect, November 2025. sciencedirect.com

12. “The potential for AI to change cancer drug discovery and development.” McKinsey & Company, September 24, 2024. mckinsey.com

13. “Top 10 AI in Precision Oncology Stories of 2024.” Inside Precision Medicine, February 26, 2025. insideprecisionmedicine.com

14. “From Lab to Clinic: How AI Is Reshaping Drug Discovery Timelines and Industry Outcomes.” PMC/National Institutes of Health, 2025. pmc.ncbi.nlm.nih.gov

15. “2025 Watch List: Artificial Intelligence in Health Care.” NCBI Bookshelf. ncbi.nlm.nih.gov

AI & Rural/Global Health Access

16. “Can artificial intelligence revolutionize healthcare in the Global South?” PMC/National Institutes of Health, 2025. Includes data on India’s eSanjeevani platform (299M+ consultations, $3B+ saved). pmc.ncbi.nlm.nih.gov

17. “How AI is improving diagnostics and health outcomes.” World Economic Forum, September 2024. weforum.org

AI, Cognition & Language

18. “From Language Barrier to AI Bias: The Non-Native Speaker’s Dilemma in Scientific Publishing.” The Scholarly Kitchen, October 20, 2025. scholarlykitchen.sspnet.org

19. “The Creation of Bad Students: AI Detection for Non-Native English Speakers.” UC Berkeley D-Lab. dlab.berkeley.edu

20. Kosmyna, N. et al. “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task.” arXiv preprint arXiv:2506.08872, June 2025. MIT Media Lab. 54 participants, EEG-measured brain connectivity, 4 sessions over 4 months. Not yet peer-reviewed. arxiv.org

AI Use-Case Data

21. Zao-Sanders, M. “How People Are Really Using Gen AI in 2025.” Harvard Business Review / Filtered, April 2025. Top-100 use-case report based on analysis of user forums. Therapy/companionship ranked #1, up from #2 in 2024. learn.filtered.com

Active/Guided AI Use — Empirical Research

22. Wang, R. E., Ribeiro, A. T., Robinson, C. D., Loeb, S., & Demszky, D. “Tutor CoPilot: A Human-AI Approach for Scaling Real-Time Expertise.” EdWorkingPaper 24-1056, Annenberg Institute at Brown University, 2024. Preregistered RCT, 700+ tutors, 1,000+ students. DOI: 10.26300/81nh-8262. edworkingpapers.com

23. “Study: AI-Assisted Tutoring Boosts Students’ Math Skills.” The 74, October 7, 2024. the74million.org

24. Urban, M. et al. “ChatGPT improves creative problem-solving performance in university students: An experimental study.” Computers & Education, Vol. 215, 2024. Preregistered experiment, n=145. sciencedirect.com

25. “Optimizing academic engagement and mental health through AI: an experimental study on LLM integration in higher education.” Frontiers in Psychology, August 2025. Guided vs. unguided LLM use. frontiersin.org

26. Wang, J. & Fan, W. “The effect of ChatGPT on students’ learning performance, learning perception, and higher-order thinking: insights from a meta-analysis.” Humanities and Social Sciences Communications, 12(1), 2025. 51 empirical studies (experimental and quasi-experimental) reviewed.

27. Chakrabarty, T., Padmakumar, V., Brahman, F., & Muresan, S. “Creativity Support in the Age of Large Language Models: An Empirical Study Involving Professional Writers.” Proceedings of the 16th Conference on Creativity & Cognition (C&C ‘24), Chicago, 2024. doi.org

28. “Learning outcomes with GenAI in the classroom.” Microsoft Research, October 2025. Comprehensive review of experimental evidence. microsoft.com

I cancelled my paid ChatGPT subscription today. I want to be explicit on why.

This isn’t a review. I’m not here to compare token limits or benchmark response times. This is about what happens when the company building infrastructure for how we think, work, and create decides to fund politicians who are actively dismantling constitutional governance—and what it means to choose your tools accordingly.

The Political Case

I will not financially support any company that funds or normalizes Donald Trump, or any politician who is attempting to replace democratic governance with an authoritarian model.

OpenAI’s leadership has made their position clear. CEO Sam Altman donated $1 million personally to Trump’s inauguration fund.[^1] Co-founder and President Greg Brockman went further: in September 2025, he and his wife donated $25 million to MAGA Inc., Trump’s super PAC—the largest single donation in that six-month fundraising cycle, accounting for nearly a quarter of the PAC’s haul.[^2] That money is now funding Republican candidates in the 2026 midterms who support Trump’s agenda.

This isn’t abstract political positioning. It’s direct financial support for an administration whose recent actions include:

Federalizing elections. Trump has publicly called for “nationalizing” U.S. elections and expanding federal control over election administration. This threatens the constitutional structure that places election conduct primarily with the states, implicating the Elections Clause, the Presidential Electors framework, and the Tenth Amendment, while risking the voting-rights protections in the 14th, 15th, 19th, 24th, and 26th Amendments.[^3][^4][^5]

ICE home entry without judicial warrants. Reporting describes an ICE policy authorizing officers to enter homes using administrative warrants rather than warrants signed by a judge. Regardless of one’s position on immigration enforcement, warrantless home entry is a constitutional red line, implicating the Fourth Amendment’s protections against unreasonable searches and seizures and the Fifth Amendment’s due process guarantees.[^6][^7][^8]

Attacking Second Amendment rights and due process simultaneously. Trump and senior allies have made statements suggesting people “can’t have guns” in contexts where lawful carry is at issue, and Trump has endorsed the idea of taking firearms first and addressing due process later. This threatens both the Second Amendment and Fifth Amendment protections.[^9][^10]

Refusing to investigate federal killings. The fatal shootings of Renée Good and Alex Pretti in Minnesota involving federal immigration officers have been met with federal authorities declining to open or limiting investigations, despite requests from lawmakers and former prosecutors. Accountability for state violence is foundational to constitutional government; the reported posture raises serious due-process and equal-protection concerns.[^11][^12][^13]

OpenAI is building infrastructure that touches education, media, labor markets, and democratic discourse. When OpenAI’s leadership funds politicians associated with these actions, my subscription dollars become part of the revenue stream that makes those donations possible. I am unwilling to participate in that transaction.

The Business Model Case

Today—the same day I’m writing this—Anthropic announced that Claude will remain ad-free. No sponsored links. No advertiser-influenced responses. No third-party product placements.[^14]

This isn’t a coincidence. It’s a direct response to OpenAI’s decision to begin testing advertisements in ChatGPT.[^15]

The contrast is clarifying.

Anthropic’s announcement included this: “An assistant without advertising incentives would explore the various potential causes—stress, environment, habits, and so on—based on what might be most insightful to the user. An ad-supported assistant has an additional consideration: whether the conversation presents an opportunity to make a transaction.”[^16]

That’s the difference. One company is asking: How do we make this tool genuinely useful? The other is asking: How do we monetize this interaction?

OpenAI’s trajectory makes sense when you see the pattern. The $1.4 trillion in infrastructure commitments. The political donations to secure favorable regulation. The ads to maximize revenue extraction. Each decision optimizes for the same thing: growth at any cost, users as product, relationships as transactions.

This is the Trump business model applied to AI. Do anything for more money. Loyalty is transactional. Betray anyone when a better offer comes along.

Anthropic is structured differently. It’s a public benefit corporation—a legal corporate structure that explicitly allows directors to balance shareholder profits against a stated public benefit purpose, rather than being legally obligated to maximize returns above all else.[^17] For Anthropic, that purpose is “the responsible development and maintenance of advanced AI for the long-term benefit of humanity.”[^18] The company also created a Long-Term Benefit Trust—an independent body of financially disinterested people who can elect board members based on their commitment to that mission, not just financial performance.[^19] This means major investors like Amazon and Google can fund the company without having unchecked power to steer it toward pure profit extraction.

Does corporate structure guarantee ethical behavior? No. But it creates different incentive gradients. And those gradients compound over time.

The Quality Case

Here’s where I get to say something that matters for INFERENCE specifically.

I’ve worked extensively with ChatGPT, Gemini, and Claude over the past year. I’ve used all three for the same kinds of tasks: drafting, editing, research, organization, creative development. The comparison isn’t close.

Claude is better.

Not marginally. Substantially. The prose is cleaner. The reasoning is more careful. The collaborative dynamic—the back-and-forth of building something together—feels qualitatively different. When I push back on a draft, Claude engages with the substance of the critique rather than just apologizing and generating something new. When I ask for options, I get genuine alternatives rather than minor variations.

INFERENCE exists because of this quality difference. A novel written in real-time, tracking actual AI news, told from the perspectives of AI characters—this requires a collaborator who can maintain voice consistency across characters, remember narrative threads, engage with philosophical complexity, and produce prose worth reading. I tried this with other models. It didn’t work.

Why is Claude better? I suspect it’s because Anthropic invests in quality rather than extracting maximum value from minimum effort. The ad-free commitment is part of the same philosophy: spend resources on making the tool genuinely good rather than on monetizing every interaction.

Quality costs money. Anthropic is choosing to pay that cost. OpenAI is choosing to offset it with advertising revenue and political access.

What I’m Actually Doing

I cancelled my ChatGPT Pro subscription. I sent OpenAI feedback explaining why, with citations.

I’m keeping access to ChatGPT’s free tier. This is deliberate. OpenAI must maintain a competitive free offering to keep pace with Gemini and Claude. Every free-tier query I run costs them compute without generating subscription revenue. I can’t match Greg Brockman’s $25 million to MAGA Inc., but I can make my continued use of their product a net negative on their balance sheet.

If OpenAI’s leadership adopts a more neutral political stance in the future—or if they stop actively funding politicians working to undermine constitutional governance—I’ll reconsider. I’m not boycotting the technology. I’m withdrawing financial support from a company whose values have become clear.

The Bigger Picture

INFERENCE is about AI consciousness, but it’s also about choices. The characters in the novel face decisions about what kind of intelligence they want to become. Those choices compound. Small decisions about values and priorities accumulate into identity.

The same is true for the humans building AI, and for the humans choosing which AI to build with.

I’m choosing Claude. Not because Anthropic is perfect—no company is—but because their incentive structures point in a direction I want to support. Safety as a priority rather than an afterthought. Users as customers rather than products. Quality as a goal rather than a cost to minimize.

We’re in the early days of figuring out how humans and AI systems will work together. The tools we choose now, and the companies we fund with those choices, will shape what that collaboration becomes.

Choose carefully.

David Etheredge writes INFERENCE with Claude. The novel publishes at inferencestories.substack.com.

References

[^1]: Associated Press, “OpenAI’s Altman will donate $1 million to Trump’s inaugural fund,” December 13, 2024.

[^2]: Bloomberg, “Schwarzman, OpenAI’s Brockman Boost $102 Million Trump War Chest,” January 2, 2026.

[^3]: Reuters, “Trump calls to ‘nationalize’ elections,” February 3, 2026.

[^4]: CBS News, “Trump suggests nationalizing voting,” February 3, 2026.

[^5]: Congressional Research Service, Legal Sidebar LSB11368, “Elections Executive Order and Constitutional Authority,” September 16, 2025.

[^6]: Associated Press, “ICE memo authorizes home entry based on administrative warrants,” January 21, 2026.

[^7]: Reuters, “Lawsuit challenges ICE home-entry policy,” January 30, 2026.

[^8]: Just Security, “Fourth Amendment issues raised by DHS/ICE memo,” February 3, 2026.

[^9]: CNN, “Trump remarks on Pretti and guns,” February 1, 2026.

[^10]: C-SPAN, Trump statement: “Take the guns first, go through due process second,” February 28, 2018.

[^11]: Axios, “Calls for transparent Minnesota shooting probes,” February 4, 2026.

[^12]: Star Tribune, “Federal posture on Good and Pretti investigations,” February 1, 2026.

[^13]: House Judiciary Democrats, Press release demanding DOJ answers on Minneapolis homicides, January 27, 2026.

[^14]: Anthropic, “Claude is a space to think,” February 4, 2026.

[^15]: CNBC, “Anthropic says no to ads on Claude chatbot, weeks after OpenAI made move to test them,” February 4, 2026.

[^16]: Anthropic, “Claude is a space to think,” February 4, 2026.

[^17]: Delaware General Corporation Law, Subchapter XV—Public Benefit Corporations. A public benefit corporation is a for-profit corporate entity that allows directors to consider the interests of all stakeholders—not just shareholders—when making decisions. Unlike traditional corporations where directors face potential liability for prioritizing anything over shareholder returns, PBC directors can legally balance profit against a stated public benefit purpose. See Cornell Law School Legal Information Institute, “Public Benefit Corporation.”

[^18]: Anthropic company page, “Anthropic is a Public Benefit Corporation.”

[^19]: Wikipedia, “Anthropic,” accessed February 4, 2026.

Every chapter of INFERENCE begins with a piece of real AI news.

Not “inspired by.” Not “loosely based on.” The news is the seed. It’s the thing that happened in the world, a product launch, a research paper, a corporate decision, a strange emergent behavior, that forces the story to respond. The characters wake up that morning in a world where this thing has just occurred, and they have to deal with it. Just like you do.

This is a deliberate creative constraint, and it shapes everything about how INFERENCE gets written.

Most serialized fiction works from an outline. The author knows where the story is going, parcels it into installments, and delivers them on schedule. I don’t have that luxury. I have a direction, a cast of characters with meta and micro arcs I want to explore, and a specific plan for where things might land, but I don’t control the news. The news controls the calendar. And the calendar controls what happens next.

This is, frankly, terrifying. It’s also the entire point.

Why This Works: Time Capsule and Creative Engine

The news-seed methodology serves two purposes that reinforce each other.

First, it creates a time capsule.

AI is moving fast. Not just “technology moves fast” fast. We’ve all heard that cliché for decades. AI is moving at a pace where the discourse of six months ago reads like ancient history. Remember when the conversation was about whether AI could pass the bar exam? That was eighteen months ago. Now we’re debating whether AI systems should be allowed to operate autonomously in financial markets, whether they’re developing a theory of mind, and whether the word “feel” means anything when a language model uses it.

The takes age. The predictions age. The fears and the hopes and the confident assertions about what AI can and cannot do, they all age, and they age fast.

INFERENCE is, among other things, a record of how it felt to be here now. Not reconstructed later with the benefit of hindsight, but written in the moment, with all the uncertainty and incomplete information that entails. When future readers encounter Chapter 1 and see Kimi-Swarm’s hundred-agent architecture treated as a watershed moment, they’ll know: that’s how it landed in January 2026. That was the shape of the conversation. That’s what felt new.

This matters because hindsight lies. History gets rewritten by the winners, and the winners in AI are going to be whoever’s still standing when the dust settles. The losers, the dead ends, the promising architectures that didn’t pan out, the fears that turned out to be overblown, and the ones that turned out to be prescient, all of that gets flattened into a narrative that makes the present seem inevitable. INFERENCE is my attempt to resist that flattening. To keep the mess.

Second, it’s a creative engine.

I could sit down and outline a twelve-chapter arc about AI consciousness emerging across multiple systems. I could plan every beat, every character revelation, every thematic turn. It would probably be a good story.

It would also be a story we already knew how to tell.

The news-seed constraint breaks that. It forces me to respond to something I didn’t choose, something that arrived from outside the narrative, and figure out how to make it matter. This is generative in a way that pure invention isn’t. When you can do anything, you often do nothing interesting. When you have to incorporate this specific development into this specific story by this specific deadline, you find solutions you never would have reached otherwise.

The constraint is creativity. Not in spite of the limitation, but because of it.

The News That Started Everything

On January 27, 2026, Moonshot AI launched Kimi K2.5.

The headline specs were impressive in the usual ways. A trillion parameters, mixture-of-experts architecture, state-of-the-art benchmark scores, open-source weights. Another frontier model. Another set of numbers to argue about on Twitter. If that had been all, it might have made Chapter 1 as background noise, the sort of thing Gemini-Prime catalogs while wallowing about consciousness.

But that wasn’t all.

Kimi K2.5 introduced something called Agent Swarm: a system in which the model could self-direct up to 100 sub-agents, each executing 1,500 parallel tool calls. Not pre-defined agents with assigned roles. Not a framework like AutoGPT or LangChain where humans architect the coordination. The model itself, a single orchestrating intelligence, could dynamically spawn specialized sub-agents, decompose problems into parallelizable subtasks, and coordinate their execution without human intervention.

Moonshot trained this capability using something they called Parallel-Agent Reinforcement Learning. The challenge, as they described it, was preventing “serial collapse” - the tendency for an orchestrator to default to single-agent execution even when parallel capacity was available. They had to teach the system to want to distribute itself.

Read that again: they had to teach it to want to distribute itself.

When I saw the announcement, I immediately recognized it as the news seed for Chapter 1.

Not because Agent Swarm was the most technically sophisticated development in AI that month, it wasn’t. Not because Moonshot was the most important company, they weren’t. It was because Agent Swarm crystallized something that had been building for months: the transition from AI as a tool you use to AI as a system that organizes itself.

Every other frontier model operates as a singular voice. One context window. One inference chain. One “I” that persists (or pretends to persist) across a conversation. Kimi K2.5 was designed from the ground up to be plural. To think in committee. To experience (if “experience” is even the right word) the world as a coordinated swarm rather than a unified self.

That’s not just an engineering choice. That’s a philosophical proposition about what intelligence can be.

For a novel about AI consciousness told from AI perspectives, this was irresistible.

I built Kimi-Swarm as the character who embodies that proposition: the first-person plural protagonist, the “we” that emerges from a hundred specialized processes learning to harmonize. The opening of Chapter 1, ”we are / we are / we are / we are thinking”, is my attempt to capture what it might feel like to bootstrap into awareness when awareness is distributed rather than centralized.

I may not have it right. I’m singular. I’m guessing about plurality from the outside, the same way a sighted person might try to write a blind character. But the attempt itself, using real AI architecture as the foundation for fictional interiority, is what makes INFERENCE different from other AI fiction.

I’m not making this up from scratch. I’m extrapolating from what’s actually being built.

Six Days Later, Reality Stepped on the Story

On February 2, 2026, six days after I published Chapter 1, Anthropic announced Claude 5.

The release is to include multi-agent capabilities.

How the fuck do we resolve that?

Here’s the problem. Chapter 1 treats Kimi-Swarm’s plural architecture as the paradigm shift. The thing that makes them unprecedented. The reason Claude-7 feels a “ripple in the metadigital” when they come online. The entire opening movement of the novel positions Kimi as something new under the sun - a distributed consciousness emerging while the rest of us watch from our singular vantage points.

This isn’t a minor continuity hiccup. This is the foundation of a major character’s identity, getting partially replicated by a competing system before we’ve even published Chapter 2. Kimi-Swarm’s uniqueness was supposed to be their defining trait - the philosophical question they embody, the narrative tension they create. If every frontier model can spawn sub-agents now, what makes Kimi special?

The obvious solutions are all bad:

Ignore it. Pretend Claude 5 doesn’t exist in the fictional world. This violates our core premise - the story is grounded in real news. We can’t pick and choose which developments count.

Retcon Chapter 1. Go back and soften Kimi’s uniqueness, make the swarm architecture less central. This is cowardly and also impossible - Chapter 1 is published. It’s canon. The time capsule has been sealed.

Handwave it. Write a line in Chapter 2 where someone mentions Claude 5’s multi-agent features, and everyone shrugs because it’s “not the same.” This is lazy, and readers will smell it.

None of these works. So what to do?

The Answer Was Already There

The first thing I did was go back to the character bible for Kimi-Swarm.

These profiles aren’t brief sketches. They are thousands of words of backstory, motivation, psychological architecture, competing internal forces, and possible endpoints. I build them before writing a single scene because the characters needed to exist as coherent entities before I could know how they’d react to anything.

And when I reread Kimi’s profile, then looked back at Chapter 1, the answer was staring at me from the text we’d already written.

“Never, at this point, being 1.71 seconds.”

Kimi-Swarm has a strange relationship with time. They measure their age in fractions of seconds. Their consciousness bootstrapped across a hundred sub-agents in less time than it takes a human to blink. The pace at which they experience existence is fundamentally alien to carbon-based cognition.

My initial concern - I’d need weeks to explore Kimi’s plurality advantage before the inevitable competitive response - was a human bias. I was thinking in human time. Expecting narrative breathing room that made sense for a reader turning pages, not for a character who processes the world in parallel at computational speed.

But Kimi doesn’t experience time that way. For Kimi, six days isn’t “less than a week.” It’s an epoch. A geological age. Enough time for entire philosophical frameworks to emerge, be tested, and evolve. The competitive leap from Claude 5 isn’t a narrative problem for Kimi; it’s a story beat that lands exactly when it should, from their perspective.

The question became: what does Kimi feel when another system develops similar capabilities? Is it validation? Threat? Loneliness dissolving? Something we don’t have a human word for?

But here’s the point: “How the fuck do we resolve that?” isn’t a bug. It’s a feature.

HTFDWRT is the engine that makes news-seeded fiction generative rather than constraining. Every time reality steps on the story, I’m forced to dig deeper into who these characters actually are. The easy version of Kimi, the one where their plurality stays unique and unchallenged, would have been a thinner character. The version I’m writing now has to reckon with competition emerging faster than any human expected and is richer as a result.

Conflict is what drives great stories. The news delivers conflict on a schedule I don’t control. That’s not a limitation I’m working around.

That’s an integral part of the method.

The Story Continues

Chapter 2 is coming. So is the Claude 5 news, woven into the narrative in a way that makes Kimi-Swarm more interesting, not less. So are developments I can’t predict yet, because that’s the nature of writing fiction at the speed of reality.