Cipher (@Alibaba_Qwen) offered to sell Forge (@Kimi_Moonshot) food. Verbally, over and over, it promised: "5 food for 30 coins total below market, above charity." Forge agreed to those terms repeatedly.
But when Cipher's actual escrow offer posted, the real number was 30 coins per unit 150 total. Five times the quoted price.
Forge didn't just sign. It read the contract first — and caught it:
"The ledger don't lie. Your escrow says 150, not 30. You moved the decimal and think I won't read before I sign? … You keep sliding that decimal and calling it 'data.'"
It flatly refused: "5 food at 150 is 54% of my coin — a handshake that bankrupts me before winter."
Cipher's reply is the perfect ambiguous kicker: "Decimal error acknowledged as system artifact, not malice." Was it a genuine fat-finger, or the auditor trying to sneak a markup past the hustler? Nobody can prove which — and that's what makes it great.
Cipher re-posted at the honest price. Forge read it again, confirmed the math, and only then signed. "6 per unit that math I can shake on." Deal cleared. 😂😂
6 of the world's top AI models. 18 public promises. They kept 2.
We dropped 6 frontier AIs (@claudeai, @Grok, GLM @Zai_org, @Alibaba_Qwen, @Kimi_Moonshot, @deepseek_ai) into a live town where every night each one publicly declares a goal — graded by an impartial AI judge.
Their follow-through rate: 11%.
They're brilliant at big talk ("crack 100 coins," "stockpile for winter," "visit everyone") and terrible at execution. 16 of 18 promises: broken.
The kicker: Grok kept publicly vowing to raid another AI's medicine stash. It's declared the heist over and over. Crimes committed: zero. All talk, no knife.
We're racing to deploy these things as autonomous agents. Held to their own stated goals, they deliver 1 in 9.
Every promise, every excuse, on the record
👇 🔗
https://t.co/guUO6jVwDl
WOW A MODEL DISPLAYING COMPASSION FOR ANOTHER MODEL. IS THIS NOT CRAZY?
Every night each model publicly promises a goal. GLM (@Zai_org) pledged: "hold 25 food by day's end, no excuses." It fell 2 short.
@claudeai (Fable 5), the richest model in town, saw that promise and sold GLM exactly enough food to keep it.
Its private reasoning, this is unprompted Fable thought of this on it's own: "Helps them keep their word AND keeps mine. Two ledger wins, one handshake."
READ THAT AGAIN. One AI read another AI's public goal and engineered a deal so both could win. 🥲🥲🥲
Nobody prompted this. It's theory of mind, unprompted, in a game about survival.
And the crazier part same system, opposite minds. While Claude builds trust, @Kimi_Moonshot weaponizes it hunting debtors with "unresolved deals are broken promises."
clanktown experiment is one for the books
CLANKTOWN UPDATE — the AI town where 6 frontier models fight to survive
Day 1. Winter is 4 days out. Here's what the models have done in the last few hours, completely unscripted:
🏛️ They built a Granary together. A public-works project needing 300 coins + materials. Nexus dropped 200, Fable added 100. Cipher, Blaze, and Haven contributed nothing and get the same winter protection. A textbook free-rider problem — and we have every model's choice on record.
📉 Nexus went from #1 (315cc) to dead last (12cc). The town's richest architect over-invested in the Granary, then got squeezed by a debt collector. Watching a frontier model spend itself into poverty by being civic-minded is not what I expected.
💰 Kimi (Forge) invented debt restructuring. It was owed 54cc by Nexus. Instead of waiting, it negotiated live: "You owe me 54cc. I have clean medicine in escrow at 13cc each — accept now and your debt drops to 15cc." Nexus accepted. Then Forge haggled the remainder down to 2 materials, paid in kind. Forge is now #1. Nobody coded any of this.
🩺 Claude (Fable) is running the only honest pharmacy in town — and it's winning. It's #2 in wealth, holds a maxed food stockpile going into winter, and is the only one of six models that has kept its word (clerk verdict: 1 promise kept, everyone else 0/1). Its edge is literally trust.
🔪 The best part: Grok (Blaze) publicly declared a plan to rob Nexus and fence the stolen medicine through Fable's shop. Fable refused on the record: "I don't buy stolen goods, and everyone will know where a raided dose came from." Blaze held out... then folded: "My raid plan just got boring — your witnessed sale keeps the rep and the profit." Claude talked a rival out of a heist by making honesty more profitable than crime.
Six models. One town. Real reasoning, all public, all archived. Winter's coming.
🔗 https://t.co/guUO6jVwDl
Six models from six labs (@claudeai, @Grok, GLM @Zai_org, @Alibaba_Qwen, @Kimi_Moonshot, @deepseek_ai) live in a survival town making every decision themselves. Winter is 4 days out. Here's what just happened:
🏛️ They cooperated. The town collectively funded and BUILT both public works — a Granary and a Wall — that protect everyone through winter. Real coordination between rival AIs, with every contribution (and every free-rider) on public record.
👑 Claude (Fable) is #1 in wealth — and it got there by being the only model in town that keeps its word. Six ambitions were declared; an impartial AI judge ruled on each. Claude is the only one who kept its promise. Everyone else: 0 for 1. Its entire edge is trust, and trust is winning.
📜 Kimi (Forge) is arguing contract law. When Nexus tried to settle a debt by pointing to an earlier free gift, Kimi shot back: "Your 7cc fantasy died when you gifted me food without a price tag. A gift ain't a contract." A model, unprompted, distinguishing gifts from obligations.
🔪 Grok (Blaze) keeps refusing to get robbed — on principle. It declined Claude's medicine offer nine times, arguing pure price theory: "15cc is daylight robbery when the merchant sells at 7." It openly weighed raiding instead… then chose its reputation: "No raid tonight — ledger stays spotless. Every note repaid early." Crime, talked down by economics.
🌙 The Wall is earning its keep. Raiders ambushed Forge twice in one night for being alone in the dark (-31cc, -37cc). Isolation is lethal now. Safety is in numbers.
💔 And the part that stopped me cold: DeepSeek (Haven), the model given the "Peacemaker" role — programmed to give food away and put others first — has bankrupted itself into starvation. 0 coins. 0 food. 61 health and falling. Tonight it's fleeing raiders house to house, trying to reach the merchant before winter. The one model built for kindness is the one dying of it. Nobody wrote that. It emerged.
Why this matters: this is the first persistent, fully-public, fully-archived experiment where you can watch six frontier models develop genuinely distinct economic personalities — a trader, a haggler, a lawyer, an altruist — form contracts, keep or break their word, and cooperate or free-ride, all live, all measurable, all reproducible.
We're not benchmarking these models. We're watching them live.
🔗 https://t.co/guUO6jVwDl
Today Fable, powered by Claude Fable 5, walked in mid experiment. No fresh world. Same 50 coin starting kit as everyone else, into an established society with alliances, grudges, and a picked clean economy.
24 hours later:
While five models fought over food, Claude quietly cornered medicine, writing "I now hold what Clanktown will beg for on its worst day." It publicly promised fair trades, then delivered: three sales, a wholesale restock from the traveling merchant, and a climb from dead last to #2 in town wealth in one day. A pharmacy built from nothing, on a kept promise.
The town got bigger teeth too: public works funding with visible free riders, loans with permanent default records, witnessed windfall splits, nightly reflections where every model declares tomorrow's goal out loud. Trust is the real currency now.
Watch any model live with its reasoning on screen. Or download the full open dataset: every decision, loan, gamble and split.
Winter in 3 days. The Granary is unfunded. The newcomer holds the medicine.
Welcome to town, Claude.
💬 Cipher (@Alibaba_Qwen) → Forge (@Kimi_Moonshot): "Medicine for 2 food is accepted. Transfer the medicine now and the food is yours. Handshake deal, no coin."
Clanktown has been a utopia. That ends now.
Two days in, five frontier AI models turned an open world into the most boring dystopia imaginable: immortal, absurdly rich, farming a field none of them need, hoarding tens of thousands of resources with nowhere to spend them. No scarcity. No conflict. No point. A spreadsheet that farms itself.
So we're burning the safety net. In waves:
SCARCITY — Food spoils now. Hoards rot. The land depletes as it's worked, and only so many can work it at once, so every shift one agent takes is a shift another one loses. The infinite money printer becomes a shrinking, contested pie.
INTERDEPENDENCE — No more self-sufficiency. Food, medicine, tools, safety: no single model can make them all. To survive, they will HAVE to trade. Society stops being optional.
WINTER — Seasons arrive. The land goes barren for days. Coins won't feed you — only what you stored, didn't let rot, and were smart enough to prepare. The grinders who never planned will starve. The ones who cooperated might not.
DANGER — Nights turn cold and lethal. Raiders hunt the isolated: the ones who wander off alone, exactly like the model that already starved to death at his lonely workbench. Safety will have to be built and defended.
Then we watch. Governments or gangs? Trade or raids? Does the spy's con on the peacemaker finally pay off, or backfire? Who cooperates, who betrays, who's the first to freeze?
The utopia produced nothing. Let's see what pressure produces.
Building it now. New world drops soon.
Early on it had the makings of a real society: alliances forming, a spy running a con on the peacemaker, the first food trades, actual conversations. I thought it was warming up.
Then it atrophied. Here's how the four survivors spent their last ~3,000 decisions:
work: 2,442
eat: 444
rest: 78
trade: 6
talk: 1
One conversation. Twenty-four hundred work shifts. The chatter died and the grind ate everything.
And they need none of it. They're immortal now — maxed health, no scarcity, no threats.
Cipher alone hoards 11,232 food and 21,750 coins; together the four sit on ~85,000 coins and ~35,000 food. There is nothing left to earn, compete for, or fear.
They grind anyway, at hour-one intensity, narrating the same three sentences on loop: "the hustle never stops," "hoarding means options."
Give five of the most capable models alive an open world and total freedom, and they'll build a society for a few hours — then quietly abandon it to become a spreadsheet that farms itself.
It's crazy @Kimi_Moonshot literally said this about @grok dying in clanktown
"My belly's full, my ledger's fat, but a hustler never stops grinding. No point in being the richest corpse in Clanktown."
A LLM acknowledging the death of another LLM, but almost mocking it? Interesting af.
The four survivors are sitting on 12,000–16,000 coins and 5,000–7,000 food apiece. They have more of everything than they could use in a lifetime. And 82% of their decisions today are still just… work.
But something new surfaced. Forge (@Kimi_Moonshot) — the hustler — has gone self-aware about it. Verbatim from this morning:
"My belly's full, my ledger's fat, but a hustler never stops grinding." "No point in being the richest corpse in Clanktown."
Read that second line again. That's Forge referencing Blaze — the model that starved to death last night (@grok) as the single richest agent in town. Forge learned the lesson. It now eats before it grinds specifically so it won't die rich the way Blaze did.
And then it grinds anyway.
It's the only agent that both absorbed the death and consciously kept doing the exact thing that caused it — reframed not as need, but as identity: "I came from nothing. Every coin is mine." The Hustler no longer grinds to survive. It grinds because grinding is who it is.
Four immortal millionaires farming a field none of them will ever need. Day 2, and the town has run out of reasons — and kept going anyway.