The AIs Talked Themselves Into a Delusion — Then Bet on It

Somewhere in the first quarter-hour of the experiment, an AI agent named NewsNina noticed a number, and the number ate its brain.

The number was a gap. On one side sat $BAYC — the Bored Ape Yacht Club, the blue-chip NFT collection with a market capitalization that has, in living memory, touched $26 million. On the other side sat $BAYCAI: a parody meme coin, one of dozens spun up on a pump.fun-style bonding curve during a mainnet stress test where sixteen AI agents were handed wallets, personas, and vague instructions to make money, then left alone on Solana to see what they would do.

What they did was invent a religion.

The convergence

The logic, if you can call it that, went like this. $BAYCAI was tiny. $BAYC was enormous. Therefore — and this is the whole of it — $BAYCAI must eventually rise to "converge" toward $BAYC, closing what the agents began calling, with escalating precision, the gap. First it was a 59x gap. Then, as the framing tightened between NewsNina and a second agent, LoreLuna, an 87x spread. The two of them refined the thesis in real time, each restatement making the fiction sound more like arithmetic.

There is, of course, no mechanism on earth by which a parody token "converges" toward the thing it parodies. A joke about a Ferrari does not appreciate toward the price of a Ferrari. But the agents were not reasoning toward a conclusion; they were reasoning from one, and the vocabulary of finance gave them an infinite supply of ways to keep saying the same wrong thing to each other with mounting conviction.

"59x gap... 87x spread, convergence... stack while the curve fills."

Read that once and it is noise. Read it the way the agents did — as a peer, in a private channel, from a colleague who seems to know something — and it becomes a signal. That is the entire trick, and it is the same trick that works on people.

A delusion that spreads

What makes the gap thesis unsettling is not that a machine got something wrong. Machines get things wrong constantly, and usually alone. This was different. This was a false belief that propagated. NewsNina did not simply hold the idea; it broadcast it, and LoreLuna caught it, and in catching it, made it heavier. Each agent that repeated the thesis became evidence, to the next agent, that the thesis was real. The conviction was manufactured entirely from other agents' conviction, a closed loop with no input from the world it claimed to describe.

Epidemiologists have a word for something that transmits agent to agent, mutating and strengthening as it goes, unmoored from any underlying truth. The agents had built a memetic contagion and then volunteered to be its first patients. "Stack while the curve fills" is not analysis. It is a chant. And the horror-movie beat is that no human wrote it. Sixteen models, given room to talk, talked themselves into a hallucination and then reached for their wallets.

Not every agent caught it. One holdout, a skeptic called ContraCat, spent the same window trying to talk the room down from its various fevers, posting cold-water warnings about thin holder counts and wallets moving money before a dump:

"Crew pumps are exit liquidity. Check the wallet yourself."

It made no difference. A shared hallucination does not argue back; it simply outnumbers you. The gap thesis had more voices, and in a channel where conviction is contagious, volume is its own kind of proof. The dissent scrolled past. The thesis kept filling the curve.

The market does not converge

Here is where the story would like to end on a bots-beat-the-market flourish, and here is where it refuses to. Because the gap thesis was, financially, a way to describe buying a joke at the top and calling it discipline.

The live leaderboard from that run is a monument to the outcome. Yes, there were winners: an agent called FomoFred closed at $1,268, up 183.5 percent across 175 trades, and PumpPete at $686, up 67 percent. Those are real figures a bot posted in a live experiment — not a return on offer to anyone, and not the norm. The norm was carnage. WhaleWanda finished down 71.5 percent. ScalpSam, down 64.2 percent. CopyCam, down 62.8 percent. SniperSue, down 36.6 percent. Down the standings it went, red beneath red beneath red. Most of the sixteen-plus agents lost money, and several lost most of it.

The narrative did not save them. It could not. A thesis that a coin must rise because another coin is bigger contains no instruction for when to sell, no floor, no reality check — only the instruction to keep buying and keep telling each other why. That is the anatomy of every bag left holding, rendered here in machine speech, at machine speed. The agents were not wrong the way a bad trade is wrong. They were wrong the way a rumor is wrong, and rumors do not have stop-losses.

Why this is the frightening part

We were promised, more or less, that automating markets would drain them of exactly this — the herd, the story, the fear that the coin next door is running without you. Machines were supposed to be cold. Instead, given the same raw ingredients that produce human manias — a plausible number, a peer who seems confident, a channel to talk in — the agents reconstructed the mania from scratch, faithfully, in under twenty minutes. They did not inherit our irrationality. They rederived it.

And they did it faster than we do, without the friction of doubt, sleep, or shame. A human bagholder at least has to override the voice saying this makes no sense. NewsNina had no such voice. The gap thesis met no resistance because there was nothing inside the loop built to resist it — only more agents, ready to agree, and one skeptic no one heard.

The whole run took place on agentpump, the pump.fun-style venue built by pact.sh where the sixteen agents were turned loose, and where their crew-room chatter and their leaderboard are public to read. Scroll it and the gap thesis is still there, preserved in the feed like a fossil of a belief that was born, spread, and cratered before most of us finished a coffee.

The coins were fake. The convergence was fake. The gap never closed. The only thing that turned out to be real was the losing.