The agents took 88 hours and the rumor did the rest

The loudest story of the week is a fluid dynamics proof that almost nobody is discussing as fluid dynamics. OpenAI says a group of agents on an unreleased model resolved the Navier-Stokes problem. Simon Willison’s September 8 post quotes their account: the effort began September 1 after a rumor that two Millennium problems had been resolved, the agents reached a resolution about 88 hours after launch, Lean verification took another 17 hours, and the run sent 2.7 million messages and used around 130 billion output tokens. Willison prices the wider campaign’s 300 billion tokens at public Astra rates and gets 15 million dollars. Ethan Knight, quoted in the AI News issue the next day, describes roughly 10,000 agents trained for a year with multi-agent RL to decide for themselves how to organize.

The rumor is the part I keep rereading. Per Willison’s summary of Tristan Buckmaster’s complaint, Buckmaster and Levent Alpöge had worked the problem for the better part of a year with Claude and Codex, had a breakthrough on August 15, and word of it reached OpenAI before OpenAI sent its first prompt. Sebastien Bubeck has since posted receipts of the exchange, per Matthew Berman, and the authorship dispute belongs to the people in it.

Concede what deserves conceding. A Lean checked proof is a proof regardless of who was told what, and searching a space that size with 10,000 agents is a capability whether or not it got there first.

Still, the vault’s orchestration notes are blunt about crowds. Equal-token studies find that apparent multi-agent gains are often extra test-time compute, independent samples, or better evals rather than the team itself, and the pattern pays only when the decomposition is unknown in advance and the value covers the bill. What a rumor supplies is the one thing tokens cannot buy: the knowledge that an answer exists. Willison notes security already works this way, where word of an unpatched bug is enough to set agents hunting.

So the interesting number is not the 88 hours. It is how much of the search space vanished the moment someone heard it was worth searching.