OpenAI Says 10,000 Agents Cracked Navier-Stokes Blowup in 88 Hours; NYU Mathematician Cries Foul

A million-dollar Millennium Prize problem wasn't cracked by a lone genius in an attic this time. It was swarmed by ten thousand AI agents over a week, and it came with a credit war attached.

Three key facts

On September 8, OpenAI said an unreleased model running roughly 10,000 agents found a Navier-Stokes blowup proof in about 88 hours. The claim: the 3D incompressible Navier-Stokes equations can develop a finite-time singularity from smooth initial data under a smooth external force. Per TechCrunch, the project burned about 300 billion output tokens in a week, roughly $22.5 million of compute at current prices.

The Clay Mathematics Institute hasn't signed off. It still lists the problem as unsolved, the proof hasn't been independently verified, and many mathematicians note it leans on an external force term, while the version they care about has none.

NYU's Tristan Buckmaster says OpenAI jumped the queue. He and Anthropic researcher Levent Alpöge had spent about a year on the problem with Claude, proving blowup for the simpler Euler equations on August 15. OpenAI only launched its agent push on September 1, after hearing rumors the problem was about to fall. OpenAI denies seeing their work but concedes de-identified data from their use of its products may have helped improve its models.

WangDou's Take

Park the "AI solves Millennium Problem" headline and read the ledger: $22.5 million and 300 billion tokens for a proof the Clay Institute won't stamp and that relies on a forcing term. The prize is $1 million. OpenAI spent more than twenty times that chasing a title, which makes this a marketing budget, not a research grant. The playbook is what should worry people: academics grind for a year, get close, a rumor leaks, and ten thousand agents flood their direction in 88 hours. Who will ever discuss unfinished work with peers again? And that line about de-identified data is the real tell: if you do research with our product, your ideas may fuel our model, which we then use to beat you to publication. Math used to reward whoever proved it first. Now it rewards whoever burns through GPUs fastest.

Source: TechCrunch · Scientific American · Interesting Engineering