AINEWS
The result targets the Navier Stokes equations, used to model everything from weather to blood flow, and could be only the second of the seven famous Clay Millennium Prize problems ever solved — and the first by an AI lab.
OpenAI says its internal model produced a certified proof that the Navier-Stokes equations — the math used to model how fluids move, from weather systems to blood flow — can in rare cases predict infinite fluid speed, a physically impossible event mathematicians call a "blow-up." If the result holds, it would be only the second solution in history to one of the seven Clay Millennium Prize Problems, each carrying a $1 million reward, and the first by an AI lab.
The announcement landed Tuesday, and so did the dispute. Two human mathematicians who had been working the same angle — Tristan Buckmaster at New York University and Levent Alpöge, a researcher who is also an Anthropic employee — publicly alleged that OpenAI only pursued the full problem after learning, within roughly the prior week, that they were using a technique called "forcing" to break the related Euler equations. Buckmaster's public statement the night before the announcement framed the timing as evidence that OpenAI was aware of the parallel work.
Córdoba and Martínez-Zoroa developed the forcing approach. Diego Córdoba, asked about the OpenAI result by Scientific American, said the team was "a little bit in shock" and that a confirmed resolution would be "a big surprise."
OpenAI's Sébastien Bubeck denied the allegation in a press briefing and on X, saying the Euler-stage solution was reached by a totally different method. He acknowledged that the full Navier-Stokes step followed a similar method to Buckmaster and Alpöge's, and said the proof was assembled over the weekend before the announcement. OpenAI did not publish the full proof on its blog. The model-disproof page currently hosts a separate discrete-geometry result and its supporting PDFs, alongside an independent arXiv preprint on the same unit-distance question.
OpenAI says the result was checked in Lean, a programming language that turns a mathematical proof into machine-verified code. If the certification holds, it would rule out the most common failure mode for AI-generated math: a result that looks right but is wrong. Lean is the same formal-verification system used in high-profile proofs such as the liquid tensor experiment, and it is one of the few public signals that an AI-produced result can be checked without a human re-reading every line.
Within a single news cycle, an AI lab and two named human researchers produced results using the same method for the same long-standing problem. Buckmaster alleges OpenAI was aware of the parallel work; Bubeck denies it. The field has no settled norm for who counts as the author of a result in that situation, or for what "priority" means when a model can replicate a research approach in a weekend.
The Millennium Prize framing — a $1 million reward on a problem mathematicians have chased for two decades — gives the result its public stakes. A Navier-Stokes blow-up proof would tell physicists and engineers where the standard equations break down, and where models of turbulence, ocean currents, and cardiovascular flow need to be patched.
Whether the Lean certificate is published and survives independent re-checking is the next concrete test. If it does, the proof stands regardless of who claims credit. If it doesn't, the priority argument may turn out to have been fought over a result that never quite landed.