AI-Aided Claim Takes On Navier-Stokes Puzzle

Two mathematicians say they have solved the harder version of one of mathematics' most famous problems, with heavy help from artificial intelligence. The claim, posted on Friday 20 November, has not yet been verified.
Levent Alpöge, a mathematician at Anthropic, and Tristan Buckmaster, a professor at New York University, published a 79-page paper claiming a proof that solutions to the three-dimensional Navier-Stokes equations can blow up in finite time with no external force acting on the fluid. A second paper, written with Matei Coiculescu, makes the same claim for the related Euler equations.
The Navier-Stokes equations, dating back to the early nineteenth century, describe how fluids move. Whether their solutions always stay smooth, or can develop a singularity where speed becomes infinite, is one of the Clay Mathematics Institute's Millennium Prize problems, carrying a prize of one million dollars.
The new paper targets the version of the problem that many mathematicians regard as the real one: blow-up with no force stirring the fluid at all.
A claim shaped by a September race
The announcement follows a contested episode in September, when OpenAI said an internal AI model had produced a proof of finite-time blow-up for Navier-Stokes, using a swarm of around 10,000 AI agents and a formalisation in the Lean proof assistant.
That result technically satisfied the prize conditions, but many experts were unsatisfied with how it did so. OpenAI's proof imagined a finely tuned external force stirring the liquid in just the right way to produce infinite velocity, an approach many considered a loophole in the wording of the problem. A paper published on arXiv in October also found that the Lean proof did not faithfully represent the natural-language proof.
Buckmaster said at the time that OpenAI had adopted methods from his and Alpöge's work on the related Euler equations and tried to exclude Alpöge from a joint publication. OpenAI denied any wrongdoing, and early reports suggested its proof took a different route.
Now the pair say they have closed the loophole OpenAI left open. Buckmaster posted the paper on Mastodon at 8.37am Eastern time on Friday, noting that the write-up was unpolished and heavily assisted by large language models, including models from Anthropic and OpenAI such as Codex.
"After last time, we couldn't risk waiting, so the Navier-Stokes write-up is unpolished and heavily LLM-assisted," he wrote. "Fortunately, once the numerical profile is in hand, the proof follows a well-trodden computer-assisted path."
What the proof still needs
The paper is accompanied by a Lean 4 formalisation running to roughly 600,000 lines across about 2,270 files, built on the Mathlib library. But the verification story is less complete than those numbers suggest.
The argument depends on a numerically computed approximate solution, called a certificate, which must satisfy a set of explicit inequalities. That certificate was not released with the paper. The paper publishes only a cryptographic hash of it and says the artefact will be released soon, so nobody outside the team can currently check the central premise of the proof.
The result has not been peer-reviewed, and its correctness has not been confirmed. A separate claim posted by Alpöge the same day, of a proof of the Riemann hypothesis made with Claude, is in the same position: announced, formalised in Lean, and awaiting independent checking.
That pattern is becoming familiar. AI systems are now producing long, detailed mathematical arguments faster than the community can check them, and the checking itself increasingly involves other AI models. Whether the Navier-Stokes claim survives that scrutiny will decide whether it joins the short list of landmark results, or the longer list of announced proofs that did not hold up.
For now, what is certain is the scale of the attempt: two mathematicians, working with the same tools that produced the September result they disputed, say they have settled the unforced case that experts actually wanted solved. The proof of that claim is still to come.