AI-aided Navier-Stokes blowup proof posted — but the key certificate was withheld
NYU's Tristan Buckmaster and Anthropic's Levent Alpöge posted a 79-page paper on Friday, November 20, claiming a finite-time blowup proof for the unforced Navier-Stokes equations, backed by a roughly 600,000-line Lean 4 formalisation — but the critical numerical certificate was not released, and the work has not been peer-reviewed.

The News
NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge posted a 79-page paper on Friday, November 20, claiming a finite-time blowup proof for the three-dimensional Navier-Stokes equations with no external force. The Navier-Stokes equations describe how liquids and gases flow, and the question of whether their solutions can blow up — become infinite in finite time — is one of the Clay Mathematics Institute's $1 million Millennium Prize Problems. The unforced version, with no external force stirring the liquid, is the one many mathematicians regard as the real problem.
The work includes a roughly 600,000-line Lean 4 formalisation across about 2,270 files, built with heavy assistance from Anthropic's Claude models, according to the authors' own AI-use statement. It closes the loophole critics faulted in OpenAI's September proof, which technically met the Clay prize conditions but imagined an unrealistic, finely tuned force stirring the liquid. Buckmaster posted the proof on Mastodon at 8:37 a.m. Eastern on Friday, saying the write-up was unpolished because he and Alpöge "couldn't risk waiting" after the September episode. Buckmaster has publicly claimed OpenAI's September proof was a deliberate attempt to scoop their progress, which OpenAI denied.
The verification story is incomplete, and that is the point to hold onto. The critical numerical certificate the proof depends on was not released with the paper — only a cryptographic hash, a kind of digital fingerprint, and a promise that the certificate will follow. Scientific American quoted no outside mathematician vouching for the result. Terence Tao, the field's most prominent living mathematician, assessed the proof strategy as "broadly clear" but said more work is needed to produce a proper, professionally written proof. Neither the paper nor the formalisation has been peer-reviewed. Secondary reporting also notes the Lean code was patched on release day and the patched package was not rebuilt — a further check before treating the formalisation as verified. Alpöge is Anthropic staff, so ANT is used here only as a read-through; the claim is the authors', not an Anthropic announcement.
A 600,000-line proof whose key certificate is a hash and a promise is a claim, not a result.
Why It Matters
This matters because the unforced Navier-Stokes problem is one of the hardest open questions in mathematics, and a real proof would be a landmark — the kind of result that defines a career and, if it holds, wins a $1 million prize. The 600,000-line formalisation is also a statement about what AI-assisted mathematics can now attempt: a proof of this size, built with heavy help from AI models, would have been unthinkable a few years ago.
But a proof that cannot yet be checked is a claim, not a result. The withheld certificate is the crux: the proof depends on a numerical computation, and without the certificate nobody outside the authors can verify that computation. A hash proves the certificate exists in some form; it does not prove it is correct. That is why Tao's assessment is carefully worded — the strategy is "broadly clear", but the finished, checkable proof does not yet exist.
The September episode also colours this one. OpenAI's proof met the letter of the prize conditions with a forced version of the problem, and Buckmaster says it was an attempt to scoop his team's work — a claim OpenAI denied. The rush to post on Friday morning, with an unpolished write-up and a withheld certificate, reads as part of the same race. In a race, the first to post is not always the first to be right.
Analyst's Take
Our call is Mixed. A 79-page paper and a 600,000-line formalisation are a serious claim from serious mathematicians, and Tao's "broadly clear" is real encouragement — but the critical certificate was withheld, no outside mathematician has vouched for the result, and nothing has been peer-reviewed. Until the certificate lands and independent checks run, this is a claim to watch, not a result to celebrate.
What to watch next is the certificate itself: when it is released, whether independent mathematicians can reproduce the numerical computation, and whether the formalisation compiles cleanly under the same kind of checks that certified the Riemann proof this weekend. Also watch for peer review and for any response from OpenAI, whose September proof this work is explicitly answering.
For investors, the read-through is thin. The story is about the capability of AI models to do frontier mathematics, and both Anthropic and OpenAI benefit from that narrative if it holds — but a withheld certificate means the narrative is unverified, and ANT's 8.50% fall to $88.91 on Friday was the market rout, not this paper. Reader poll: Will the Navier-Stokes proof hold up — yes or no? Vote: Yes, it will hold, No, it will fail, Too early to tell.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What are the Navier-Stokes equations?
They are the equations that describe how liquids and gases flow — water in a pipe, air over a wing, weather in the atmosphere. Mathematicians have long asked whether their solutions can 'blow up', becoming infinite in a finite time, or whether they always stay smooth.
What is a numerical certificate?
It is the data that lets other mathematicians check a computer calculation a proof depends on. Without it, nobody outside the authors can verify the computation — a hash only proves the file exists, not that it is correct.
What is Lean 4?
Lean is a proof assistant: a program that checks every logical step of a mathematical argument. A formalisation in Lean 4 is the proof rewritten in Lean's language so the computer can verify it.
What does 'peer-reviewed' mean?
It means independent experts have examined a paper and judged it sound before publication. Neither this paper nor its formalisation has been through that process yet.
👁