The version without the stirring

The Navier-Stokes equations describe how a fluid moves. The Millennium Prize question, stripped down, is whether a smooth flow in three dimensions always stays smooth, or whether some starting flow drives a quantity such as the velocity gradient to infinity in finite time. That second outcome is called blowup.

In September 2026, OpenAI announced a blowup result for equations with a smooth external force: in effect, a carefully tuned force stirring the fluid in exactly the right way. Scientific American reported that the result technically satisfied the Clay Mathematics Institute's prize conditions, while many experts objected to the finely tuned forcing as a loophole in the problem's wording. The unforced case, where the fluid is left to its own dynamics, is the one this new claim addresses.

That is the case Buckmaster and Alpöge now claim. Their paper, posted on November 20, constructs smooth initial data, both in open space and on the torus, and argues that the resulting solutions blow up in finite time.

Why it was posted unfinished

Buckmaster said so himself. Announcing the paper on Mastodon, he wrote that the write-up is unpolished, and Scientific American reports the reason: after last time, the pair "couldn't risk waiting," over concerns about being scooped. Last time is the September episode: Buckmaster had publicly suggested that his and Alpöge's progress had reached OpenAI before its announcement, and raised questions about how. OpenAI denied allegations of foul play. That history is the stated reason for the rush: a 79-page human-readable paper went out the moment it could, alongside a Lean formalization of about 600,000 lines, rather than after the slower work of polishing either.

The same-day context matters too. The same day, Alpöge also announced a claimed proof of the Riemann hypothesis, likewise AI-assisted (covered by ALFW on November 20). November 20 produced two Millennium-scale claims from overlapping authors. That is a reason to read each one separately, on its own artifacts, rather than letting either borrow credibility from the other.

Three gaps, all stated in the repository

What separates this from a finished proof is not hidden. It is written in the project's own notes.

First, the certificate. The claim depends on a numerical certificate: the paper gives its hash, and the repository states plainly that the formalization contains no data, that the certificate is not in the repository, and that it will be released. A hash is a commitment to a specific file. It lets nobody check anything until the file appears and matches it.

Second, the check on the certificate. Whether the certificate satisfies the conditions in the paper's Appendix B is verified, the repository says, by a rigorous interval-arithmetic computation at 128 bits of precision, using the Arb library through FLINT. That computation is not part of the formalization and is not in the repository. So even after the certificate appears, reproducing its verification means rebuilding a separate computation that has not been published either.

Third, the build state. A patch was applied on November 20 to the cell-layer statements and to the library proofs that depend on them: a bound's weight changed, one constant was allowed to be zero, and a set of harmonics was restricted. The repository states that the patched package was not rebuilt when the repository was assembled.

None of this makes the claim false. Computer-assisted proofs are a legitimate and established route in fluid equations, and the authors have disclosed each gap themselves, which is more than many claimants do. It does fix the claim's current status precisely: a detailed argument whose central numerical object, and the computation that validates it, are not yet available for anyone else to run.

What would settle it

Three events, in order. The certificate is released and matches the published hash. The interval computation is published or independently reproduced, so the certificate's properties do not rest on the authors' own run. And the patched Lean package is rebuilt, so the formalization on offer corresponds to the statements in the paper as patched. Expert reading of the 79 pages runs in parallel with all of that, and it is the reading, not the line count, that decides whether the analytical argument has holes the numerics cannot fill.

Until the first of those happens, the accurate headline is the narrower one. A serious team, working in the open and under competitive pressure it has described on the record, has posted a claim to the unforced version of a famous problem, with the central file promised rather than attached. The promise may well be kept, and quickly. It has not been kept yet.

Sources

  1. [1] Alpöge and Buckmaster, "Finite-time blow-up for the Navier-Stokes equations" — paper and Lean repository, including the repository README, GitHub, posted November 20, 2026Read source
  2. [2] Scientific American, "AI just solved the 'real' Navier-Stokes problem", November 20, 2026Read source