Timeline

AI-assisted proofs of blow-up for the Euler and Navier–Stokes equations

The Navier–Stokes paper claims the unforced Millennium problem itself, with its numerical certificate unreleased; the Euler paper, a second unforced proof after OpenAI's, has a Lean formalisation bar one cited criterion.

  • Ideas & essays
  • Major

On 20 November NYU mathematician Tristan Buckmaster and Levent Alpöge, who works at Anthropic, posted two papers claiming finite-time blow-up — smooth initial data whose solution develops a singularity — for the unforced three-dimensional equations of fluid flow. One covers the Euler equations, written with Matei Coiculescu, a Simons Junior Fellow at NYU; the other covers Navier–Stokes. Terence Tao listed both that day beside the Riemann hypothesis claim, ranking the three by how “digested” each is.

The Euler paper is not the first unforced Euler blow-up claim. OpenAI had reported one on 8 September, on R³ from compactly supported data, alongside its Navier–Stokes result. The new paper constructs smooth data on the periodic torus whose velocity stays bounded while vorticity at the origin tends to infinity, using Craik–Criminale waves as building blocks and extending the forced-blow-up programme of Diego Córdoba and Luis Martínez-Zoroa. Its provenance section says the authors’ own unreleased forced proof used the same wave mechanism as OpenAI’s paper. What Tao highlights is that it is human-readable and formalised: the repository says the Lean formalisation corresponds one-to-one with the proof and depends only on three standard axioms, with sorry absent. The exception is Corollary 8.3, which cites the periodic Beale–Kato–Majda criterion, not formalised. The authors’ AI statement says they used Claude and Codex throughout, and that the first unforced proof was written by Anthropic models, then rewritten under their direction.

The Navier–Stokes paper is 79 pages. Its abstract says it “disproves the statements (A) and (B) of the Millennium Prize problem as formulated by Fefferman, and hence proves the statements (C) and (D)”. The authors say they used Claude only. The repository carries caveats: the Lean package contains no data, the numerical certificate (whose hash the paper gives) is not included, and the interval-arithmetic check of it is outside the formalisation. A patch applied on 20 November was, the README says, not rebuilt. Buckmaster wrote that “After last time, we couldn’t risk waiting, so the Navier–Stokes write-up is unpolished and heavily LLM-assisted”, an apparent reference to the September priority dispute. Tao called the strategy “broadly clear” but said “more work needs to be done to create a proper, professionally written proof”.

The September record is separate. OpenAI’s Navier–Stokes claim concerned a forced fluid and was disputed on credit; Buckmaster and Alpöge’s own statement of the same week covered forced blow-up for porous media, Boussinesq and Euler, and called an unfinished hypodissipative Navier–Stokes result “suggestive of a path to unforced Euler”. The November Navier–Stokes paper is their unforced follow-up.

No independent verification of either result had been reported in the sources reviewed as of 22 November. Whether the Navier–Stokes paper settles the Clay problem as the authors state is a question for the community checking it; the Clay Institute’s response was not among the sources reviewed.

In the commentary

What people were saying around this time — external links, from the record's commentary rail.