Two months following OpenAI's announcement of its own result, Tristan Buckmaster of New York University and Levent Alpöge of Anthropic made their Navier-Stokes singularity proof available on GitHub on Friday. The submission includes a 79-page paper dated 20 November and a formal verification written in Lean, though certain components remain incomplete.
The Navier-Stokes equations model the behavior of fluids including water and air. A "blow-up" occurs when a previously smooth solution becomes singular at some point within a finite time interval. Whether such singularities can develop represents one of the Clay Mathematics Institute's seven $1 million Millennium Prize Problems.
OpenAI released its own blow-up proof in September, but that construction required applying an external force to the fluid. The new work by Buckmaster and Alpöge demonstrates that singularities can form even without any such forcing.
What was released
The GitHub repository contains a 79-page paper and a Lean formalization that verifies each logical step. The Lean library comprises approximately 2,270 files and 600,000 lines of code.
The repository documentation identifies several incomplete elements. A numerical certificate whose cryptographic hash appears in the paper has not yet been included, though the authors indicate they will publish it. The computation validating this certificate operates independently of the formal proof. Additionally, the authors modified the Lean files on 20 November but have not yet recompiled the updated package.
Buckmaster addressed the incomplete state of the submission in a Mastodon post:
After last time, we couldn't risk waiting, so the Navier-Stokes write-up is unpolished and heavily LLM-assisted. Fortunately, once the numerical profile is in hand, the proof follows a well-trodden computer-assisted path.
Tristan Buckmaster
Tao's first read
UCLA mathematician Terence Tao evaluated the paper on his blog, concluding that it is technically accessible to human readers. He described the overall approach as clear: construct an approximate solution through numerical methods, then employ rigorous perturbation theory to convert it into an exact solution. However, Tao noted that a fully polished proof would require additional refinement.
An Euler paper too
On the same day, Alpöge and Buckmaster, together with Matei Coiculescu, a junior fellow at NYU, released a separate 55-page paper. This work establishes blow-up for the Euler equations, which govern the motion of inviscid fluids, also without external forcing. Tao characterized this proof as human-readable, and it includes its own Lean formalization.
OpenAI had already addressed the unforced Euler problem. In September, the company reported that approximately 100 AI agents had solved this question in roughly 50 hours. The new paper notes that before OpenAI's announcement, Alpöge and Buckmaster had completed an unpublished alternative proof that included external forcing and employed the same underlying mechanism.
Throughout their work, the authors employed Claude and Codex. Anthropic's models generated the initial unforced proof, which the team subsequently refined through multiple iterations using the same models.
After last time
In September, Buckmaster and Alpöge shared their forced results a single day before OpenAI unveiled its proof. This timing discrepancy sparked controversy. OpenAI subsequently stated that its investigation determined Buckmaster's Codex prompts could not have affected its own system.
Alpöge's Friday announcements extended beyond the Navier-Stokes and Euler results. He also claimed to have discovered a proof of the Riemann hypothesis using Claude.
Source: The Next Web



