ThinkFacility Sign in
  1. Home
  2. News
  3. Anthropic

Alpöge and Buckmaster claim Navier–Stokes can blow up with no outside force, unlike OpenAI's result

This time the fluid gets no push from outside, and the authors say a key numerical certificate isn't public yet.

On November 20, 2026, at 8:37 am Eastern, NYU mathematician Tristan Buckmaster posted two links on Mastodon. One was a GitHub repository holding a 79-page paper by Levent Alpöge and Buckmaster, Finite-time blow-up for the Navier–Stokes equations, along with a Lean 4 formalization of it. The other was a paper on the Euler equations, written with Matei P. Coiculescu.

Both papers claim a singularity with no outside force pushing the fluid. That's the version of the problem OpenAI's September result didn't reach. Alpöge works at Anthropic, and the pair say they used Claude and OpenAI's Codex throughout.

Buckmaster's own post is short, and it's blunt about the write-up.

After last time, we couldn’t risk waiting, so the Navier–Stokes write-up is unpolished and heavily LLM-assisted.

From https://mastodon.social/api/v1/statuses/117649035783718539
Who
L. Alpöge and T. Buckmaster
Paper
79 pages, version of 20 November 2026
Lean library
about 2,270 files and 600,000 lines
Not in the repository
the certificate whose hash the paper gives
Lean
v4.33.1 with Mathlib

What changed since September

On September 8, 2026, OpenAI said an internal system had found a singularity in the Navier–Stokes equations, the equations that describe how fluids such as air and water move. OpenAI's post says on the order of 10,000 agents reached it in about 88 hours. Its own description of the result also says the fluid has a smooth force applied to it, and that it establishes statements C and D of the official Clay formulation.

Buckmaster's post files the new paper under unforced Navier–Stokes. The Lean notes in the repository describe its main conclusion with no force in it: for every viscosity, a smooth, compactly supported starting state for which the equations have no smooth solution with bounded energy that lasts forever.

The Euler paper makes the same move for fluids with no viscosity. Its authors say they start from smooth periodic data on a three-dimensional torus and that the solution breaks down in finite time, with the speed staying bounded while the spin at the center grows without limit. OpenAI's September post says its agents had already done unforced Euler, and that nearly 100 agents took about 50 hours on it. The new Euler paper says Alpöge and Buckmaster had a second forced Euler proof using the same wave mechanism as OpenAI's, finished before OpenAI released its unforced Euler paper.

What the formal proof covers, and what it doesn't

The repository is more candid than most announcements. Its Lean README says the formalization came first and a first draft of the paper was obtained from the Lean proof, so the two prove the same statements by different routes.

The formalization contains no data. The certificate whose hash the paper gives is not in this repository; the paper says that it will be released.

From GitHub

That certificate is the numerical part, and the README says its check runs on interval arithmetic (128 bits of precision, using the Arb library through FLINT) outside the formalization. It also says one patch dated 20 November 2026 was applied to the Lean text and that the patched package wasn't rebuilt when the repository was put together. In plain terms, nobody outside can check the full chain yet, and the authors say so themselves.

The README lists a gap on the formal side too. Lean's version of one theorem stops at a weak profile and doesn't record that the solution is smooth. The paper gets smoothness from standard regularity theory, while the library takes another path to the final statements. I read that as careful bookkeeping. Still, it's the authors' own list, and anyone checking the proof will want it.

Who wrote it, the people or the models?

The Euler paper has an AI statement. Its authors say the first proof without a force was written by Anthropic models from their own drafts on the forced problems, and that they then rewrote it with Claude and Codex, swapping in ideas of their own that the models implemented. The paper's acknowledgments thank Anthropic for supporting the collaboration. In September, Buckmaster's own statement had described it as a purely personal collaboration with no involvement from either employer.

The Navier–Stokes post leans the other way on effort. Buckmaster writes that once the numerical profile is in hand the rest follows a familiar computer-assisted route, which puts most of the weight on finding that profile.

It was Alpöge's second claim on a famous problem that day. We covered his Riemann hypothesis claim, which he published only in Lean, on November 20 as well.

Why a fluid equation matters outside math

OpenAI's September post notes that these equations are used for aircraft design, weather forecasting and the study of blood flow. The Clay Mathematics Institute's page on the problem puts it in everyday terms, with waves behind a boat and turbulence behind a jet.

Although these equations were written down in the 19th Century, our understanding of them remains minimal.

Navier-Stokes Equation - Clay Mathematics Institute

A blow-up proof won't change a weather model tomorrow. What it settles is narrower. OpenAI's post explains that a real fluid can't move infinitely fast, so a singularity would mark a point where the equations stop describing the fluid at all. A proof without a force says that can happen from a smooth start, with nothing feeding energy in. None of these sources says how soon that would touch working forecast or design software, and I won't guess.

The other effect is on how mathematics gets checked. A 600,000-line formal library with an unreleased certificate is a new kind of thing to referee. OpenAI said in September that it doesn't intend to claim the Millennium Prize for its result.

The repository says the paper promises to release the certificate. Neither the repository nor Buckmaster's post gives a date for it.

More on Anthropic

All Anthropic stories