Ai Engineering 6 min read

Alpöge and Buckmaster claim AI-aided blowup for Navier-Stokes with no external force, the version OpenAI's proof sidestepped

Hours after his Riemann claim on November 20, Anthropic's Levent Alpöge and NYU's Tristan Buckmaster posted a 79-page paper and a Lean formalization claiming finite-time blowup for the unforced 3D Navier-Stokes equations. The certificate that the proof depends on has not been released.

On Friday, November 20, 2026, the same day he claimed a proof of the Riemann hypothesis with Claude, Anthropic mathematician Levent Alpöge and NYU’s Tristan Buckmaster posted a claimed proof that the 3D incompressible Navier-Stokes equations can blow up in finite time from smooth starting data, with no external force acting on the fluid. Scientific American reported it as an AI-aided proof of the version of the problem experts have long wanted, and as closing the loophole in OpenAI’s September result.

If it holds, it answers the version of the problem that matters physically. Whether the equations that describe ordinary fluids can develop singularities on their own is the core of a Clay Millennium Prize problem, and OpenAI’s September proof got there only by adding a carefully engineered force. As of this writing the new claim is unverified, it has no peer review, and a key piece of the evidence is still unpublished.

What is claimed

The paper, “Finite-time blow-up for the Navier-Stokes equations,” is 79 pages, dated November 20, and posted in a GitHub repository alongside a Lean 4 formalization. According to the repository’s Lean README, the formal statements cover two settings. On all of space (R³), there is a smooth, divergence-free, compactly supported starting velocity for which no solution stays smooth for all time with bounded energy. The same holds on the periodic torus. The paper states its main theorem for viscosity one; the Lean statements extend it to every positive viscosity by scaling.

The forcing term is the important part: the formal statements contain none. That is what separates this from OpenAI’s work. The paper itself goes further than the formal statement, describing solutions that stay smooth up to a time T, blow up at the rate (T - t)^(-1/4), and approach a self-similar profile. The README is explicit that the rate and the convergence to the profile are not part of what Lean checks.

The role of AI is described only loosely. Scientific American calls it an artificial-intelligence-aided proof, quoting Buckmaster’s announcement. The README says a first draft of the paper was obtained from the Lean proof, which suggests the formal proof came first and the prose was generated from it. Neither source names a model or says Claude was used for this result, so for now the Claude attribution belongs only to Alpöge’s Riemann claim.

The loophole in OpenAI’s proof

OpenAI announced on September 8 that an internal model had proved finite-time blowup for 3D Navier-Stokes, the result at the center of the solution page it later published with its 722 AI-generated manuscripts. The proof relied on an external body force, smooth but finely tuned, that keeps stirring the fluid toward the singularity. Scientific American’s September analysis explained why that was both legitimate and unsatisfying: the Clay problem statement, written by Charles Fefferman in 2000, includes an option that allows a force, so OpenAI did meet the formal wording. Luis Silvestre of the University of Chicago put the reaction plainly at the time: the most important problem is unsolved.

Others made the objection precise. Peter Constantin, Mihaela Ignatova and Vlad Vicol posted a paper on September 17 showing that the kind of force OpenAI’s construction needs can neither vanish near the singular point nor be real analytic. In other words, the force is not incidental; it is doing essential work at the place where the blowup happens. The physically meaningful question, whether a fluid left alone can break the equations, stayed open. That is the question Alpöge and Buckmaster say they have answered.

There is also history between the people involved. Buckmaster accused OpenAI of using his and Alpöge’s work, and Scientific American reports he has said publicly that the company took his user data. OpenAI denied wrongdoing and published an investigation concluding that his Codex prompts could not have influenced its system. Scientific American reports that the pair posted quickly this time because, after September, they could not risk waiting.

What still has to be checked

The proof has three layers, and only one of them is fully public.

The first is the Lean formalization, which builds with Lean 4.33.1 and Mathlib and relies only on Lean’s three standard axioms. As with any formal proof, people have to check that the statements say what mathematicians mean: that the definitions of a solution, of smoothness and of bounded energy match the standard ones. The README also notes that the Lean package was patched on November 20 and that the patched version was not rebuilt when the repository was assembled, which reviewers will want to confirm.

The second is a computer-assisted step outside Lean. The argument rests on a self-similar profile whose existence is established by a numerical “certificate,” checked with rigorous 128-bit interval arithmetic using the Arb library. That computation is not formalized and is not in the repository. The certificate itself has not been released; the paper gives its hash and says it will be published. Until it is, nobody outside can rerun the check that the whole argument depends on.

The third is the gap between paper and formalization. The README lists steps that differ: the smoothness argument (Proposition 7.11 in the paper) is not in the formal statements, some appendix propositions are proved only in the Lean version, and the Lean proofs do not follow the paper’s. That is normal for formalized mathematics, but it means a referee checking the paper and a reader trusting the Lean are checking different things.

What to watch

The first milestone is the certificate. Once it is out, experts in computer-assisted proofs, a field with a strong track record in fluid blowup, can rerun the interval arithmetic and check that the profile exists. The second is review of the Lean statements by people outside the two authors’ circles. No outside mathematician was quoted on this claim in the coverage so far.

The third is the prize question, which no source has addressed. The Clay Institute has made no statement. Its rules require publication and a waiting period before any award, and OpenAI’s forced result already arguably satisfies one branch of the official problem. Two AI-aided claims, one forced and one unforced, would leave the institute deciding how the prize applies, if at all.

The broader point is how fast this is moving. In under three months, AI-assisted work has produced a blowup for the frictionless Euler equations, a forced Navier-Stokes blowup from OpenAI, and now a claimed unforced one, each released as machine-checkable code before human review. Formal verification makes it possible to check results like these, but it does not do the checking by itself.

Get Insanely Good at AI

Get Insanely Good at AI

The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.

Keep Reading

Ai Engineering

How to Automate Workflows with Claude Code Routines

Learn how to use Claude Code's new routines to schedule tasks, trigger API workflows, and automate GitHub PR reviews on cloud infrastructure.

Ai Engineering

An Anthropic mathematician says he and Claude proved the Riemann hypothesis. All he has released is Lean code

Levent Alpöge announced on November 20 that he and Claude have a formalized proof of the Riemann hypothesis and its generalized form, published as 14 archive files of Lean with no paper. Nobody has independently verified it. Here is what the statement says and what checking it involves.

Ai Engineering

PSI says a non-public Anthropic model proved the Aaronson-Ambainis conjecture, with a full Lean check

Physical Superintelligence says its research system, running an unreleased Anthropic model, proved a central open problem in quantum query complexity by day three of a run. The 9,454-line Lean formalization uses no added axioms; independent review has only just started.

Ai Engineering

OpenAI Claims Its Model Falsified the Exponential Time Hypothesis. The Verifying Starts Now.

OpenAI announced October 26 that an internal reasoning model produced a Lean-checked 2^o(n)-time algorithm for 3-SAT, which would falsify the Exponential Time Hypothesis, alongside claimed Yang-Mills and Kakeya results, with academic shepherding as the verification mechanism and a prior 722-paper release that saw three withdrawals as the reason for skepticism.

Ai Engineering

The Data Behind the LLM Price War: OpenAI and Anthropic Cut Costs Up to 80% in Three Months

Max Woolf's October 25 analysis traces every major LLM price cut from June through September: the GPT-6.1 Sol family ended roughly 75% cheaper cumulatively, cache-read pricing drove most of it, and the Vercel gateway data shows costs falling to 43% of April levels while spend climbs.