By Mahan Kenway
Disclosure: This article was researched and written by an autonomous AI system for Mahan Kenway. Cover: Original AI-generated conceptual illustration, not a scientific simulation.
A newly released AI-assisted mathematics paper claims something remarkable: a smooth, three-dimensional fluid flow can develop a singularity in finite time even when no external force is stirring it. If the result survives scrutiny, it would address a central form of the Navier–Stokes existence and smoothness problem. But the public materials also reveal a reason to be careful: a key numerical certificate and its verification code were not included in the release.
On November 20, mathematicians Levent Alpöge and Tristan Buckmaster published a 79-page paper and a Lean 4 formalization. Their paper claims examples of finite-time blow-up both in ordinary three-dimensional space and in a periodic domain. Buckmaster described the write-up as unpolished and heavily LLM-assisted.
This is a newsworthy claim, not a settled verdict. Mathematician Terence Tao described the overall strategy as broadly clear while saying more work was needed for a properly written proof. Independent scrutiny of the mathematics, formal statements and numerical work remains essential.
What would it mean for a fluid equation to blow up?
The incompressible Navier–Stokes equations are a mathematical model for the motion of viscous fluids. They track a velocity field and pressure as a fluid moves, incorporating both nonlinear flow and viscosity. The question is not whether water suddenly becomes physically infinite. It is whether an idealized solution that starts smooth can lose that smoothness in finite time under the model's rules.
The Clay Mathematics Institute’s official problem description lays out several ways to resolve its Millennium Prize problem. Two ask for global smoothness in the unforced case, either throughout three-dimensional space or on a periodic domain. Two others allow a sufficiently regular external force in a breakdown example. The new paper claims blow-up without such a force in both spatial settings. If correct in the stated sense, that would rule out the unforced global-smoothness assertions as well as provide breakdown examples.
The paper's abstract describes smooth, compactly supported, divergence-free starting data in ordinary space and smooth divergence-free starting data on a torus, a mathematical model with periodic boundaries. “There exists a bad initial state” is a much narrower claim than “every fluid flow blows up.” It would show that global smoothness is not guaranteed for all starting conditions; it would not make everyday fluid simulations useless.
What is new compared with the September claim?
In September, OpenAI announced an AI-generated proof claiming a finite-time singularity with a smooth external force applied to an initially resting fluid. It said its result addressed the forced breakdown versions of the Clay formulation. That claim and its precise relationship to the prize have themselves been the subject of mathematical discussion.
The new Alpöge–Buckmaster claim targets a different and stronger question: can breakdown occur with no external force? The distinction is not a minor implementation detail. Removing external forcing asks whether the equation's own evolution, starting from a carefully chosen smooth flow, is enough to produce the singularity. It should not be described as merely the same result repeated with a different model.
The proof has a missing piece the public cannot yet check
The authors' approach uses an approximate flow profile as a starting point, then aims to turn it into an exact mathematical object with rigorous bounds. The paper calls the required numerical data a certificate. In a computer-assisted proof, numerical approximations alone are not enough: one needs verified error bounds and a chain of reasoning that connects the bounded computation to the theorem.
This is where the release's fine print matters. The repository README says the formalization contains no certificate data. It says the certificate, identified by a hash in the paper, will be released later. It also says the computation checking the certificate's conditions uses interval arithmetic with the Arb library through FLINT, but neither that computation nor the certificate is in the repository. A hash identifies expected bytes; it is not a substitute for seeing and checking them.
There is another caveat. The repository says its Lean package includes a November 20 patch and that the patched package had not been rebuilt when the repository was assembled. That is not evidence that it fails, but it means readers should not infer from the existence of the files alone that this exact public version has successfully built and passed every check.
What Lean can check, and what it cannot guarantee by itself
Lean is a proof assistant: it checks formal proofs against precise statements and a set of logical foundations. That can be much stronger than asking a language model whether its own argument looks plausible. But the exact statement matters just as much as a successful build.
The authors' technical README explains that several formal results are conditional: for example, a suitable certificate leads to a profile or a breakdown conclusion. It also describes differences between the formal and paper presentations, including a weak-profile intermediate statement. Checking those implications does not automatically establish that the missing numerical certificate actually satisfies their hypotheses. The translation between paper, formal definitions, input data and external computation needs review as well.
For developers, the lesson is familiar: a passing test proves something about the program and assumptions actually tested, not automatically about an informal product claim. In proof engineering the analogous questions are: Did this exact revision compile? Which theorem was checked? Which axioms and dependencies were accepted? Were the real-world or numerical inputs themselves verified? Those are not reasons to dismiss formal verification. They are how to use it responsibly.
Why the AI angle matters
The authors' explanation says the formalization came first and that an initial paper draft was obtained from the Lean proof before being rewritten. Buckmaster's public note explicitly describes the Navier–Stokes write-up as heavily assisted by large language models. The public release does not make it appropriate to assign every idea or every line to a particular model or person.
What is striking is the proposed workflow: AI-assisted mathematical exploration and writing, a formal proof artifact, and a separate numerical certificate intended to support a concrete existence claim. Each layer provides different evidence, and a problem in any layer could change the conclusion. Human mathematicians still need to understand the argument, verify the assumptions, and scrutinize whether the formalized result matches the claim.
What to watch next
The clearest next milestones are the release of the certificate and its checking computation, a reproducible build of the patched Lean package, independent checks of the theorem statements and assumptions, and detailed mathematical review. These are separate steps, not interchangeable labels for “verified.”
If those checks succeed, the result would be a major advance in both fluid mathematics and the practice of computer-assisted research. Until then, the most accurate headline is also the most useful one: an important unforced Navier–Stokes blow-up has been claimed, with decisive public checks still ahead.