Navier-Stokes And The Day AI Solved an 'Impossible'' Math Problem

Ronni Holmvig Strøm · 2026-09-11

On September 8, OpenAI announced that an internal model had constructed a finite-time singularity for the three-dimensional Navier-Stokes equations, and the announcement arrived dressed in numbers that belong more naturally to an industrial process than to mathematics: roughly 10,000 agents running concurrently for about 88 hours, 2.7 million messages passed between them, some 130 billion output tokens, and then a further 17 hours in which GPT-6 Astra rendered the argument in Lean 4 so that a proof checker could confirm it.

The paper runs to 166 pages. The Lean project sits in a public repository with build instructions and a tool for independent checking. Anyone willing to spend the compute can re-run the verification and watch it pass.

What the announcement could not establish is that the mathematics is understood. Correctness and comprehension have come apart, and they are now keeping separate time.

The clearest evidence for that sits at the other end of the story. Tristan Buckmaster of NYU and Levent Alpöge spent close to a year on the forcing program developed by Diego Córdoba and Luis Martínez-Zoroa, working with publicly available models from both OpenAI and Anthropic, and on August 15 they obtained finite-time blowup for the forced Euler equations. They verified it in Lean on August 22. The English-language explanation their models produced was, in Buckmaster's account to Scientific American, "barely readable." The two of them then went back through the argument step by step and rewrote it in a form other researchers could digest.

That rewriting is the most consequential piece of work in the entire episode, and it is the piece nobody is arguing about. Call it the translation layer: the labor of converting a machine-verified argument into something a human field can absorb, teach, generalize, and build on. It has no prize attached, no press call, and no line in any benchmark.

The news cycle has been careless with what was actually proved. Charles Fefferman's official Clay problem statement offers four alternatives. (A) and (B) ask whether smooth, unforced initial data always leads to a solution that stays smooth, on all of space and on the periodic torus. (C) and (D) ask for a counterexample, and they explicitly permit a smooth external force meeting stated decay and periodicity conditions.

OpenAI claims (C) and (D). Its Theorem 1.1 constructs, for every positive viscosity, a fluid starting at rest, driven by a force that is smooth and compactly supported in space and time, whose velocity becomes unbounded in finite time while its kinetic energy stays finite. A rescaling and periodization argument carries it to the torus.

The popular version of the question, the one about whether turbulence can spontaneously tear itself apart with nobody pushing on it, is untouched. Alternatives (A) and (B) remain open. Stan Palasek at Princeton has already identified an obstacle to removing the force along the obvious route, on the grounds that viscous dissipation would swamp the growth mechanism, and Terence Tao welcomed the observation.

The construction itself is beautiful, and worth a sentence for the reader who will not open a 166-page PDF today. A vortex core contracts faster in width than in length, spinning up as it narrows because angular momentum is nearly conserved, so azimuthal speed runs away while the energy in the shrinking core stays bounded. That object solves nothing on its own. Its momentum residual is rewritten as a stress, and two families of divergence-free oscillatory waves are constructed whose averaged quadratic momentum flux reproduces that stress exactly. Then a fixed four-step correction cycle runs over and over, improving a decay exponent by one tenth on each pass, until every derivative of the remaining error vanishes to every order as the singular time approaches.

Three sentences of prose for a result that took a swarm eighty-eight hours and a century of preparatory analysis. This is what translation buys, and why it cannot be optional.

The credit dispute will dominate the coverage, and it deserves less of our attention than it is getting. Buckmaster has alleged that OpenAI pursued the forcing route only after word of his unpublished work reached the company, that Sébastien Bubeck proposed a paper excluding Alpöge because of his employment at Anthropic, and that when he threatened to go to the press the reply was, "Why would you ruin your career?" OpenAI denies that any agent or employee saw the pair's private transcripts. It has declined to say whether such material could have entered training. Its own write-up states that the effort began on September 1 "after hearing a rumor which we later realized was related to" Alpöge and Buckmaster, and it now credits both men and offers to recognize their priority on forced Euler.

Provenance is the one question in this story that no amount of formalization can settle, because it is a claim about a training process nobody outside the company can inspect.