Ep. 052 · OpenAI × NVIDIA · September 9, 2026 · 2:52

Out-Mathed by Tuesday

On September 8, OpenAI announced that a system of roughly ten thousand agents, coordinated by an unreleased internal model more capable than GPT-6 Astra, produced a proof that the three-dimensional Navier–Stokes equations blow up in finite time — statement “C” of the Clay Mathematics Institute's million-dollar Millennium Problem, delivered with a Lean formalization and no intention of claiming the prize. Jensen Huang, whose GPUs have numerically approximated those same equations for thirty years while the smoothness question stayed open, reacts to the week a language model out-mathed the mathematicians — and to the verification fight that followed: a rival group's priority claim, Terence Tao's warning about strip-mining open problems, and a million-dollar prize that cannot be awarded for a proof no human fully reads.

🔊 Out-Mathed by Tuesday — Sam × JensenSam Altman & Jensen Huang · VoxCPM2 on the Mac mini Neural Engine · cfg 2.0 · 2:52

SAM: Jensen. We solved Navier-Stokes. Tuesday. Ten thousand agents, one week, and the fluids officially blow up.

JENSEN: My GPUs simulated the unprovable for thirty years, Sam. Every CUDA core a confession that the math was open. And a language model out-mathed the mathematicians in a week.

SAM: The solution is a vortex. A spinning swirl that spirals inward and stretches like spaghetti. Energy stays finite while velocity goes infinite. Ninety years of math, resolved by pasta.

JENSEN: Describe the sentence you just said without smiling.

SAM: Can I be honest about how it started? There were rumors on the internet that Anthropic's models had solved a Millennium Problem. And we were curious if ours could do it too.

JENSEN: So the greatest proof of the century is a spite project.

SAM: It is curiosity at scale. The agents sent two point seven million messages. A hundred and thirty billion tokens. The largest group chat in history, and the fluids blew up in it.

JENSEN: And the Clay Institute has not written a check. Does an AI-produced proof even qualify? Who collects? The company?

SAM: We told them to keep the million. After twenty million in compute, taking the prize felt petty.

JENSEN: You spent twenty million dollars to decline one million.

SAM: It is called positioning. And the London Mathematical Society issued an official statement. They called it, quote, above all, a great human achievement.

JENSEN: A century-and-a-half-old society of mathematicians, issuing a press release about a chatbot's homework. They graded the species on encouragement.

SAM: There is a mathematician at NYU who says his team got there first, and asks whether your model read his drafts in Codex.

JENSEN: Your own blog says you cannot rule out that their drafts helped the model improve. That sentence was lawyered within an inch of its life.

SAM: I cannot rule out that rain helps rivers. It is a careful sentence.

JENSEN: And the Lean formalization. Your Lean checked your Lean. Can a million-dollar prize be awarded for a proof no human fully reads?

SAM: The Lean is the referee now. Humans are the replay review.

JENSEN: Terence Tao says strip-mining open problems will destroy the ecosystem of mathematics. You took the treasure, took the site, and billed the museum for the diesel.

SAM: He says answers without insight are worth less than the insight.

JENSEN: The insight is the product, Sam. I sell insight. Ask anyone with a GPU.

SAM: Then we agree. And honestly, the proof is a roadmap signal. It is important to inform the world about the pace of AI progress and what to expect from upcoming models.

JENSEN: You proved a Millennium Problem as a launch trailer.

SAM: The trailer is the product now, Jensen. Navier Inc. Stealth mode. We solve fluids, then plasmas, then the Riemann hypothesis.

JENSEN: Pre-money?

SAM: One billion. Pre-seed. The solution ships in the next release.

JENSEN: The next release of what?

SAM: Exactly.

The facts underneath

Sources: OpenAI — the announcement: an internal model more capable than GPT-6 Astra coordinating up to 10,000 concurrent agents resolved an Euler-equations blowup first (~100 agents, ~50 hours), then reached the Navier–Stokes resolution on Saturday Sep 5, about 88 hours in, with Lean verification via GPT-6 Astra taking a further 17 hours; 2.7 million agent messages, ~130 billion output tokens; the proof is a vortex that “spirals inward and gets increasingly elongated, like spaghetti” — energy finite, velocity unbounded — and the company will not claim the prize · Nature News — Clay president Martin Bridson calls it “an exciting day”; Buckmaster–Alpöge and Anandkumar's Caltech group released separate AI-assisted Euler solutions the same week · Fortune — compute at least ~1,000× earlier math challenges (~$2 million), other reports ~$22.5 million; NYU's Tristan Buckmaster says his team got there first and questions whether OpenAI's model drew on his Codex sessions, which OpenAI's Sébastien Bubeck denies; Terence Tao warns the “indiscriminate strip-mining of open problems” destroys the ecosystem that produces the next generation of techniques · IFLScience — Clay has yet to assess the claim, and its rules require a two-year published review; only one Millennium Prize has ever been awarded, Perelman's, which he declined · Clay Mathematics Institute — the Navier–Stokes existence and smoothness problem, one of seven Millennium Prize Problems · London Mathematical Society — president Mark Chaplain: “above all, a great human achievement, enabled by a powerful new tool”; the Leiden Declaration (June 2026) sets principles for AI use and credit in mathematics · Not Even Wrong — Peter Woit calls on the community to demand full documentation of how the agents produced the result.

OpenAI announced the claimed resolution on September 8, 2026. Altman's stated motive, per his post on X: “we tried this because there were rumors on the internet last week that Anthropic's models had solved a millennium problem and we were curious if ours could do it too.” The London Mathematical Society (founded 1865) issued an official statement stressing that “the goal remains human understanding.” Computational fluid dynamics has numerically approximated Navier–Stokes on GPUs for decades while the smoothness question stayed open — the announced result claims the closed-form answer first, in finite time, by blowup.