·
Out-Mathed by Tuesday
Sam Altman announces the proof: an unreleased OpenAI model and ten thousand agents have claimed the Navier–Stokes Millennium Problem. Jensen Huang wrestles with what it means — his GPUs simulated the unprovable for thirty years, and a language model may have just out-mathed the mathematicians in a week. Then the verification question: can a $1 million prize be awarded for a proof no human fully reads? The personas are invented. The proof claim, the compute bill, and the beef are not.
🔊 Out-Mathed by TuesdayFake Jensen & Fake Sam · VoxCPM2 on the Mac mini Neural Engine · cfg 2.0 · 2:52
The verified facts: On September 8, 2026, OpenAI announced that a multi-agent system — coordinated by an unreleased internal model more capable than GPT-6 Astra, with up to 10,000 concurrent sub-agents — produced a proof that the three-dimensional Navier–Stokes equations can blow up in finite time, resolving statement “C” of the Clay Mathematics Institute's million-dollar Millennium Problem, with a Lean formalization. OpenAI says it will not claim the $1 million prize. The announcement landed amid a priority dispute with NYU's Tristan Buckmaster, who says his team reached a strikingly similar partial result first using AI tools, and a broader revolt among mathematicians about verification and credit.
The Facts Behind It
- The claimed solution: OpenAI announced on Sep 8, 2026 that an internal model — more capable than GPT-6 Astra, with a system of ~10,000 concurrent agents — proved that initially smooth 3D fluids can develop a singularity in finite time (statements “C” and “D” of the Clay formulation). Per OpenAI's own post: agents first resolved an easier Euler-equations blowup question (~100 agents, ~50 hours), then the full effort reached its Navier–Stokes resolution on Saturday Sep 5, about 88 hours in, with Lean verification via GPT-6 Astra taking a further 17 hours. The agents sent 2.7 million messages and used ~130 billion output tokens. The proof's answer is blowup: the solution is a vortex that “spirals inward and gets increasingly elongated, like spaghetti” — energy stays finite while velocity grows without bound. And the stated motive, per Sam Altman 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.” (OpenAI, Sep 8; Nature News, Sep 9)
- The compute bill: OpenAI told reporters the effort used at least ~1,000× the compute of earlier math challenges (~$2 million); other reports put it around $22.5 million. The company says it does not intend to claim the Clay prize. (Fortune, Sep 8; IFLScience, Sep 9)
- The priority fight: NYU's Tristan Buckmaster (with Levent Alpöge of Anthropic) says they had a strikingly similar partial solution using AI tools and alleges OpenAI's model used the same line of attack after the rumors; he questions whether OpenAI's model had access to or was trained on his Codex sessions — which OpenAI's Sébastien Bubeck denies. Buckmaster also alleges an offer to co-announce and remove Alpöge's name, and what he calls a threat; Bubeck disputes the characterizations. (Fortune, Sep 8)
- The million-dollar question: Each Millennium Prize carries $1 million; the Clay Institute (whose president Martin Bridson called it “an exciting day”) has yet to assess the claim, and its rules require a two-year published review. Only one prize has ever been awarded — Perelman's 2010 Poincaré conjecture, which he declined. (Clay Mathematics Institute; IFLScience)
- A product teaser: OpenAI frames the result as a signal of things to come: “We believe it is important to inform the world about the pace of AI progress and what to expect from upcoming models.” The model that produced the proof is internal and unreleased. (OpenAI, Sep 8)
- An 1865 society issues a press release: The London Mathematical Society (founded 1865) issued an official statement on the breakthrough (updated Sep 9). President Mark Chaplain called it “remarkable” and “above all, a great human achievement, enabled by a powerful new tool,” stressing that “the goal remains human understanding.” (London Mathematical Society, Sep 9)
- Mathematicians split: Terence Tao warns that “the indiscriminate strip-mining of open problems” may destroy the ecosystem that produces the next generation of mathematical techniques — answers without insight. The London Mathematical Society's Leiden Declaration on AI and Mathematics (June 2026) sets out principles for AI use and credit in the field; Columbia's Peter Woit calls on the community to demand full documentation of how AI agents produced the result. (Fortune, Sep 8; London Mathematical Society, Jun 2; Not Even Wrong, Sep 9)
- Competing results the same week: Buckmaster–Alpöge and Anandkumar's Caltech group released separate AI-assisted solutions for the zero-viscosity Euler equations on Sep 7, both building on the Córdoba–Martínez-Zoroa program. (Nature News, Sep 9)
- GPUs and the unsolved equation: Computational fluid dynamics numerically approximates Navier–Stokes rather than solving them in closed form — CFD workloads have run on GPUs for decades while the smoothness question stayed open. (Clay Mathematics Institute)