OpenAI's Mysterious Model Allegedly Cracks Navier-Stokes Millennium Prize Problem
nashnova research
Mathematician Tristan Buckmaster disclosed that an internal OpenAI model has produced a 100-page proof of finite-time blowup for the Navier–Stokes equations under external forcing — if it holds, AI may have independently touched the core of a million-dollar Millennium Prize problem for the first time.
Why is this problem worth a million dollars?
The Navier–Stokes equations — the master equations describing the motion of all fluids, from ocean currents to airflow over a wing — are one of the seven Millennium Prize Problems set by the Clay Mathematics Institute in 2000. Solving any one of them carries a $1 million prize.
The central open question: does a solution to these equations always stay smooth, or can it "blow up" — reaching infinity in finite time? In plain terms = can the math describing a fluid suddenly break down and lose all meaning?
Mathematicians call such a breakdown "blowup" (a singularity). Proving it happens, or proving it never does, would both count as a solution. The problem has been open for over two decades.
What exactly did OpenAI's 100-page proof prove?
NYU mathematician Tristan Buckmaster disclosed that a mystery model inside OpenAI independently generated a 100-page mathematical manuscript, completing a finite-time blowup proof for the N-S equations under "external forcing."
This means → the model did not solve the full N-S problem. It proved that, under a specific condition (an applied external force), the solution does blow up.
The critical detail: the Clay Institute's official problem statement allows constructing smooth forcing to show that no global smooth solution exists. If this proof meets all the formal conditions, it could in principle resolve the official prize problem. The model is believed to be OpenAI's next-generation product, "Bel."
Three more fluid equations broken on the same day?
Buckmaster and Levent Alpöge (an Anthropic employee) announced that, with large-language-model assistance, they completed finite-time blowup results for three classes of fluid equations — the incompressible porous media equation, the Boussinesq equation, and the 3D incompressible Euler equations.
Their papers and Lean formalized code — a system that lets a computer verify every step of a mathematical proof — are all publicly available. The researchers used Claude, Codex, and GPT-5.6 Sol to assist the work, with Astra used for manuscript editing and proof review.
Fields Medalist Terence Tao called the results "a remarkable achievement" but was explicit: the Euler breakthrough cannot directly be declared as solving N-S — the viscosity term in N-S smooths out high-frequency structures, and extending the blowup mechanism still faces major technical obstacles.
How does Terence Tao explain the proof strategy?
Tao's reading of the core idea: layer high-frequency perturbations repeatedly onto an existing flow, so that the large-scale flow keeps amplifying fine structures.
In plain terms = imagine a river where you keep throwing in smaller and smaller whirlpools. The river amplifies each tiny whirlpool, while the whirlpools' feedback on the river is kept tightly controlled — until the whole system spirals out of control in finite time.
This reflects a "cascade amplification" construction strategy. AI's role was to assist and verify within this extraordinarily intricate construction process.
What is the authorship dispute about?
Buckmaster disclosed that on September 3, news spread online that Anthropic's Claude had solved a major N-S problem. That same day, his and Alpöge's research progress reached OpenAI internally. He immediately contacted a prominent mathematician at OpenAI, stressing the work was unaffiliated with any institution.
On September 6, OpenAI's Sébastien Bubeck joined the conversation. During two phone calls that afternoon, Buckmaster learned of the 100-page proof for the first time. He directly asked whether OpenAI had used his project drafts or conversation data previously entered into Codex for training. OpenAI said the model did not search user data, but gave no answer when pressed on the training question.
During negotiations over publication, OpenAI twice asked to remove Alpöge's name from the paper, citing his status as an Anthropic employee. Buckmaster refused. The two sides remain at an impasse.
What is the real unresolved question here?
Buckmaster says he has not yet received the 100-page manuscript. OpenAI's full proof remains unpublished. This means → the mathematics community currently has no way to verify whether the proof holds.
Even once it is made public, two questions will determine its historical significance: whether the proof survives rigorous peer review, and how to define the boundary of AI's independent contribution.
This reflects a wholly new tension created by AI's entry into frontier mathematics — when a model may have been exposed to a researcher's data, and its output closely overlaps with a human team's work, the question of "who got there first" becomes unprecedentedly complex.
市场有风险，内容仅供研究参考，不构成投资建议。