OpenAI President Announces Major Breakthrough on Second Millennium Prize Math Problem
nashnova research
OpenAI president Greg Brockman says the company has made major progress on a second Millennium Prize problem; if it turns out to be P vs NP, the mathematical foundations of modern cryptography and financial security infrastructure face a fundamental challenge.
How was the first problem solved?
OpenAI deployed roughly 10,000 reasoning agents over 88 continuous hours, generating 2.7 million messages and consuming about 130 billion output tokens.
The conclusion: smooth fluid flow can develop a singularity — a mathematical "blow-up" where velocity or pressure becomes infinite — in finite time.
The entire proof was formally verified in Lean, a machine-proof language that lets a computer check every logical step line by line. This means → the result is not "the AI says it's right"; every step passed machine-auditable logical closure.
What might the second problem be — and why should markets care?
Five Millennium Prize problems remain unsolved: P vs NP, the Riemann Hypothesis, Yang–Mills and the mass gap, the Hodge Conjecture, and the Birch and Swinnerton-Dyer Conjecture.
The most closely watched speculation inside the tech world points to P vs NP. In plain terms = if P equals NP is proven true, encryption algorithms currently considered "practically unbreakable" could, in theory, all be cracked quickly.
This means → asymmetric encryption — the core security mechanism behind internet logins, bank transfers, and digital signatures — would lose its mathematical foundation. The entire financial and internet security infrastructure would need reassessment.
What is the academic-credit controversy?
NYU mathematician Tristan Buckmaster and Anthropic mathematician Levent Alpöge had already made preliminary progress on Navier–Stokes, using Codex and Claude in the process.
Buckmaster claims his progress was relayed to OpenAI before publication. OpenAI then launched a parallel effort, using an unreleased next-generation model to consume roughly 300 billion output tokens — about $22.5 million at Astra rates — and published a full proof within about a week.
Buckmaster also alleges that OpenAI mathematician Sébastien Bubeck pressured him to remove Alpöge's name from the work, with threats including "destroy your career." This reflects a rapidly intensifying conflict between AI companies and academia over intellectual-property attribution.
How has OpenAI responded?
Brockman stated publicly that the model used for the final proof was trained on data cut off in early July this year — before external researchers' results were made public — and denied any use of private data.
The academic backlash is still widening: hundreds of mathematicians signed a petition opposing OpenAI's sponsorship of Caltech's inaugural "Math Hackathon," and OpenAI has been forced to withdraw its support entirely.
NYU professor Scott Armstrong summed it up: "The singularity has already begun — it's just extremely unevenly distributed."
Should top-tier compute go to math puzzles or to saving lives?
Brockman revealed that on the eve of GPT-5.5's launch, OpenAI was agonizing over compute quotas and painfully cutting API rate limits, calling compute allocation "the hardest capital-and-engineering problem we face."
Public criticism centers on a stark question: should the world's most powerful supercomputing resources be spent solving abstract conjectures, or tackling Alzheimer's and shortening drug-development timelines?
Brockman's argument: a system that can reason through dozens of logical steps without error — and close the loop entirely in Lean code — will produce exponential gains once transferred to life sciences. In plain terms = prove that AI can reason rigorously on math first, then move that same capability to drug discovery. Math is the test track, not the finish line.
Should top-tier compute be spent on abstract math?
BULL
Reasoning ability transfers
The math proof validated AI's long-chain reasoning rigor — transferring it to life sciences could yield exponential gains.
Formal verification is a moat
Lean-based closure means reasoning is traceable and reproducible — that capability itself has enormous commercial value.
BEAR
Opportunity cost is steep
A single proof consumed roughly $22.5 million in compute; the same resources could go directly to disease research.
Transfer is unproven
The cross-domain leap from mathematical reasoning to molecular dynamics has no publicly demonstrated results yet.
In plain terms = Brockman's logic is 'prove AI can reason rigorously first, then move it to save lives.' The logic is sound — but the 'move it' step remains unvalidated, and that is the core of the current debate.
市场有风险，内容仅供研究参考，不构成投资建议。