GPT-6 Hits FrontierMath, Claims Navier-Stokes Proof in 88 Hours; 25 Fields Medalists Push Back

新智元 · wechat · 2026-09-13

WSJ columnist Ben Cohen warns AI's math progress is accelerating frighteningly fast, as 25 Fields Medalists (incl. Terence Tao) issue a joint statement on a 'severe misalignment of AI in mathematics.'
Benchmark saturated: GPT-6 Astra scored 97.6% on FrontierMath Tier4 (a level only 5% when launched in July 2025); on Sept 10 EpochAI declared Tier4 fully solved and is moving evals to open problems with no known answers.
88-hour AI research org: OpenAI ran 10,000 agents in collaborating groups (proving, refuting, sharing intermediate results, switching models mid-run), producing a 100-page Navier-Stokes singularity proof in 88 hours, with claimed Lean formal verification; OpenAI won't claim the $1M prize, and the math community hasn't verified it (official review takes years).
Attribution fight: NYU's Tristan Buckmaster and Anthropic employee Levent Alpöge were working on related problems in August; OpenAI offered concessions but refused to credit Alpöge, and Buckmaster suspected his private Codex prompts fed OpenAI's model — OpenAI denies it.
Core worry: machines now generate knowledge hundreds of times faster than humans can verify, attribute, and absorb it; GPT-6 Astra is also OpenAI's first model hitting the cybersecurity 'Critical' threshold, and OpenAI itself says it may need to be more deliberate about its pace.

More from AGI Musings