AI推翻80年数学猜想，菲尔兹奖得主一夜没睡：以为要出局

TL;DR - A report on how frontier LLMs (OpenAI internal models, Anthropic's Claude/Fable) disproved or advanced several long-open math problems in three months, and how the math community — including Fields medalists and Terence Tao — is reacting. It matters because it marks AI shifting from tool to source of research-level mathematical results.

• Claimed results: an unreleased OpenAI model refuted Erdős's 1946 unit-distance conjecture (open ~80 years); Anthropic's Levent Alpöge reported a high-dimensional counterexample to the Jacobian conjecture via "Claude Fable"; OpenAI's internal "Astra" announced 10 advances (sphere packing, group theory, lattice cryptography) at roughly $2,000 in API-priced tokens, with 5 reproduced on public Fable within 24 hours.
• Fields medalist Timothy Gowers initially feared mathematicians were obsolete, relaxing only on learning the model disproved rather than proved the conjecture; 2026 medalist 邓煜 argues AI will handle "technical details" while humans set theories and frameworks.
• Evaluation signal: the independent "First Proof" benchmark (May 28) gave 4 AI systems 10 unpublished research-level problems; experts judged 7 of 10 solved to publishable quality by at least one system.
• Tao's ICM talk frames a result's lifecycle in six stages (generation, verification, exposition, publication, digestion, canonization); AI accelerates only the first two, producing "proof indigestion" — a shift from proof scarcity to proof surplus, with dozens of unvetted AI proofs already piled up on erdosproblems.com.