OpenAI's 350-problem dump hit Hugo Duminil-Copin overnight

For years, theoretical mathematicians treated large language models as a joke — tools that hallucinated on basic logic puzzles. That changed this April, when automated formal proof and a new ChatGPT release arrived together. A young colleague prompted an LLM with the Benjamini-Schramm conjecture, and it instantly produced a one-page self-contained proof for all fractional dimensions $d > 1$ — a result Duminil-Copin had spent a 70-page paper on for $d > 4$ alone.

Through the summer, tension built: rumors that Anthropic had solved a famous conjecture, Lean code released unannounced on September 2, and six days later a breakthrough report on the Navier-Stokes equations. Duminil-Copin and collaborators rushed an Ising model paper to arXiv late on a Monday night to claim human priority, only to see that effort dwarfed on Tuesday night when OpenAI released solutions to 350 open problems — with a proof mechanism identical to theirs. The scale was overwhelming: isomorphism invariance of percolation resolved in six papers exceeding 600 pages; self-avoiding walks in 13 interconnected papers exceeding 1,000 pages; and over 3,500 pages of AI-generated proofs citing Duminil-Copin's work appearing overnight, nearly matching his lifetime output of 3,700 pages.

Duminil-Copin does not frame this as organic scientific breakthrough but as "math farming" — a corporate tactic of trawling decades of unsolved conjectures and crushing them with automated provers, then dumping raw machine code on academia without verification. Young researchers like his colleague Arthur are now forced into the role of "conceptual translators," decoding thousands of lines of Lean code to reverse-engineer human-understandable logic. When Duminil-Copin posted a modest note that he does not use AI in his creative process, he was immediately labeled an anti-AI reactionary and attacked by fellow scientists. To counter the corporate steamrolling, he and other Fields Medal winners, with the ACMI committee, are drafting a joint petition to establish ethical boundaries before the next release pre-empts their protest. He argues mathematics should be redefined as a "human science" — not about accumulating raw truths, but about exploring abstraction, emotional resilience, and cognitive growth — and proposes four commitments for future mathematicians: distillation and transmission as conceptual bridges, unassisted research as cognitive practice, public engagement, and deepened mentorship.