25 Fields Medalists Warn AI Race Undermines Mathematics as 10,000-Agent Proof Pushes Credit Disputes
Updated
Updated · O'Reilly Media · Sep 15
25 Fields Medalists Warn AI Race Undermines Mathematics as 10,000-Agent Proof Pushes Credit Disputes
3 articles · Updated · O'Reilly Media · Sep 15
Summary
Twenty-five Fields Medalists issued a Sept. 11 declaration warning that the rush to solve benchmark problems with AI is distorting mathematics by prioritizing fast results over understanding, explanation and training.
RLVR has accelerated that shift by rewarding machine-checkable outputs, helping drive a recent surge in AI-generated proofs, counterexamples and formalized results that humans may verify without fully understanding.
OpenAI’s reported 10,000-agent, 88-hour run on a Millennium Prize problem crystallized the concern: UC Berkeley’s Abhishek Nagaraj estimated token costs at $20 million to $30 million, underscoring how compute-rich labs can dominate discovery.
The push has also sharpened disputes over credit and publication, with NYU mathematician Tristan Buckmaster saying OpenAI pressed for publication before fuller explanations were ready; OpenAI’s Sébastien Bubeck disputed that account.
The broader warning is that if universities, journals and funders keep rewarding solved problems over exposition and teaching, control over mathematical knowledge could shift toward companies that own the models, data and compute.