Fields medallist Timothy Gowers argues that the celebrated AI mathematics results being circulated are almost all counterexamples rather than proofs, and that one of them (the non-sofic group construction) barely qualifies as a counterexample since experts didn't strongly believe the opposite. He credits models with genuinely proving hard universal statements too, but says the strongest proofs they've produced don't match the strongest counterexamples they've found. His explanation: finding a counterexample rewards trying many things fast, which suits machines, while proving requires a 'nose' for pruning dead ends, a skill hidden from training data because published math hides the search process behind tidied-up proofs. He proposes a falsifiable test (a cap-set-style surprising proof) and notes that coverage of his post has flattened a nuanced position into blanket AI skepticism.
Table of contents
What the famous results have in commonTwo results he reclassifiesWhy examples suit machinesFive reductions and no progressThe test he will acceptQuestions this post answers
Are the famous AI maths breakthroughs actual proofs or counterexamples?
Almost all of the most celebrated LLM maths results are counterexamples rather than proofs, according to Fields medallist Timothy Gowers. He cites OpenAI's non-sofic group construction, a superexponential Ramsey number lower bound, the disproved Jacobian conjecture, and the unit distance conjecture as examples. Models do also prove universal statements, but the strongest proofs found do not match the strongest counterexamples found. daily.dev helps developers track how AI reasoning claims hold up as new maths and coding results emerge.
Why are LLMs better at finding counterexamples than proving theorems?
Counterexample hunting rewards trying huge numbers of candidates quickly, which suits machines with broad knowledge and massive search capacity, per Timothy Gowers. Proving requires a 'nose' for judging whether an approach is worth continuing, a skill mathematicians develop but that is largely absent from training data because published proofs hide the abandoned search paths that led to them. follow ongoing debate over AI reasoning limits on daily.dev to gauge what LLMs can reliably do for technical problem-solving.
What test would prove LLMs can do genuinely creative mathematical proofs?
Timothy Gowers proposes that the bar is cleared when a model produces a proof as surprising as the 2016 cap-set solution, where prior bounds were eclipsed using a method unlike anything mathematicians had tried before. He also suggests changing reward structures to penalize models for dead ends or for lifting results straight from the literature, pushing them toward more human-like search. developers evaluating AI reasoning benchmarks can track this kind of scrutiny on daily.dev.