Tim Gowers: LLMs solve maths by getting lucky at scale, and nearly every recent win was a counterexample, not a proof.

The Fields medalist published his assessment August 12, arguing models excel exactly at problems amenable to "try lots of things till you get lucky" and not at the search-tree pruning humans use for proof discovery (Gowers). His observation about the recent run of ten solved major problems (first non-sofic group, superexponential growth for multicolor Ramsey numbers, Jacobian conjecture, unit distance conjecture) is that they predominantly involved finding counterexamples. Testing ChatGPT 5.6 Pro he found it frequently offers "promising approaches" that don't survive scrutiny. Read against Anthropic's Riemann zeta result the same week, this is the sharpest available framing of what these systems actually do in mathematics: a very fast, very broad search, not a pruner.