Fields Medalist Timothy Gowers says most famous mathematics problems solved by large language models so far have almost all been handled through counterexamples rather than proofs, according to a Techmeme item pointing to Gowers’s Weblog. The available summary does not name the specific problems, models, or papers Gowers is discussing. It also does not provide examples of the counterexamples or proofs at issue. The only supported takeaway from the cluster is the distinction Gowers is drawing: LLM progress on notable mathematics problems, in his view, has leaned toward finding counterexamples rather than constructing proofs. That distinction matters because the two tasks are not the same. A counterexample can disprove a conjecture by finding one case where it fails; a proof must establish a claim across the relevant mathematical domain. The cluster does not say how Gowers evaluates the relative difficulty of those tasks, so that point should not be overstated. For AI readers, the item is best treated as expert commentary rather than a new benchmark result. It signals how a leading mathematician is framing reported LLM successes in mathematics, but the provided material is too thin to support broader claims about model capability, reliability, or the direction of AI-assisted research. Who benefits: Teams evaluating AI for mathematical or formal reasoning benefit from sharper language around what counts as progress. Counterexample search and proof generation should not be treated as interchangeable from this cluster alone. Who's exposed: Anyone marketing LLMs as broadly solving famous mathematics problems is exposed to scrutiny over the type of solution being claimed. The provided item supports that scrutiny only at a high level.