Claude does real math, raises real questions about what that means

An Anthropic thread about Claude's mathematical capabilities got serious attention. A research version of Claude improved on a longstanding lower bound related to the Riemann hypothesis, specifically by combining existing results from Baluyot, Goldston, Suriajaya, and Turnage-Butterbaugh in a way no human researcher had tried. The initial attempt failed, but after a non-mathematician staffer prompted it to 'take a real stab,' it made genuine progress.

The comments were careful. People noted this is an improvement on a bound, not a proof of the hypothesis itself, and that the mathematical choices were guided but not pre-specified. Still, this is the kind of result that shifts priors. It's not that Claude understood the problem the way a mathematician does; it's that it was able to synthesize across a literature large enough that no single human researcher holds it all in working memory at once.

The implication being debated: is this the beginning of AI as a genuine research collaborator in hard mathematics, or a one-off that reflects the particular structure of this problem? The HN thread leaned toward cautious optimism, with people noting that the methodology, prompting a non-expert to push the model harder than a domain expert would, is itself an interesting finding.