Gowers' GPT-5.5 Pro math claim questioned

This is a real signal about model capability, but the strongest version of the headline looks overstated. The useful story here is not “AI solved math,” but “frontier models can compress literature digestion and proof assembly fast enough to confuse rediscovery with discovery.”

• –Gowers’s writeup suggests the model produced a non-trivial combinatorics result with minimal human input, which is still notable
• –The criticism is about provenance and timing: if the key theorem or method was already public, the claim shifts from breakthrough to recombination
• –Even a rediscovery matters, because research productivity often comes from finding the right existing technique faster than a human would
• –For AI builders, this is a reminder that capability claims need provenance-aware evaluation, not just impressive anecdotes
• –The broader implication is that math research may be moving toward human-plus-model workflows rather than fully autonomous theorem proving