OpenAI researcher announced GPT-5 math breakthrough that never happened (the-decoder.com)

OpenAI researchers briefly claimed on X that GPT-5 had “found solutions to 10 (!) previously unsolved Erdős problems” and made progress on eleven more, prompting excitement that generative AI had independently cracked longstanding number‑theory questions. The post was quickly deleted after mathematicians pushed back: Thomas Bloom (who runs erdosproblems.com) clarified that “open” on his site meant the problem was unknown to him personally, not provably unsolved, and that GPT-5 had surfaced existing papers Bloom had missed rather than producing new proofs. OpenAI researchers acknowledged the misstatement, and critics including DeepMind CEO Demis Hassabis called out the sloppy communication—highlighting how ambiguous phrasing (notably from figures like Bubeck) can amplify hype. The substantive takeaway is more modest but important for AI/ML: GPT-5 appears to be a strong literature‑search and synthesis assistant—useful at trawling scattered or inconsistently labeled math literature and accelerating tedious discovery tasks—rather than an autonomous theorem prover cracking deep open problems. Prominent mathematicians like Terence Tao view this as the immediate, practical role for generative models: speeding workflows, organizing results, and “industrializing” parts of mathematical research. The episode underscores two technical and cultural lessons for the field: rigorous verification and transparent communication about model outputs are essential, and human expertise remains crucial to validate, classify, and safely integrate AI‑generated findings into real research.