OpenAI posts 722 AI math papers, 3 pulled next day

This was not a case of the AI making three separate mistakes. According to the account, one manuscript contained a proof for a highly advanced geometric result, and a sign error in its middle made a key explanation fail; two other manuscripts had been built on that explanation, so all three were withdrawn together. OpenAI also revised 14 other manuscripts the same day, filling gaps in proofs and rewriting necessary conditions, and updated references for 13 more to point to the corrected versions. The 722 manuscripts were not 722 distinct breakthroughs: OpenAI grouped them into 372 themes, with main results, companions and supporting material together, out of roughly 4,000 math problems given to the model. What remains unclear, and what the body frames as unsettled, is whether releasing so much at once counts as a contribution to science or shifts the job of finding out which results are right onto outside researchers.