Terence Tao says AI models have produced proofs of hundreds of open problems with nobody to answer mathematicians' questions about them, so understanding them falls to volunteer labour; some junior researchers' projects were made obsolete, and some now keep their research directions to themselves for fear someone will feed them to an AI first.
The proofs arrived, the explanations did not
Tao calls the models 800-pound gorillas "feasting on our open problems". Maths suits AI, he says, because an answer can be checked objectively on a computer; but the problems themselves are mostly lighthouses, and what matters is the understanding, collaboration and partial results gathered on the way. A person who solves a major problem gives talks and answers calls; for the hundreds of papers OpenAI released at once, he says, "there's nobody at OpenAI I can call", so mathematicians must work through them on their own, far more slowly than usual. The record already holds OpenAI's 722 manuscripts of 6 October and the three withdrawn a day later, and his 10 October "Math 2.0" lecture arguing that proofs are now plentiful and understanding is the goal; what the interview adds is his verdict on how the results were released: "at the very least, very inconsiderate".
What it does to junior researchers
He says many junior people's projects were made obsolete by this release and they are scrambling for new ones; some have become secretive about their research directions, which the field normally shares, for fear that mentioning one gets it fed to an AI and solved first. Priority, he argues, is degenerating into who can feed a problem to an AI fastest, so prizes and promotions should weigh being first less and reward collaboration. These are his observations and judgments; the episode gives no numbers or named cases.
The tools sit with a few companies
Tao says the technology that solves these open problems sits with only a few tech companies, often as internal models, while a practising mathematician may have access only to a much cheaper and less capable one. He did not mention the SAIR Foundation he is part of; the record holds its September preview of an open maths model governed by the mathematical community, which answers this gap but is so far only a call for interest.