Mathematical AI Must Separate Discovery From Retrieval
A model can find an unexpected link between mathematical fields, but novelty depends on proving it hasn't merely recovered an overlooked result.
Developed from a conversation between Pete Winn, Anthony and Andy David
OpenAI once claimed that a model had solved ten problems associated with mathematician Paul Erdős. Within 24 hours, the company withdrew the claim. The model had performed a sophisticated literature review and found that mathematicians had already solved the problems elsewhere, although those results hadn’t been formally recognised as answers to the Erdős problems. The work was useful, but it wasn’t novel.
Anthony described a reasoning model producing a new solution to a famous outstanding Erdős problem, with mathematicians including Fields medallists accepting it as legitimate. From his understanding, the key move was to apply insights from one branch of mathematics to a combinatorial problem in another. The model then developed that cross-field connection through the logical steps needed for a sound argument.
The contrast identifies both the promise and the test for mathematical AI. Searching across fields can reveal connections that specialists working within one framing may miss. Yet a plausible proof and a novel proof are separate achievements. Each result must be checked for mathematical validity and searched against the literature under other names or formulations. Otherwise, retrieval can look like invention.