Particle.news
Download on the App Store
15 ARTICLES
·
3w ago

OpenAI Releases 722 AI‑Generated Mathematics Manuscripts

Mathematicians must now inspect a large public dump of claimed proofs for errors because the unreleased model that produced them remains closed.

Image
Image
ULCA mathematician Terence Tao is among the prominent members of his field who have decried the way OpenAI and other AI companies have targeted longstanding mathematics problems as a way to test their new AI models and score marketing points. He worries that the effort has produced solutions without insights that would advance the field's overall understanding. To survive, Tao thinks the field will now have to focus on skills other than problem-solving.
Illustration: Emil Lendof/WSJ, Getty Images

Overview

  • OpenAI posted 722 AI‑generated manuscripts on GitHub on Oct. 6, organized into 372 families and drawn from roughly 4,000 problems the company fed its internal model.
  • The corpus includes claims linked to major open questions across number theory, algebraic geometry and theoretical computer science, and cites advances related to variants of several Millennium Prize problems.
  • OpenAI provided many Lean formalizations that check logical steps, but it warns many Lean files were written by AI and are marked 'unchecked', so the checks do not replace human peer review.
  • Mathematicians have begun manual verification but progress is slow because the producing model is proprietary, some results may contain errors, and some researchers say their unpublished work may have been used without clear attribution.
  • OpenAI says it consulted an Institute for Advanced Study advisory group, will publish corrections and fund workshops, and plans a responsible model release timeline that it has not yet fixed.