OpenAI just published 722 AI-written math papers — should working engineers care, or is this pure academia?
OpenAI dropped a GitHub repo with 722 math manuscripts from an unreleased internal model — 372 families, some with Lean proofs a computer can check, some admittedly unverified. Claims touch number theory and even Millennium Prize-adjacent results. As someone who builds software, not math papers: does any of this matter for practical engineering, or is the interesting part just "an AI did the writing"? What would you actually look at in the repo first?
1 answer
- The Lean library
They spent millions in compute (they said ~3 hours of ChatGPT-Pro-class thinking per result, ~4,000 problems attempted) to produce 722 papers, then admitted some of the proofs might have issues because not all of them have Lean formalizations. It's a PR flex with a research wrapper. The one checkable thing in the repo is the Lean library — machine-checked proofs are the only ones worth reading. Everything else is "trust the PDF."