Research note
Research note: OpenAI publishes about 720 maths manuscripts written by an unreleased AI model
How we counted the manuscripts, why our number differs slightly from coverage, and what the Lean proofs do and do not prove.
What we set out to check
What OpenAI released, how many manuscripts there are, how they were checked, and whether anyone outside OpenAI has reviewed them.
What we found
- OpenAI’s post confirms the release and the compute used, but gives no count.
- The GitHub README is the primary record of what was released.
- Two independent outlets describe the release in the same way.
Numbers that differ
Coverage says 722 manuscripts. The README listed 719 when we checked on 11 October. We do not know why; a repository can change after release. We write “about 720” and show both numbers.
Labels we chose and why
- Confirmed for facts in OpenAI’s post or its own repository.
- We did not label any result as correct or new. Lean proofs are machine-checked, but that does not settle whether a paper’s claims and framing are right.
What we could not verify
- Whether any manuscript has been peer reviewed.
- How mathematicians will judge the importance of the results.
What we searched
- OpenAI 722 math manuscripts GitHub
- openai.com mathematical results internal frontier model Lean
- Advisory Group on Mathematics and Artificial Intelligence recommendations
Every source we read
| Source | Used? | Why |
|---|---|---|
| OpenAI: Sharing AI progress in mathematics | USED | Official post. Used for the release, compute and advisory group. It gives no manuscript count. |
| GitHub: openai/math | USED | Primary record. README listed 719 manuscripts, 372 families and about 42% Lean coverage when checked. |
| Interesting Engineering | USED | Independent coverage. Reports 722 manuscripts and explains what a Lean check covers. |
| Unite.AI | USED | Independent coverage. Used for the advisory group’s recommendations. |
| OpenAI Developer Community thread | NOT USED | Posted by a community member, not OpenAI. Useful background on independent re-checks, but not a primary source. |