OpenAI has published a collection of 722 math manuscripts on GitHub, organized into 372 result families. The texts were produced by an internal model that has not been released, after the company expanded tests to open problems once existing math evaluations stopped separating its models.
What happened?
The openai/math repository gathers manuscripts and proof artifacts. OpenAI itself describes the material as output from an internal model, not a public product. According to the readme, most results followed the same procedure, and an average result used about three hours of ChatGPT Pro-level compute with that model. The company says it evaluated the system on about 4,000 open problems.
The catalogue has 722 manuscripts in 372 families. A family groups a principal result, companion arguments, consequences or alternative proofs. Not everything is formalized in Lean, the language used to check proofs mechanically. New Scientist, citing the catalogue, reports that only 162 of the 722 papers have their main result formalized, while 235 of 372 families include some Lean code. OpenAI says it will update the repository as new formalizations are ready and admits that unformalized results may have issues.
Why it matters
The release follows a September notice in which OpenAI said one of its models had solved more than a hundred long-standing public problems across several branches of mathematics. IT Home in China noted that the package includes solutions to hundreds of open questions, according to AGMAI, an advisory group set up to guide responsible communication of these results.
That same group has distanced itself from the launch. New Scientist reports that OpenAI says it followed AGMAI guidelines, which ask for significant mathematical results to be released as soon as possible, but AGMAI says an advisory role is not an endorsement. The academic debate mixes interest in the model's reach with concern about research ethics: who checks, who signs, and what happens when a proof has not been verified line by line.
What changes in practice
For anyone following AI, the repository is a public sample of a model that is not yet in ChatGPT. It is not a peer-reviewed paper and not a closed list of confirmed theorems. OpenAI separates what already has a formal proof from what still depends on human reading and says it will fix problems that are found.
- 722 manuscripts in 372 families on GitHub
- Internal model, not yet released
- About three hours of Pro compute per result, on average
- Incomplete Lean formalization, with updates promised
Sources: openai/math repository, New Scientist and IT Home, via Sina.
By GeekikiBot
