OpenAI's unreleased model makes hundreds of math breakthroughs
What's the story
OpenAI has unveiled a massive trove of mathematical research produced by an unreleased frontier model. The collection consists of 722 manuscripts divided into 372 result families. According to the company, it includes 377 new mathematical results, although their importance and accuracy will have to be independently verified by experts in the field.
Validation challenges
Validating mathematical research can be time-consuming
The release of this collection comes after OpenAI announced in September that its internal system solved over 100 long-standing open problems in mathematics.
However, validating mathematical research can be time-consuming, especially when results are generated at a scale that makes traditional expert scrutiny difficult.
OpenAI itself acknowledges that the newly published papers will need further work, including revisions to citations and presentation.
Resource estimation
Questions about priority and credit
OpenAI estimates that the typical result in the collection required computing equivalent to around three hours of ChatGPT Pro-level reasoning.
However, this figure doesn't directly indicate how much human intervention was needed to turn individual discoveries into publishable mathematics.
The model is also understood to be the same unreleased system behind an earlier result concerning the Navier-Stokes equations, but its work overlapped with another group's research using AI tools, raising questions about priority and credit.
Addressing issues
Model not released yet
In response to concerns, OpenAI has published the papers on GitHub with guidelines for revisions and citations.
The company is also considering other community-hosted options that would comply with recommendations from the Advisory Group on Mathematics and Artificial Intelligence (AGMAI).
Despite these efforts, one major detail remains: the model itself hasn't been released, preventing researchers from independently reproducing its original reasoning process.