OpenAI releases 722 research documents showing its AI model solved 372 Maths problems
OpenAI has unveiled a collection of 722 research manuscripts covering a wide range of mathematical fields. Among these documents are pivotal findings linked to the Millennium Prize Problems, notably the Riemann hypothesis. While some experts remain doubtful about the validity of machine-generated mathematics, OpenAI is actively seeking feedback from an advisory panel to refine their approach.
Just weeks following the controversy surrounding their claim to have solved a 90-year-old Millennium Prize Problem, OpenAI has now released a repository of 722 research manuscripts. These documents, divided into 372 distinct problem families, span a wide range of mathematical disciplines, including number theory and algebraic geometry. The release includes significant advancements tied to three out of the five remaining Millennium Prize Problems, notably the Riemann hypothesis.
The move comes after a period of skepticism from the mathematical community, with many expressing concerns over the potential implications of machine-generated mathematics on the field. In response, OpenAI has collaborated with an advisory group on mathematics and AI, based at the Institute for Advanced Study, to ensure the responsible dissemination of these results. The company emphasized that they have taken the independent advice and public recommendations of this group into account when releasing the materials.
Inside the 700-page repository, OpenAI has organized the material into 722 manuscripts across 372 families. Each family group encompasses primary findings, along with companion arguments, corollaries, and alternative proof strategies. OpenAI has also acknowledged that not all papers have undergone complete verification or paired with computer-checked formal code.
To address this, the company is publishing the formalisations in Lean, a programming language designed to enable computers to verify proof logic, and has committed to adding more in the future.
OpenAI admitted that some of the unformalised results may contain errors, and they have pledged to address identified flaws promptly while exploring community-hosted repositories. To foster scientific openness, the lab shared technical data outlining the resources required to generate the proofs. The average mathematical result within the repository consumed compute equivalent to approximately three hours of ChatGPT Pro's thinking time.
Additionally, OpenAI published ten breakdowns of the model's internal reasoning chains, estimates of total compute expenditures, and aggregate statistics tracking the total volume of attempted problems.
Rewritten by urgent.news; may contain errors — check the original.