[ITmedia Business Online] OpenAI releases ``722 mathematical papers'' all at once The reason why the mathematics world strongly objected to ``AI companies' technology monopoly''

On October 6th, OpenAI in the US released 722 mathematical papers generated by its previously unpublished in-house AI model on GitHub. The papers are categorized into 372 groups based on related research results, and deal with unsolved problems in a wide range of fields, including number theory, geometry, and theoretical computer science.

A mysterious AI model that generates papers with one instruction

According to OpenAI's official announcement, this achievement was the result of giving the company's cutting-edge AI model approximately 4,000 unsolved math problems and attempting to answer them. On average, the computational resources used to obtain one result are equivalent to running ChatGPT Pro's advanced inference function for about three hours.

Published research includes results related to the Riemann hypothesis, one of the most famous unsolved problems in mathematics, a proof of the CM-type Abelian variety, which is part of the Hodge conjecture, and a paper claiming to solve the Kaketani conjecture in four-dimensional space. However, not all of the Riemann conjecture and Hodge conjecture have been solved.

This time, OpenAI also released formalized data for many of its papers using Lean, a programming language that allows computers to verify whether mathematical proofs are correct. However, not all papers have been formally verified, and some proofs may contain problems. The published paper has not been peer-reviewed by academic journals, and independent verification by the mathematical community is yet to come.

The name and detailed specifications of the AI ​​model that produced the results have not been made public. OpenAI explains that almost all of the results were achieved by giving instructions to a single AI agent once. Researchers are not yet able to reproduce the results using the same model.