OpenAI internal frontier model math release

by Vojtech: OpenAI

The volume of new mathematics produced in one week exceeds typical annual departmental output, and it came from an unreleased internal frontier model OpenAI just shared. We are looking at 722 mathematical manuscripts covering pure math, theoretical computer science, and mathematical physics that contain 372 families of new results. The model processed approximately 4,000 problems, with each result requiring an average of three hours of ChatGPT Pro compute time. Reports indicate the system solved over 100 open problems, including the Navier-Stokes equation, which is identified as a Millennium Prize problem. Some proofs include Lean formalization for machine verification, and the release was reviewed by the Institute for Advanced Study's math advisory group. There is a major catch though: the model remains unreleased, preventing external reproduction. Not all 722 manuscripts have been verified, and unformalized versions may contain errors. Major claims circulating about the work lack independent confirmation. It is a massive technical development, but the verification status varies between formalized and unformalized proofs. #OpenAI #ArtificialIntelligence