OpenAI has released 722 manuscripts on GitHub, claiming solutions and progress toward 372 major math problems. The results stem from an unreleased ChatGPT pioneer model, confirmed by OpenAI's advisory group. Key breakthroughs include a solution to the four-dimensional Kakeya conjecture and improvements to computer algorithms.
OpenAI states the average result required around three hours of ChatGPT Pro use. The release includes model reasoning, compute estimates, and prompt details. This follows OpenAI's previous announcement that models resolved over 100 long-standing open problems.
The company adhered to guidelines from its independent Advisory Group on Mathematics and Artificial Intelligence (AGMAI), ensuring traditional academic channel releases. However, OpenAI did not release specific compute times or exact prompts used. Nearly every paper resulted from a single prompt to a single AI agent.
Results remain subject to mathematician assessment. Following the Navier-Stokes imbroglio, skepticism persists. As Andrew Sutherland noted, until the model is released for replication, claims about solving problems with a single agent remain unverified.
Source: Engadget · Summarized by HeadlinesBriefing