OpenAI publishes 722 manuscripts of AI-generated mathematics results from unreleased model

Left · 1OpenAI has published 722 manuscripts, grouped into 372 result families, containing solutions to long-standing mathematics problems produced by an unreleased frontier model. 1 Left · 1Scientific American reported that OpenAI posted the results to a GitHub repository at 6 P.M. EDT on 6 October, and that the company says each result resolves or makes substantial progress on a major open question in mathematics or theoretical computer science. 2 Left · 1Among the claimed results are a solution to the four-dimensional Kakeya conjecture, improvements to important computer algorithms, and progress toward the Riemann hypothesis. 2

Left · 2The release follows an earlier OpenAI result on the Navier-Stokes problem, which Scientific American says came from a 10,000-agent swarm costing millions of dollars in computing; in September OpenAI said its model had resolved more than 100 long-standing open problems. 2 1 Left · 2An OpenAI spokesperson told Scientific American that the model produced almost every new result from a single prompt given to a single AI agent, though some may have taken multiple attempts; OpenAI separately says the average result used the equivalent of three hours of ChatGPT Pro thinking. 2 1 Left · 1MIT mathematician Andrew Sutherland said claims of single-agent results should be treated as unverified until the model is released and results can be replicated. 2

Left · 1Many results have already been verified in Lean, a proof-checking language, according to Scientific American, but mathematicians will need months to judge whether the proofs contain novel and important ideas. 2 Left · 2The Advisory Group on Mathematics and Artificial Intelligence (AGMAI), a newly formed independent group of mathematicians announced by OpenAI on 21 September, published recommendations in late September urging labs to release results promptly through established academic channels and to disclose the model, prompts and compute costs. 1 2 Left · 2The group also urged companies to avoid treating releases as marketing vehicles and criticised the use of proprietary internal models that mathematicians cannot access. 1 2

Left · 1Scientific American reports OpenAI is disclosing only average compute time and some statistics, with no prompts; the company says it takes the guidelines seriously and is not bound by them, and is working to release the model as quickly and responsibly as possible. 2 Left · 1OpenAI said it is publishing in a GitHub repository with protocols for paper revisions and citations while exploring community-hosted alternatives that meet the committee's guidelines. 1 Left · 1Mathematicians are divided: Daniel Litt of the University of Toronto sees no reason for results to be kept secret and expects the release to be good for mathematics, while Terence Tao has criticised frontier labs for the “insane” pace of AI-generated results. 2

Left · 1OpenAI's spokesperson said many of the new results are not yet understood by its own mathematicians, and the company says it will not slow down because such problems are a test of whether its AI is improving. 2 Left · 1The Verge notes that rivals such as Anthropic are also producing results, including on a Millennium Prize problem, and that the pace has provoked debate over research ethics and how companies credit the human mathematicians whose work they build on. 1