OpenAI publishes 722 AI-generated math manuscripts, sparking ethics debate
Version 6: The report now covers OpenAI’s release of AI-generated mathematics results, a separate event from the previous version’s conference report.
In short
OpenAI has published 722 mathematics manuscripts, grouped into 372 result groups, produced by an unreleased internal frontier model, saying they solve or make substantive progress on hundreds of long-open problems. The release on GitHub includes Lean formal proofs for some results and compute estimates, with OpenAI saying an average result used compute equivalent to three hours of deep reasoning on ChatGPT Pro. An independent advisory group, AGMAI, called on AI labs to publish through academic channels and not to use math releases as model marketing, while mathematicians say the full impact will take time to assess.
Read the full story 2 min read
OpenAI published 722 mathematics manuscripts, grouped into 372 result groups, produced by an unreleased internal frontier model, according to IT Home and Huxiu. The manuscripts were released via GitHub with revision and citation protocols and include Lean formal proofs for some results, according to CLS, 36Kr and Huxiu. OpenAI said the results either solve long-standing open problems or achieve substantive breakthroughs, cnBeta reported. [ 4 , 5 , 6 , 7 , 8 ]
Huxiu reported that the manuscripts involve the quasi-Riemann hypothesis, the unique games conjecture, special cases of the Hodge conjecture and the free group factor problem. It said the quasi-Riemann work proves no zeros in the region with real part greater than 7/8 but does not solve the full Riemann hypothesis, and that the Hodge work addresses the rational Hodge conjecture on CM abelian varieties, not the full conjecture. [ 7 ]
OpenAI said an average result consumed compute roughly equal to three hours of deep reasoning on ChatGPT Pro, according to CLS, IT Home and 36Kr. Huxiu reported that 162 manuscripts’ main conclusions have Lean formalization material, but the formalization scope is limited and does not cover all conclusions. [ 4 , 5 , 6 , 7 ]
IT Home reported that AGMAI, an independent advisory group of mathematicians, said the release contains answers to “hundreds” of unsolved problems. AGMAI called on AI labs to publish math results through established academic channels and disclose model names, prompts and compute overhead, and urged companies not to use math releases as model marketing. [ 5 ]
IT Home reported that OpenAI had said in September its model solved more than 100 long-unresolved open problems, covering most branches of mathematics, without providing proofs. Huxiu reported that OpenAI convened about 40 mathematicians in August, announced a claimed Navier-Stokes equation solution in September, and that 25 Fields Medal winners published an open letter criticizing AI companies for ignoring understanding and talent cultivation. [ 5 , 7 ]
IT Home said mathematicians still need to evaluate and digest the results, so the full impact may take time to appear. Huxiu reported that OpenAI put the manuscripts on GitHub, set revision and citation norms, published reasoning summaries and promised to fund workshops, but the generative model remains closed, so mathematicians cannot yet use the same tool to continue research. [ 5 , 7 ]
Why it matters
The release could accelerate mathematical research if the results withstand scrutiny, and it intensifies a debate about AI’s role in mathematics, research ethics and credit for human mathematicians. Because the model behind the results remains unreleased, mathematicians cannot yet use the same tool to continue the work, leaving the practical impact uncertain.
Key facts
- OpenAI published 722 mathematics manuscripts, grouped into 372 result groups, produced by an unreleased internal frontier model, according to IT Home and Huxiu. [ 5 , 7 ]
- The release includes Lean formal proofs for some results, allowing proofs to be computer-verified, according to CLS, 36Kr and Huxiu. [ 4 , 6 , 7 ]
- OpenAI said an average result consumed compute roughly equal to three hours of deep reasoning on ChatGPT Pro, according to CLS, IT Home and 36Kr. [ 4 , 5 , 6 ]
- AGMAI, an independent advisory group, said the release contains answers to “hundreds” of unsolved problems and called on AI labs to disclose model names, prompts and compute overhead, according to IT Home. [ 5 ]
- Huxiu reported the manuscripts touch on the quasi-Riemann hypothesis, the unique games conjecture, special cases of the Hodge conjecture and the free group factor problem, while full Riemann and Hodge conjectures remain unsolved. [ 7 ]
- OpenAI had said in September that its model solved more than 100 long-unresolved open problems, according to IT Home and Huxiu. [ 5 , 7 ]
- OpenAI put the manuscripts on GitHub with revision and citation norms and promised to fund workshops, but the generative model remains closed, Huxiu reported. [ 7 ]
Confirmed by several sources
- OpenAI published a batch of new mathematics research results generated by an internal frontier model. [ 4 , 5 , 6 , 7 , 8 ]
- The release includes Lean formal proofs for some results, allowing proofs to be computer-verified. [ 4 , 6 , 7 ]
- OpenAI said an average result consumed compute roughly equivalent to three hours of deep reasoning on ChatGPT Pro. [ 4 , 5 , 6 ]
- OpenAI said the results either solve long-standing open problems or achieve substantive breakthroughs, covering hundreds of unsolved problems. [ 5 , 7 , 8 ]
- The content was published via GitHub with paper revision and citation protocols. [ 4 , 6 , 7 ]
- OpenAI had said in September that its model solved more than 100 long-unresolved open problems. [ 5 , 7 ]
Still unclear
- Which specific problems were solved and the full list of results. OpenAI did not say which problems it solved before the release, and the documents only give examples such as the quasi-Riemann hypothesis and unique games conjecture.
- Whether the results are mathematically correct. IT Home reported that mathematicians still need to evaluate and digest the results, so the full impact may take time to appear.
- Whether the unreleased model will be made available to mathematicians. Huxiu reported that the generative model remains closed, so mathematicians cannot use the same tool to continue research.
- The exact time of the release. cnBeta says OpenAI published on local time October 6, while other documents were published on October 7 in China and do not give an exact release hour.
- The status of OpenAI’s September claim about solving a Navier-Stokes equation problem. Huxiu mentions it as a claim and reports criticism from Fields Medal winners, but the documents do not say whether it was verified.
- The location of the announcement. The documents do not state where OpenAI published the results.
What local media are saying
Timeline, local time
- OpenAI publishes a batch of new mathematics research results generated by an internal frontier model, according to cnBeta. [ 8 ]
- CLS reports OpenAI is publishing the results via GitHub with Lean formal proofs and compute estimates. [ 4 ]
- IT Home reports 722 manuscripts covering 372 result families and AGMAI’s call for academic publication and disclosure. [ 5 ]
- 36Kr relays the GitHub publication, Lean proofs and compute cost. [ 6 ]
- Huxiu details specific mathematical problems, limited Lean formalization and the ethics debate. [ 7 ]
- cnBeta reports OpenAI published 372 result groups covering long-standing unsolved problems. [ 8 ]