OpenAI Publishes 722 AI-Generated Math Manuscripts as Mistral Previews 1T-Parameter Le Chonk

OpenAI Releases 722 AI-Generated Math Manuscripts

OpenAI has published a broad range of new mathematical results produced by an internal frontier model. The release includes 722 manuscripts organized into 372 families, covering 377 math problems. The company claims the model fully solved 90 of the top 500 open problems in mathematics, with partial progress on several Millennium Prize problems including the Riemann Hypothesis, the Hodge Conjecture, and the Birch and Swinnerton-Dyer Conjecture.

Among the highlighted results is a quasi-Riemann hypothesis showing that the zeta function has no zeros with real part above 7/8. The model also produced a matrix multiplication exponent of no more than 2.25, improving on the previous record of approximately 2.37. Other claims include a resolution of the Kakeya conjecture in R^3 and progress on the Hadwiger-Nelson problem, where the lower bound moved to 6 or 7. OpenAI stated that nearly all results came from a single prompt handed to a single AI agent, with the internal Pro model averaging about three hours per problem.

The company consulted with the independent Advisory Group on Mathematics and Artificial Intelligence at the Institute for Advanced Study. Several mathematicians on social media expressed astonishment at the results, though none have been independently confirmed by outside researchers.

Mistral Previews 1T-Parameter Le Chonk Model

Mistral launched a public preview of Mistral Large 4, nicknamed Le Chonk, a one-trillion-parameter model with 49 billion active parameters. The company claims it is the best open-weight model from the US or Europe on aggregated benchmarks, with state-of-the-art performance on cyber defense, manufacturing, and finance workloads.

Mistral said the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs at its European datacenters, including a cluster in Bruyères-le-Châtel. Much of the training data was multilingual. On a test asking models to reproduce a real vulnerability in open-source software and then patch it, ML4 scored 82%, which the company said was the highest of any model. The model also scored 62% on DeepSWE v1.1.

Open-weight release is scheduled for October 27. Mistral CEO Arthur Mensch said the model was trained and served on the company's own compute and that reinforcement learning shows no sign of saturation.

Sources: Techmeme