OpenAI Advances Mathematical Research with New Internal Model

Choi reports that OpenAI's new internal model has successfully addressed over 100 long-standing mathematical challenges. This feat was accomplished by 10,000 AI agents working collaboratively over a 24-day period.

OpenAI announced these mathematical achievements on September 21. Due to the rapid pace of progress, the company has established an independent mathematics advisory group to oversee the disclosure and utility of these AI-generated findings.

The research spanned 24 days, utilizing large-scale reinforcement learning on a pre-trained model. OpenAI continuously updated the agents with improved model versions throughout the process to optimize problem-solving efficiency.

Prior to this, OpenAI showcased Astra's capabilities in fields like group theory and quantum complexity. These earlier efforts involved researchers formalizing AI-generated proofs into Lean, costing approximately $2,000 in API tokens.

On September 9, OpenAI announced the resolution of the Navier-Stokes millennium problem. The model provided a counterexample demonstrating that fluid flow can develop singularities where velocity increases infinitely while energy remains finite.

The solution was not the work of a single model but a collective effort of 10,000 agents. Codex helped synthesize intermediate results, and the final proof was verified using the Lean formal verification tool.

Reports from The Information suggest OpenAI is nearing a solution to the Hodge conjecture. While unconfirmed, this indicates the model is simultaneously tackling diverse and complex mathematical domains.

While rumors mention a model called 'Bel,' OpenAI officially refers to the system only as a 'new internal model.' The company claims it significantly outperforms the previously known Astra model.

OpenAI is moving toward 'automated research interns' capable of performing complex tasks. This capability extends beyond math to include writing research code, designing experiments, and analyzing results.

Experts like Jay Cummings and Mark Kisin emphasize the importance of success rates over raw output. OpenAI's internal evaluations show the new model maintains higher pass rates than Astra under similar computational loads.

The new advisory group includes prominent mathematicians like Timothy Gowers and Edward Witten. They will provide independent guidance on academic standards and the ethical disclosure of AI-generated mathematical proofs.

With OpenAI DevDay approaching on September 29, there is interest in whether these research methodologies will be integrated into public-facing products. The collaborative agent approach could significantly expand developer capabilities.

Choi suggests the most significant impact will occur when these tools are available to external researchers. This could allow individual scientists to deploy thousands of agents to accelerate their own specific inquiries.

OpenAI has provided further details regarding their advisory group on mathematics and AI via their official website.