OpenAI releases a range of new mathematical results produced by an internal model, with details like estimations of compute spent in terms of ChatGPT Pro usage

AI models can now produce novel mathematical proofs, potentially accelerating scientific discovery. This output represents a new capability for AI beyond code generation.

OpenAI has released a collection of mathematical manuscripts and proof artifacts generated by an internal AI model. The repository, hosted on GitHub, contains 722 manuscripts organized into 372 families, classified by mathematical discipline. These outputs were produced using an unreleased internal OpenAI model, with an average of three hours of ChatGPT Pro compute time dedicated per result. The model was tasked with approximately 4,000 problems during the evaluation period. While most results followed a standardized procedure, exceptions include work on the Riemann zeta function's zero-free region and the Hodge Conjecture for CM abelian varieties. Some manuscripts include Lean formalizations, with plans to add more, though not all unformalized results are guaranteed to be error-free. OpenAI intends to update the repository with new findings and formalizations as they become available.

OpenAI's release of AI-generated mathematical proofs signals a significant leap in AI's capacity for abstract reasoning and scientific discovery. This development suggests that AI models are moving beyond pattern recognition and code generation into complex problem-solving that was previously the exclusive domain of human experts. The quantified compute cost, framed in terms of ChatGPT Pro usage, provides a tangible metric for the resources required to achieve these advanced reasoning capabilities, offering insights into the economics of frontier AI development.

The availability of these AI-generated proofs, especially those with formal verification in systems like Lean, could revolutionize mathematical research by providing new avenues for exploration and validation. It also raises questions about the role of human mathematicians in the future, potentially shifting their focus from generating proofs to guiding AI, verifying AI outputs, and exploring the novel mathematical landscapes revealed by these models. This could lead to a more collaborative approach between humans and AI in advancing scientific knowledge.