The model helped scientists by simplifying expressions and made a simple formula for this general case.
Then it spent nearly 12 hours reasoning the new problem, verifying the formula and «producing a formal proof.»
This was then reviewed and accepted by the research team, and incoroporated in the pre-review report at Arxiv.org. It will then go on to be peer reviewed and published.
— This preprint felt like a glimpse into the future of AI-assisted science, with physicists working hand-in-hand with AI to generate and validate new insights. There is no question that dialogue between physicists and LLMs can generate fundamentally new knowledge, Nathaniel Craig, Professor of Physics at the University of California, Santa Barbara tells OpenAI.
«Luminary» Maths exam
At the same time, OpenAI is claiming to have found solutions to many of the «First Proof» challenges, put forth by top tier mathematicians to supply a 10 question exam of brand-new problems that won’t be found in the AI model’s training data.
OpenAI now claims to have solved at least six of these problems, with an internal model with «limited human supervision.»
The problems are novel and difficult to prove, says Jakub Pachocki on x.com, saying they will publish the solution attempts at midnight pacific on Friday, which is also when the mathematicians answers decrypt and we will see if they were right.
Read more: OpenAI on results in physics and the physics paper. Jakub Pachockic claiming maths proofs, and more the «First Proof» challenge.