OpenAI releases 249-page paper: Cutting-edge inference model overcomes ten major mathematical problems in a single shot

OpenAI released a 249-page research collection on July 31, unveiling ten advances in its cutting-edge inference models in the direction of advanced mathematics and theoretical computer science. The most eye-catching one is that the model solved a famous open mathematical problem that had been unsolved before in a "single attempt" method, without the need for multiple retries and manual intervention. From "being able to solve problems" to "being able to independently advance open problems", this is perhaps the most tangible demonstration of the reasoning model moving towards "autonomous mathematical research".

Ten Progresses and "Single Shot" Overcome

The collection covers the two major directions of advanced mathematics and theoretical computer science, and contains a total of ten specific developments. One of the landmark events was that the model solved a famous open mathematical problem with only one inference - traditional paradigms often rely on multiple attempts or human assistance, and "single shot" means that the model has pushed the inference quality itself to the point where it can be directly implemented. All relevant results are open sourced in the form of Lean formal proofs, and come with self-written papers on the models, which are verifiable and reproducible.

Cost account of USD 2000/question

Officially, the computational cost to solve each mathematical problem is about $2,000. Comparing this cost account with the time and financial investment of a team of human mathematicians, the difference in order of magnitude is very intuitive: