OpenAI has made public hundreds of new mathematical findings generated by a cutting-edge internal model, representing one of the most advanced technologies developed by the company. In total, the organization published 722 manuscripts on GitHub, grouping 372 sets of results for problems that have challenged mathematicians for several decades.
This disclosure occurs amid intense debates about the application of artificial intelligence in the scientific field. Although the company consulted an independent collective of mathematicians to establish publication guidelines, it did not accept all suggested recommendations.
The compilation covers various areas, including number theory and algebraic geometry. Among the most notable points is a solution found for the Kakeya conjecture in four dimensions, in addition to improvements in computational algorithms and progress related to the so-called 'Riemann quasi-hypothesis'.
The results also encompass three of the five Millennium Prize Problems that still lack a solution. To achieve these results, the model attempted to solve approximately 4 thousand problems.
Scientific Community Reactions
The magnitude of this disclosure attracted the attention of the mathematical community. Alex Kontorovich, a professor at Rutgers University, commented on X that if a human had presented the proof regarding the 'Riemann quasi-hypothesis', it would be enough to guarantee him an immediate Fields Medal.
However, it is crucial to emphasize that the material still requires rigorous evaluation by experts. The mathematical relevance and validity of the demonstrations depend on independently conducted analyses.
To structure the publication, OpenAI sought assistance from the Advisory Group on Mathematics and Artificial Intelligence (AGMAI), composed of independent specialists. The company claims to have followed the group's guidelines on disseminating the findings but opted not to implement a specific recommendation to suspend testing of advanced mathematical problems in its proprietary models.
The repository contains formalizations of the proofs using Lean, a programming language that enables the verification of mathematical proofs through computational means. Summaries of the model's reasoning, estimates of computational resource consumption, and statistics related to the tested problems were also made available.
OpenAI announced that the GitHub publication includes protocols for both reviewing the articles and correctly citing the work. According to the company itself, each result required, on average, a computational power equivalent to about three hours of using ChatGPT Pro.
Additionally, the company plans to fund workshops, conferences, and programs focused on analyzing AI-generated discoveries. The repository will continue to be updated with new formalizations as they are obtained.
This disclosure follows a series of discussions about the use of AI in solving mathematical problems. A previous controversy involving OpenAI intensified the demand for greater transparency and reproducibility of the results.
Stance on Advanced Testing
The advisory group had suggested that laboratories should cease testing complex mathematical problems on their own models. OpenAI ignored this guidance, defending the importance of continuing the evaluation of its cutting-edge internal models in mathematics and other sciences.
Furthermore, the company did not disclose the exact computational times for each problem nor the prompts used. This omission reinforces, for the mathematical community, the need to meticulously examine the proofs before granting them definitive impact.
The next step falls to researchers, who must verify the demonstrations, replicate the results whenever feasible, and determine which ones constitute significant advances for mathematics.