On October 7th, OpenAI inadvertently published a new repository, dubbed openai/math, on GitHub. This repository boasts 722 mathematical manuscripts and 372 sets of achievements, all generated by its internal, yet-to-be-published models. These accomplishments span a diverse array of mathematical domains, encompassing number theory and algebraic geometry. Had these papers been uploaded manually, it is estimated that the endeavor would have spanned three decades. The model curated these significant achievements by referencing roughly 4,000 unresolved open problems within the mathematical community. On average, it required merely about 3 hours of ChatGPT Pro-level computational resources to generate each set of achievements. Among these milestones are major breakthroughs, such as progressing the quasi-Riemann hypothesis to 7/8, reducing the exponent of complex field matrix multiplication to 2.25, and resolving the four-dimensional version of the Kakeya conjecture. In the past, any one of these feats would have sufficed to captivate the attention of the entire mathematical community.

The article also elucidates the distinctive facets of mathematical proofs and the utilization of the Lean formal verification tool in mathematical research. It traces the evolution from early machine proofs to the burgeoning involvement of AI in cutting-edge mathematical exploration in recent years. Furthermore, the article references the collective stance of 25 Fields Medalists, who contend that AI companies employing problem-solving as a metric for model advancement is incongruent with the essence of mathematics, which prizes human insight. Consequently, they advocate for granting the academic community additional time to assimilate these advancements. However, OpenAI proceeded to release these achievements without heeding these appeals, thereby igniting a pertinent debate.

Drawing a parallel to the past, when AlphaGo's triumph over human chess players gradually transformed related perceptions into common knowledge, the article posits that in the future, various disciplines will witness accelerated development in accordance with the cost of acquiring 'evaluators.' Mathematics and code are poised to be the first to experience explosive growth, followed by simulatable fields, while hard sciences reliant on physical experiments and fields pertaining to humans, such as medicine, will reap the benefits last. Ultimately, humans may emerge as the slowest link in the research chain.

Finally, integrating elements of mathematical history, such as Hilbert's 23 problems and Gödel's incompleteness theorems, the article posits that in the AI era, humans still uphold three core values: posing questions, comprehending, and infusing meaning.