On August 1, OpenAI presented proofs for ten mathematical problems that had remained unsolved since 2016. This is not just another step in AI development, but a potential turning point in how we perceive the capabilities of neural networks in fundamental science.

A key role in this achievement was played by an internal version of the Astra model—the name ChatGPT's developer gave to its next major architecture. By my estimates, the amount of computation spent on finding solutions is equivalent to roughly $2,000 at current API rates for Sol. This is negligible compared to the years of work a human mathematician would have needed.

The methodology is of particular interest: the manuscripts were prepared by humans together with the model, after which Astra formalized each proof in Lean—a language for machine-checked theorem proving. The company has made all certificates and reasoning records publicly available, allowing the scientific community to verify the results.

A breakthrough in solving fundamental problems

Among the key results is a construction proving the existence of non-sofic groups. This closed a central question that mathematicians had been unable to solve since 1999, when Mikhail Gromov introduced the concept of soficity. Also impressive is the refutation of Connes's rigidity conjecture on von Neumann algebras and the solution of the Ehrhart volume conjecture.

I would particularly note Erdős problem No. 183 on multicolor Ramsey numbers—a classic of combinatorics that had resisted attempts at a solution for decades. Astra also obtained new lower bounds on the complexity of computing the permanent using arithmetic circuits and proved a theorem on parallel repetition for two-player quantum games. The model improved upper bounds on packing densities of spheres in high dimensions up to the Cohn–Elkies threshold and strengthened bounds for binary codes.

Renowned mathematician Thomas Bloom from the University of Manchester called these results "big news" and a "substantial step" in the field of constructions. However, not everything is smooth: the model failed to tackle the "Millennium Problems"—seven questions that the Clay Mathematics Institute named as the most important in mathematics, promising $1 million for solving each.

In my professional opinion, this breakthrough marks the beginning of a new era in mathematics, where AI becomes not just a tool but a full-fledged partner for the researcher. However, the fact that Astra did not conquer the Millennium Problems reminds us: we are still far from creating a universal mathematical intelligence. This is more of a powerful catalyst than a replacement for human genius.