In early August, I noted a landmark event for the entire artificial intelligence industry: OpenAI presented proofs for ten mathematical hypotheses that had remained open since 2016. This is not just another neural network success — it is a demonstration that AI is beginning to work at the forefront of fundamental science.
This concerns an internal version of the Astra model — the next generation of algorithms that the ChatGPT developer positions as a separate class alongside the Sol, Terra, and Luna lines. By my estimates, the volume of computation spent on finding solutions is equivalent to roughly $2000 at standard API rates for Sol. This is a negligible amount compared to years of labor by human mathematicians who unsuccessfully struggled with these problems.
The key feature of the approach is verification. People, together with the model, prepared manuscripts, after which Astra formalized each proof in Lean, a language for machine-checking theorems. The company has made all certificates and reasoning logs publicly available, allowing the scientific community to independently verify correctness.
Breakthrough results and unsolved "Millennium Problems"
Among the most significant achievements is a construction proving the existence of non-sofic groups. This closes a fundamental question that mathematicians had been unable to solve since 1999, when Mikhail Gromov introduced the concept of soficity. Astra also disproved Connes's rigidity conjecture about von Neumann algebras and solved the Ehrhart volume conjecture, which in itself is a world-class event.
I would particularly highlight Erdős problem No. 183 on multicolor Ramsey numbers — the model obtained new lower bounds for the complexity of computing the permanent using arithmetic circuits and proved a parallel repetition theorem for two-player quantum games. Additionally, Astra improved upper bounds on sphere packing density in high dimensions up to the Cohn–Elkies threshold.
The renowned mathematician Thomas Bloom from the University of Manchester has already called these results "big news" and a "significant step" in the field of constructions. However, I must add a fly in the ointment: the model did not handle the "Millennium Problems" — seven questions for which the Clay Mathematics Institute promises $1 million each. As reported by Noam Brown, co-author of Astra's reasoning technology, attempts to approach these problems proved unsuccessful.
Nevertheless, the very fact that AI systematically solves problems over which humanity has struggled for decades signals a paradigm shift. In July, we already saw Claude from Anthropic disprove a 1939 hypothesis. Now Astra is taking the next step. I expect that in the coming years we will witness the first AI solutions to the "Millennium Problems" — and it is only a matter of time, not capability.
My forecast: if OpenAI manages to maintain its pace, Astra or its successors will become indispensable tools in pure mathematics, and the human role will shift from solving problems to asking the right questions.