OpenAI’s Astra AI model claims new breakthroughs in pure mathematics and theoretical computer science
Sebastien Bubeck of OpenAI says that non-sofic groups exist, describing the result as one of “many new beautiful results” proved by Astra, OpenAI’s next major artificial intelligence model. OpenAI has published a webpage and PDF outlining 10 mathematical and theoretical computer science advances generated by an internal version of Astra.
According to OpenAI, the total computational cost of producing the results was approximately $2,000 at Sol API rates. Human researchers subsequently developed the model-generated ideas into formal manuscripts with assistance from the AI system. The proofs reportedly include machine-checkable Lean certificates and detailed chain-of-thought walkthroughs.
AI model Astra proves the existence of non-sofic groups
A group is called sofic when finite portions of its multiplication table can be approximated, in a precise mathematical sense, by permutations of finite sets. The concept was introduced by Gromov around 1999, while the term “sofic” was later coined by Weiss.
Sofic groups generalize two important classes of groups: amenable groups and residually finite groups. Gromov asked whether every countable discrete group is sofic. That problem remained unresolved for approximately 27 years.
yes, nonsofic groups exist: this statement is one of many new beautiful results proved by Astra, our next major model.
We’re releasing 10 such Astra proofs, complete with lean certificates and CoT walkthroughs for each of them. The results are wide-ranging, from von Neumann…
— Sebastien Bubeck (@SebastienBubeck) August 1, 2026
Connes rigidity and von Neumann algebras
Another reported result concerns Connes rigidity, a major topic in the study of von Neumann algebras and II1 factors. The counterexample indicates that the mapping from groups to factors is not injective, even within this highly rigid mathematical setting.
The result could deepen researchers’ understanding of how much algebraic information a von Neumann algebra preserves. It also relates to Popa’s deformation and rigidity program, an influential area of modern operator-algebra research.
Applications across mathematics and computer science
OpenAI says the other Astra results address problems involving packing and coding, circuits, lattices, Ramsey theory, and extremal graph theory. Several of these results reportedly provide sharper quantitative bounds or resolve specific questions connected with the Erdős tradition of combinatorial mathematics.
Potential applications extend across coding theory, computational complexity, discrete geometry, combinatorial number theory, cryptography, analytic number theory, and the modular bootstrap. Improved packing bounds, stronger hardness results, and new combinatorial constructions could provide useful tools for researchers in both pure and applied mathematics.
AI-generated proofs with formal verification
The significance of the Astra project is that it reportedly produced research-level mathematical results rather than solutions to standard contest problems. The use of Lean certificates adds a formal verification layer, allowing portions of the arguments to be checked by computer.
If independently validated, these results would demonstrate how advanced AI systems can assist with novel proof discovery across distant mathematical fields. Such systems could accelerate research by proposing conjectures, identifying counterexamples, developing proof strategies, and generating formal arguments for human mathematicians to review and refine.
Source: www.nextbigfuture.com