OpenAI's Astra Solves 10 Major Math Problems, But Raises Deep Concerns

On August 1, OpenAI published a blog post claiming that an internal version of Astra, its next-generation language model, has solved or decisively advanced ten open problems in mathematics and theoretical computer science. The problems span advanced topics like high-dimensional geometry, coding theory, arithmetic circuit complexity, group theory, operator algebras, quantum complexity, lattice-based cryptography, and extremal combinatorics. OpenAI said the entire feat cost about $2,000 in tokens per problem, with results certified using the Lean formal proof assistant. The announcement sparked both excitement and concern within the mathematical community. Researcher Jay Cummins noted that for the first time, an AI solved a problem he had personally worked on for years, making the defeat hard to accept. Investor Nicolas Bustamante questioned the future of the Fields Medal, asking who gets credit when problems are a prompt away. Notably, Fields Medalist Jacob Tsimerman paused his chair at the University of Toronto to join OpenAI in late July, focusing not on math capabilities but on AI safety. Terence Tao, another Fields Medalist, warned about a dangerous shift: with AI automated proof generation, mathematics moves from scarcity to abundance, but researchers cannot keep up with verification and 'digestion' of proofs. He compared raw AI proofs to 'mystery animal carcasses' dumped at a community potluck, where no one has time to clean and cook them. Tsimerman also lamented the rise of 'mathematical slop' from users with no expertise submitting raw ChatGPT outputs to labs. Both warned that formal solutions to famous conjectures could kill funding and interest in those topics, even though no human understands the solutions, leading to a discipline that progresses without understanding.