OpenAI's mathematical AI system recently cracked an 80-year-old conjecture from legendary mathematician Paul Erdős, and researchers have now adapted that same approach to solve a second major problem that had stumped mathematicians for 50 years.
The breakthrough centers on a technique OpenAI's system discovered for tackling the cap set problem, a combinatorics challenge about finding patterns in multidimensional grids. Rather than relying on traditional mathematical reasoning, the AI identified a novel structural insight that compressed the problem into a more manageable form.
Mathematicians then applied this conceptual approach to another long-standing open problem in their field. The success demonstrates that AI systems trained on mathematical patterns can generate genuinely novel proof strategies, not merely execute existing ones.
This work extends recent progress in AI-assisted mathematics. DeepMind's AlphaProof system and similar tools have shown promise in competition mathematics, but solving decades-old research problems represents a different tier of difficulty. Professional mathematicians must verify any AI-generated insights, meaning the AI functions as a collaborator rather than an independent solver.
The technique's applicability to multiple problems suggests the underlying strategy captures something fundamental about the mathematical structure involved. Researchers can now examine why the AI's approach works and potentially extend it to related unsolved problems.
However, limitations remain. The AI's role is narrower than headlines sometimes suggest. It excels at pattern recognition within formal mathematical systems but cannot generate proofs from scratch without guidance. Human mathematicians still design the problem setup, interpret results, and determine whether solutions are rigorous.
The broader implication concerns the future of mathematics itself. If AI can consistently discover novel proof techniques, it may accelerate progress on problems that require fundamentally new ideas rather than incremental refinement of existing approaches. Yet mathematics also values understanding why theorems are true, not merely that they are. Whether AI-discovered