Tech • IA • Crypto
How a reasoning model cracked an 80-year-old math problem — the OpenAI Podcast Ep. 20
9/10TL;DR
An OpenAI reasoning model has produced a credible disproof of a major 80-year-old Erdős conjecture, signaling rapid advances in AI-driven mathematical research.
KEY POINTS
Breakthrough on an Erdős Conjecture
An advanced OpenAI reasoning model generated a disproof of the unit distance conjecture, a long-standing problem in combinatorial geometry posed by Paul Erdős. The problem examines how many pairs of points on a plane can be exactly one unit apart and how that number scales. The model’s result challenges Erdős’s original assumption that a square grid arrangement was near optimal.
A Major Open Problem with Real Stakes
The conjecture is considered a central question in discrete geometry and carried a historical prize of $500, reflecting its difficulty and importance. Unlike more obscure Erdős problems, this one has been widely studied in academic literature, making its resolution particularly significant within the field.
Unexpected Speed of Progress
AI systems were struggling with basic math as recently as late 2023, yet within months reached International Math Olympiad (IMO) gold-level performance. Researchers had expected such milestones closer to 2026, but progress accelerated dramatically, with current models now operating far beyond Olympiad-level reasoning.
General-Purpose Model, Not Specialized
The system that solved the conjecture was not trained exclusively for mathematics. It is a general-purpose reasoning model capable of coding, web browsing, and executing programs. Its mathematical success emerged from broader improvements in reasoning rather than domain-specific tuning.
Inference-Time Compute Drives Capability
A key innovation is test-time compute, allowing models to “think longer” before answering. Performance improves as more compute is allocated, with accuracy on difficult problems rising significantly and reaching roughly 50% success rates under higher compute budgets.
Creative Cross-Disciplinary Reasoning
The solution involved linking class field theory with combinatorial geometry, a connection rarely exploited in prior work. This kind of cross-domain synthesis suggests models can uncover non-obvious relationships that require both creativity and technical precision.
Human Verification and Initial Skepticism
Early reactions from mathematicians were skeptical, given the problem’s difficulty. After extended review failed to identify errors, confidence grew that the proof may be correct. The validation process involved multiple experts carefully examining the result over several days.
Immediate Impact on Further Research
Within a week, mathematicians used ideas from the AI-generated proof to tackle related problems, including disproofs in adjacent domains. This demonstrates that AI outputs can act as catalysts for new human-led discoveries rather than isolated achievements.
Changing Nature of Mathematical Work
Researchers report a shift in workflow, with AI systems handling substantial portions of coding and problem exploration. This has increased productivity and expanded the range of problems individuals can attempt, often enabling parallel exploration of multiple ideas.
Limits Remain for Deep Theory Creation
Despite progress, models are not yet capable of developing entirely new mathematical theories or solving problems like P vs NP, which may require decades of conceptual innovation. Current strengths lie in problem-solving and connecting existing ideas rather than foundational theory-building.
Broader Implications for Science and Security
The same reasoning advances could impact fields such as cryptography, where AI might validate or challenge assumptions about computational hardness. Researchers also expect applications across physics and other sciences, where models can analyze data and propose new hypotheses.
CONCLUSION
The disproof of a major Erdős conjecture by a general-purpose AI marks a turning point in mathematical research, highlighting both the accelerating pace of AI capability and its emerging role as a collaborative tool for human discovery.