OpenAI just announced a major milestone in AI-driven mathematics: an internal general-purpose reasoning model has disproved a long-held belief connected to Paul Erdős’ famous 1946 unit distance problem.
The problem asks a simple but deeply difficult question: if you place dots on a plane, how many same-length connections can you draw between them? For decades, mathematicians believed grid-like arrangements were essentially the best possible answer.
OpenAI says its model found a new family of constructions that performs better, using ideas from algebraic number theory rather than the usual geometric intuition. The result was reviewed by outside mathematicians, including leading experts in the field.
What makes this especially important is that the model was not a math-specialized system like AlphaProof. It was a general reasoning model, suggesting that frontier AI may be moving from solving prepared benchmarks toward making original contributions.
OpenAI had previously walked back claims around GPT-5 and Erdős problems, where the model had surfaced existing literature rather than creating new discoveries. This announcement is different because it claims a genuinely new proof, externally checked by mathematicians.
Why it matters: math may be one of the clearest early signals of where AI is heading. If a general-purpose model can challenge an 80-year-old mathematical assumption with a novel construction, then we may be seeing the early shape of “Level 4” AI: systems that do not just assist experts, but begin contributing new knowledge across disciplines.
https://openai.com/index/model-disproves-discrete-geometry-conjecture
Remove from favorites
