OpenAI Model Produces (Dis)proof of Unit Distance Conjecture

Marginal Revolution reports that Tim Gowers wrote a general-purpose AI model from OpenAI has produced a (dis)proof of the unit distance problem, a well-known Erdos question. Marginal Revolution quotes Gowers saying, "AI has now solved a major open problem, one of the best known Erdos problems called the unit distance problem, one of Erdos's favourite questions and one that many mathematicians had tried." The site also highlights commentary from Thomas Bloom, who argues the AI succeeded by combining patient search with familiarity across disparate areas of number theory and class field theory. The reporting frames the result as surprising and technically interesting, while noting it may not supply the new geometric tools mathematicians had hoped for.
What happened
Marginal Revolution quotes Tim Gowers reporting that a general-purpose AI model from OpenAI has produced a (dis)proof of the unit distance problem, a long-standing Erd o s question in combinatorial geometry. Marginal Revolution reproduces Gowers' line, "AI has now solved a major open problem, one of the best known Erd o s problems called the unit distance problem, one of Erd o s's favourite questions and one that many mathematicians had tried." The blog also cites Thomas Bloom's comments on the construction and its provenance.
Technical details
Thomas Bloom, quoted in the Marginal Revolution post, describes the AI's construction as relying on a confluence of conditions: deep, patient exploration of the problem space; a willingness to generalize to other number fields; and familiarity with parts of class field theory. Bloom writes that the AI's approach uses towers of number fields and existing theory to assemble a counterexample construction, and that the result does not introduce powerful new geometric machinery that a classical proof of the conjecture might have required.
Context and significance
Editorial analysis: This episode is a concrete example of a general-purpose model producing a mathematically nontrivial, publishable-level result by combining techniques across subfields. Industry observers have noted similar patterns where large models surface surprising constructions by exhaustively exploring search spaces and recombining known results. For practitioners, the event sharpens questions about reproducibility, proof verification, and how to integrate model-generated mathematics into peer review and formal verification pipelines.
What to watch
For practitioners: verification efforts and artifacts. Monitor whether the construction is (a) reproduced independently, (b) accompanied by formalized proofs or mechanized checkable artifacts, and (c) published with full prompts, model metadata, and derivations. Also watch academic responses assessing whether the result yields new methods or primarily a novel combination of existing theory.
Scoring Rationale
A general-purpose model producing a (dis)proof of a famous Erdos problem is a notable milestone for model capabilities and research workflows. It materially affects verification, formal proof practices, and expectations for model creativity, making it highly relevant for AI/ML researchers and practitioners.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems


