OpenAI model disproves Erd\u000151s planar unit distance conjecture

OpenAI wrote in a May 20 blog post that an internal OpenAI model produced a proof that disproves the planar unit distance problem, a conjecture posed by Paul Erd\u000151s in 1946 (OpenAI blog). Per OpenAI, the model constructed an infinite family of planar point sets that yield a polynomial improvement in the number of unit-distance pairs over the classical square-grid lower bound; OpenAI stated that external mathematicians have checked the argument and that a companion paper and a human-digested remarks note are available on arXiv (OpenAI blog; arXiv preprint, OpenAI remarks). Coverage in Nature and Phys.org reports the same basic sequence: an LLM-based interaction, extensive model-generated calculations, and follow-up human verification and writeups (Nature; Phys.org).

The published materials indicate the construction leverages higher-dimensional arithmetic analogues mapped back to planar arrangements; OpenAI's remarks and the arXiv companion provide the technical sketch and formalization (OpenAI remarks; arXiv). Industry-pattern observations: advanced LLMs that combine symbolic manipulations and long-form chain-of-thought can produce long, checkable mathematical artifacts; practitioners should view these outputs as machine-produced drafts that require human verification and formal checking before being accepted as proofs.

For decades the prevailing belief among combinatorial geometers was that rescaled square-grid constructions were essentially optimal; both OpenAI's release and independent writeups frame the result as overturning that expectation for the planar unit distance problem (OpenAI blog; arXiv; Nature). Industry context: this episode is one of the first widely publicised instances where a general-purpose reasoning model produced a solution to a prominent, nontrivial mathematical conjecture and where the community was able to verify and digest the argument rapidly.

Observers will follow the peer-review and community vetting process around the arXiv papers and the human-digested remarks (arXiv; OpenAI remarks). Editorial analysis: researchers and toolmakers will monitor whether formal-verification tools (interactive theorem provers, computer algebra systems, mechanized proof checkers) are applied to convert the model-produced argument into a machine-checked formal proof, and whether similar workflows scale to other open problems. Also watch for detailed critiques or simplifications from domain experts, and for workshops or replication attempts documenting the model prompts, chain-of-reasoning artifacts, and the human steps needed to reach the published form (Gil Kalai blog; Marcus commentary).

The event combines a technical mathematical advance with a high-profile demonstration of LLM-assisted frontier research. The immediate claims and supporting documents are publicly available for scrutiny, and community verification and formalisation will determine how the result is integrated into the mathematical literature (OpenAI blog; arXiv; Nature).