OpenAI Model Solves 80-Year-Old Geometry Puzzle, Disproving a Central Erdős Conjecture
Table of Contents
A breakthrough in combinatorial geometry
For nearly eight decades, mathematicians have been haunted by a deceptively simple question: if you place a set of points in a plane, what is the maximum number of pairs that can be exactly one unit of distance apart? Known as the planar unit distance problem, the puzzle was first posed by Paul Erdős in 1946. It has since become a cornerstone of combinatorial geometry, admired for being easy to explain yet notoriously difficult to resolve.
Until now, the prevailing academic consensus was that a “square grid” construction was the most efficient way to maximize these unit-distance pairs. This belief had held steady for decades, with the best known lower bounds remaining largely unchanged since Erdős’s original work in the mid-1940s. That changed this week, as OpenAI announced that one of its internal reasoning models has officially disproved this longstanding conjecture.
The AI didn’t just find a fluke; it provided an infinite family of examples that yield a polynomial improvement over the square grid. By doing so, the model demonstrated that for infinitely many values of n, there are configurations that produce significantly more unit-distance pairs than previously thought possible.
Beyond a “math-specific” tool
What makes this result particularly striking to the research community is not just the solution itself, but how it was achieved. The proof did not come from a system specifically trained on mathematical datasets, nor was it the result of a narrow search through known proof strategies. Instead, the discovery emerged from a general-purpose reasoning model.
OpenAI had been evaluating the model against a collection of Erdős problems to test if advanced AI could contribute to frontier research. In a surprising turn, the model independently synthesized a proof for the unit distance problem. The process involved bringing in sophisticated concepts from algebraic number theory—a field dealing with the factorization of integers in algebraic number fields—and applying them to an elementary geometric question. This leap in “cross-domain” reasoning is what has captured the attention of the academic world.
The proof has been rigorously vetted by a group of external mathematicians. Fields medalist Tim Gowers, who contributed to a companion paper explaining the significance of the find, described the result as “a milestone in AI mathematics.” Arul Shankar, a leading number theorist, noted that the model has moved beyond being a mere “helper” for humans and is now capable of generating original, ingenious ideas and seeing them through to completion.
The technical shift
To understand the magnitude of the shift, one has to look at the growth rates. Traditionally, placing points in a line results in a linear growth rate of n-1 pairs. A square grid performs better, but only slightly. For years, the mathematical community believed the growth rate was essentially limited to n1+o(1), where the additional term tends toward zero as n increases.
The OpenAI model disproved this, showing that configurations can achieve a growth rate of n1+δ for a fixed exponent δ > 0. While the AI’s original proof didn’t specify the exact value of δ, Will Sawin, a professor of mathematics at Princeton, has since refined the work to show that δ = 0.014.
This represents a fundamental shift in how researchers view the limits of point configurations in a plane. For comparison, the best upper bound—established in 1984 by Spencer, Szemerédi, and Trotter—has remained largely static for forty years. By narrowing the gap from the bottom up, the AI has provided a new baseline for a problem that had stalled for generations.
The result serves as a high-profile proof of concept for the “reasoning” capabilities of the latest generation of LLMs. Unlike standard chat interfaces that often hallucinate mathematical proofs, this model’s output was precise enough to be verified by world-class experts, marking a transition from AI as a linguistic tool to AI as a genuine collaborator in theoretical science.
