OpenAI Model Solves 80-Year-Old Geometry Mystery, Disproving Erdős Conjecture
Table of Contents
A Breakthrough in Combinatorial Geometry
For nearly eight decades, mathematicians have been haunted by a question that is deceptively simple to ask but notoriously difficult to answer: if you place n points on a plane, what is the maximum number of pairs that can be exactly one unit of distance apart? This is known as the planar unit distance problem, a cornerstone of combinatorial geometry first posed by the prolific mathematician Paul Erdős in 1946.
For the vast majority of the time since Erdős first identified the puzzle, the mathematical community believed that a specific arrangement—essentially a rescaled square grid—represented the optimal way to maximize these distances. This long-standing assumption wasn’t just a hunch; it was a central conjecture that shaped how researchers approached discrete geometry for generations. Now, OpenAI has announced that one of its internal reasoning models has officially disproved it.
Beyond the Square Grid
The AI’s discovery provides an infinite family of examples that yield a polynomial improvement over the previously accepted bounds. To put this in perspective, the square grid constructions achieved a growth rate only slightly faster than linear. The new result proves that for infinitely many values of n, it is possible to construct configurations with significantly more unit-distance pairs.
While the AI’s original proof did not provide a specific numerical value for the improvement, subsequent refinement by Will Sawin, a professor of mathematics at Princeton, has identified a fixed exponent of $\delta = 0.014$. This may seem like a small number to a layperson, but in the world of discrete geometry, it represents a seismic shift in understanding.
What makes this result particularly jarring is that the solution did not come from a geometric intuition. Instead, the model drew from algebraic number theory—a field dealing with the factorization of integers in complex extensions—to solve an elementary geometric question. It is the kind of “cross-pollination” of mathematical disciplines that often characterizes human genius, yet here it was performed by a machine.
The Implications for AI Reasoning
Perhaps more significant than the mathematical result itself is the nature of the tool that found it. OpenAI notes that this proof was not generated by a system specifically trained for mathematics, nor was it the result of a targeted search for this specific problem. Instead, it emerged from a general-purpose reasoning model during a broader effort to test the system’s ability to contribute to frontier research.
The mathematical community has reacted with a mix of surprise and validation. Fields medalist Tim Gowers described the event as “a milestone in AI mathematics,” while number theorist Arul Shankar suggested that this proves current AI models are no longer just “helpers” to human researchers. According to Shankar, the model demonstrated the ability to conceive original, ingenious ideas and carry them through to a verified conclusion.
A New Era of Autonomous Research
Mathematics is often viewed as the ultimate stress test for artificial intelligence. Unlike coding or creative writing, a mathematical proof is binary: it is either correct or it is not. There is no room for “hallucinations” when a proof must hold together from the first axiom to the final Q.E.D.
The fact that the proof was independently verified by a group of external mathematicians lends the result a level of legitimacy that previous AI-generated “discoveries” lacked. By autonomously solving a problem that had resisted human effort since 1946, the model has signaled a shift in the capability of large-scale reasoning systems. We are moving away from AI that summarizes existing knowledge and toward AI that generates new, verifiable knowledge.
