OpenAI Model Yields Polynomial Gain on Unit Distance Pairs

After nearly 80 years of study, a central conjecture in discrete geometry has been disproved by an internal OpenAI model, a system not specifically designed for mathematical proofs. The longstanding planar unit distance problem, originally posed by Paul Erdős, asks how many pairs of points can be exactly one unit apart when placed in a plane, and mathematicians believed constructions based on square grids were the most efficient way to maximize those pairs. OpenAI’s model has now demonstrated a polynomial improvement over these constructions, providing an infinite family of examples that challenge the previously held belief. A mathematician familiar with the work says this result is “an outstanding achievement, settling a long-standing open problem,” and notes that it has been one of Erdős’ favorite problems. This marks the first time AI has autonomously solved a prominent, open problem central to a subfield of mathematics, demonstrating a new level of reasoning capability.

OpenAI Model Disproves 80-Year-Old Unit Distance Conjecture

An OpenAI model has achieved a feat previously thought beyond the reach of artificial intelligence: disproving a longstanding conjecture in discrete geometry that has challenged mathematicians for nearly 80 years. Described in the book Research Problems in Discrete Geometry as “possibly the best known (and simplest to explain) problem in combinatorial geometry,” the question’s simplicity belies its resistance to solution, even attracting a monetary prize offered by Erdős himself. For decades, the prevailing belief centered around “square grid” constructions as being nearly optimal for maximizing these unit-distance pairs. This wasn’t simply a matter of finding a counterexample; the model’s proof has been rigorously checked by a group of external mathematicians who have also authored a companion paper explaining the argument and providing further background and context. This breakthrough wasn’t achieved by a system specifically designed for mathematical proofs.

Instead, it emerged from a general-purpose reasoning model, evaluated as part of a broader effort to assess AI’s potential to contribute to frontier research. The model’s approach is particularly striking, drawing upon sophisticated concepts from algebraic number theory to address an elementary geometric question.

Fields medalist Tim Gowers calls the result “a milestone in AI mathematics,” adding that if a human had written the paper and submitted it to the Annals of Mathematics, he would have recommended acceptance without hesitation. Number theorist Arul Shankar believes this paper demonstrates that current AI models go beyond just helping human mathematicians, they are capable of having original, ingenious ideas and carrying them out to fruition. Researchers analyzing the model’s chain of thought reveal a strong focus on constructing counterexamples, suggesting an intuitive grasp of the problem’s challenges and a willingness to explore unconventional approaches.

The disproof involves demonstrating configurations of points with at least n^1+δ unit-distance pairs, where δ is a positive exponent, surpassing the previously believed limit of n^1+o(1). This represents a significant advancement in understanding the bounds of unit distance configurations and opens new avenues for exploration in combinatorial geometry.

Erdős’s Unit Distance Problem & Prior Bounds of n^(1+o(1))

These constructions achieved a growth rate of n^{1 + C / log ⁡ log ⁡ ( n )}, where C is a constant, indicating a rate only slightly faster than linear growth. For decades, mathematicians operated under the assumption that no construction could significantly surpass this square grid limit, leading to a widely held conjecture that the maximum number of unit-distance pairs adhered to an upper bound of n^1+o(1), where the term o(1) signifies a value approaching zero as n increases. This belief persisted despite refinements to existing techniques and related structural work by researchers like Székely, Katz, Silier, Pach, Raz, and Solymosi. However, an internal OpenAI model has now demonstrated a polynomial improvement, constructing configurations of n points that achieve at least n^{1 + δ} unit-distance pairs, with δ being a positive constant. The method employed by the OpenAI model is noteworthy, bringing sophisticated ideas from algebraic number theory to bear on an elementary geometric question, including concepts such as infinite class field towers and Golod, Shafarevich theory.

New Construction Yields n^(1+δ) Unit Pairs for Infinite n

OpenAI researchers have achieved a significant advance in discrete geometry, demonstrating a construction that yields at least n^(1+δ) unit-distance pairs for infinitely many values of n, a result that surpasses the previously believed optimal bounds. This disproves a conjecture dating back nearly 80 years to the work of Paul Erdős, who first posed the planar unit distance problem and even offered a monetary prize for its resolution. The new construction moves beyond the linear growth rate previously achieved by “square grid” methods, which give about 2n pairs, and even improvements reaching n^{(1 + C / log log(n))}. For decades, mathematicians considered these square grid constructions to be near the theoretical limit, believing no substantial improvement was possible. The OpenAI model’s proof, however, demonstrates that configurations can be created with a fixed exponent δ > 0, fundamentally altering the understanding of achievable density of unit-distance pairs.

The model doesn’t provide an explicit value for δ in its initial proof, but a forthcoming refinement by Will Sawin of Princeton University has shown one can take δ = 1. The proof brings unexpected, sophisticated ideas from algebraic number theory to bear on an elementary geometric question. These fields, with their richer symmetries, allow for the creation of a significantly greater number of unit-length differences. OpenAI explains that the precise argument uses tools such as infinite class field towers and Golod, Shafarevich theory.

AI Reasoning Model Solves Problem Autonomously

The resolution of a decades-old mathematical problem demonstrates a shift in how complex challenges are approached, with implications extending beyond pure mathematics. An internal OpenAI model recently disproved a longstanding conjecture within discrete geometry, the planar unit distance problem, not through targeted design but as a byproduct of evaluating a general-purpose reasoning model. This achievement signifies a potential paradigm shift, moving beyond AI as a computational tool and towards AI as an autonomous originator of mathematical insight. For nearly 80 years, mathematicians have grappled with determining the maximum number of unit-distance pairs possible among a given set of points on a plane, a question initially posed by Paul Erdős. The problem’s simplicity belies its intractability, earning it a prominent place in combinatorial geometry. Noga Alon, a leading combinatorialist at Princeton, describes it as “one of Erdős’ favorite problems,” a sentiment echoed by many in the field.

However, the OpenAI model has provided a polynomial improvement over these constructions, achieving a growth rate of n^1+δ for some fixed exponent δ > 0, effectively disproving the long-held conjecture. The significance of this result isn’t solely in the disproof itself, but in how it was achieved. The documentation accompanying the discovery explains that the proof brings unexpected, sophisticated ideas from algebraic number theory to bear on an elementary geometric question.

Algebraic Number Theory Applied to Planar Geometry

The resolution of the planar unit distance problem, a challenge spanning nearly 80 years, wasn’t achieved through conventional geometric reasoning, but through an unexpected application of abstract algebra. While mathematicians tirelessly explored constructions within the plane itself, the breakthrough arrived via concepts originating in algebraic number theory, a field concerned with extending number systems. This surprising connection highlights how seemingly disparate branches of mathematics can unexpectedly converge to solve longstanding problems. The proof brings sophisticated ideas from algebraic number theory to bear on an elementary geometric question. These fields were brought to bear on the problem, allowing for the creation of configurations with at least n^(1+δ) unit-distance pairs, surpassing the previously believed limit of n^(1+o(1)). This isn’t merely a matter of finding a slightly better arrangement of points; it’s a shift in approach.

For decades, mathematicians believed the “square grid” constructions were approaching optimality, giving about 2n pairs, but the AI-generated proof reveals that these constructions were fundamentally limited. The model didn’t simply refine existing techniques; it imported tools from a different mathematical landscape, demonstrating a capacity for creative problem-solving previously thought exclusive to human researchers. The implications extend beyond the specific problem solved, signaling a potential paradigm shift in mathematical research where AI systems can not only verify proofs but also generate novel approaches and uncover hidden relationships between different mathematical disciplines, ultimately expanding the boundaries of human knowledge.