Beyond the Square Grid: How AI Disproved a Longstanding Conjecture in Discrete Geometry

For nearly 80 years, the mathematical community has grappled with a deceptively simple question: if you place $n$ 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, this puzzle was first posed by Paul Erdős in 1946 and has since become one of the most famous open questions in combinatorial geometry.

For decades, the prevailing belief was that constructions based on rescaled square grids were essentially optimal. However, a new general-purpose reasoning model from OpenAI has disproved this longstanding conjecture, providing an infinite family of examples that yield a polynomial improvement over previous bounds. This result marks a significant milestone: the first time a prominent open problem in a specialized mathematical subfield has been solved autonomously by AI.

The Unit Distance Problem and the Erdős Conjecture

To understand the significance of this breakthrough, one must first understand the scale of the problem. Let $u(n)$ represent the largest possible number of unit-distance pairs among $n$ points. Simple linear growth is easy to achieve—placing points in a line gives $n-1$ pairs, and a square grid provides roughly $2n$.

Previous state-of-the-art constructions used rescaled square grids to achieve a growth rate of $n^{1 + C / \log \log n}$. Because $\log \log n$ grows so slowly, this rate is only slightly faster than linear. Erdős conjectured that this was essentially the limit, proposing an upper bound of $n^{1 + o(1)}$, where the additional term tends toward zero as $n$ increases.

OpenAI's model has shattered this ceiling. The AI constructed configurations of $n$ points with at least $n^{1 + \delta}$ unit-distance pairs for a fixed exponent $\delta > 0$. While the original AI proof did not specify $\delta$, subsequent refinement by Princeton professor Will Sawin has shown that $\delta$ can be $0.014$. This polynomial improvement proves that the square grid is not the optimal configuration.

A Surprising Bridge: Algebraic Number Theory

What makes this discovery particularly remarkable is not just the result, but the method. The model did not rely on a system specifically trained for mathematics or a targeted search for this specific problem. Instead, it leveraged a general-purpose reasoning capability to connect two seemingly unrelated fields: discrete geometry and algebraic number theory.

Erdős's original lower bound was rooted in Gaussian integers ($a + bi$). The AI's proof expanded this logic, replacing Gaussian integers with more complex generalizations from algebraic number theory. By utilizing tools such as infinite class field towers and Golod–Shafarevich theory, the model identified number fields with richer symmetries capable of creating significantly more unit-length differences in the Euclidean plane.

As Thomas Bloom noted in the companion paper, this reveals that "there is a lot more that number theoretic constructions have to say about these sorts of questions than we suspected."

The AI-Human Collaborative Loop

While the AI provided the core proof, the result highlights a new paradigm of collaboration. The initial AI output was verified and expanded upon by external mathematicians, who wrote a companion paper to provide the necessary context and rigor.

This process underscores a critical point: AI can surface promising paths and unexpected connections, but human expertise remains essential for interpretation and refinement. As one commentator noted, the value of the result lies not just in the answer, but in the "new pathway the AI used," which provides a roadmap for humans to explore other problems in discrete geometry using number theory.

Technical and Philosophical Implications

The community's reaction to this breakthrough has been a mixture of awe and skepticism. Several key debates have emerged from the discussion:

Interpolation vs. Innovation

Some critics argue that LLMs merely interpolate their training data. However, others counter that mathematical discovery is often the act of recombining existing tools in novel ways.

"Proofs unfold what is already implicit in axioms, definitions, symbols, and rules... either 'recombining existing material' isn't disqualifying, or a lot of Fields Medals need to be returned."

The Nature of the "Discovery"

There is a distinction between proving a conjecture true and providing a counterexample (a disproof). Some argue that finding a counterexample is a more advanced form of search rather than "theory crafting." Yet, the fact that the AI identified an error in the intuition of one of the greatest mathematicians of the 20th century suggests a level of reasoning that transcends simple pattern matching.

The Complexity Barrier

For working mathematicians, the most practical implication is the ability to manage the "exploding complexity barrier." As the volume of mathematical research grows, it becomes impossible for a single human to maintain a mental map of all relevant progress. AI's ability to synthesize knowledge across distant domains—like algebraic number theory and geometry—could break through the super-specialization that currently limits scientific innovation.

Looking Ahead

This breakthrough suggests that AI is moving beyond the role of a "helper" and into the realm of an original contributor. The ability to maintain a coherent, complex argument over a long horizon—evidenced by a summarized chain-of-thought spanning 125 pages—indicates that reasoning models are approaching a threshold where they can tackle frontier research in physics, biology, and materials science.

As we move forward, the role of the human researcher will likely shift toward choosing the problems that matter, interpreting the results, and deciding which "unseen wonders" of the mathematical cathedral to explore next.