Breaking the Grid: How AI Disproved a Longstanding Conjecture in Discrete Geometry
For nearly eight decades, mathematicians have grappled with a deceptively simple question in combinatorial geometry: 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, first posed by Paul Erdős in 1946, it has long been regarded as one of the most accessible yet stubbornly difficult problems in the field.
For decades, the prevailing consensus was that a rescaled square grid was essentially the optimal construction for maximizing these pairs. However, a general-purpose reasoning model from OpenAI has recently disproved this longstanding conjecture, providing an infinite family of examples that yield a polynomial improvement over the square grid. This result marks a significant milestone: the first time a prominent open problem in a central subfield of mathematics has been solved autonomously by AI.
The Mathematics of the Unit Distance Problem
To understand the breakthrough, one must first understand the baseline. Let $u(n)$ be the largest possible number of unit-distance pairs among $n$ points. Simple constructions, such as placing points in a line or a square grid, yield linear growth (roughly $n-1$ or $2n$ pairs, respectively).
Previous state-of-the-art constructions based on rescaled square grids achieved a growth rate of $n^{1 + C / \log \log n}$. Because $\log \log n$ grows incredibly slowly, this rate is only slightly faster than linear. Erdős conjectured an upper bound of $n^{1 + o(1)}$, meaning any additional term in the exponent would tend toward zero as $n$ increases.
The AI's discovery shatters this ceiling. The model 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 established that $\delta$ can be $0.014$. This is a polynomial improvement, proving that the square grid is not the optimal way to organize points in a plane to maximize unit distances.
A Bridge Between Distant Fields
What makes this result particularly striking is not just the answer, but the method. The proof did not emerge from a system trained specifically for mathematics or a tool designed for geometric search. Instead, it came from a general-purpose reasoning model that bridged two seemingly unrelated areas of mathematics: discrete geometry and algebraic number theory.
Erdős’s original lower bound relied on Gaussian integers (numbers of the form $a + bi$). The AI 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 far more unit-length differences than a standard grid could.
As Thomas Bloom noted in the companion paper:
"This shows that there is a lot more that number theoretic constructions have to say about these sorts of questions than we suspected; moreover, that the number theory required can be very deep."
The AI Debate: Interpolation vs. Innovation
The announcement has sparked a rigorous debate among the technical community regarding the nature of AI "discovery."
The "Stochastic Parrot" Argument
Some skeptics argue that LLMs merely interpolate their training data. However, others counter that mathematical discovery itself is often the recombination of existing axioms and tools. As one commenter noted, if recombining existing material is disqualifying, then many human Fields Medals would be called into question, as proofs often unfold what is already implicit in the rules of the system.
Search vs. Theory Crafting
Another point of contention is whether finding a counterexample (a disproof) is fundamentally different from proving a conjecture true. Some argue that a disproof is a highly advanced form of search—finding the one configuration that breaks the rule—whereas a proof requires the creation of a new theoretical framework. Despite this, the ability to navigate a 125-page chain of thought to find a non-trivial counterexample suggests a level of reasoning that transcends simple pattern matching.
The Role of Human Verification
Crucially, the AI did not work in a vacuum. The result was verified by external mathematicians, and a companion paper was written to provide context. This highlights a burgeoning model of human-AI collaboration: AI surfaces a promising, non-intuitive path (the "ingenious idea"), and humans provide the rigorous verification and theoretical framing.
Implications for Future Research
This breakthrough suggests that AI's greatest strength in science may be its ability to overcome "super-specialization." While a human expert might be deeply entrenched in discrete geometry, they may not be thinking in terms of Golod–Shafarevich theory. AI, having been trained on the entirety of available mathematical literature, can transfer tools across domains with zero friction.
Beyond mathematics, this capability—maintaining a coherent, complex argument over a long horizon and connecting distant fields—has immediate implications for biology, materials science, and medicine. If a model can identify a counter-intuitive connection in geometry, it may eventually identify a hidden connection between a specific protein structure and a distant chemical property, accelerating the pace of automated research.