OpenAI model breaks 80-year-old Erdős conjecture in discrete geometry

An OpenAI model has disproved a central conjecture Paul Erdős posed in 1946, ending an 80-year impasse on the unit-distance problem in discrete geometry. The system produced an infinite family of point configurations that yield more pairs of exactly unit-distance neighbors than any construction on record, and a group of external mathematicians verified the result.

What is the unit-distance problem?

The question fits in one line: given n points in the plane, how many pairs can sit exactly one unit apart?

The 2005 book Research Problems in Discrete Geometry, by Brass, Moser, and Pach, called it "possibly the best known (and simplest to explain) problem in combinatorial geometry." Noga Alon, a combinatorialist at Princeton, described it as "one of Erdős' favorite problems." Erdős himself offered a monetary prize for its resolution.

For decades the standard example was a square grid, which delivers roughly 2n unit pairs for n points. A rescaled variant does slightly better: n^(1 + C/log log n) pairs, where the extra exponent drifts toward zero as n grows. The best upper bound, O(n^(4/3)), dates to Spencer, Szemerédi, and Trotter in 1984. Refinements by Székely, Katz and Silier, Pach, Raz, Solymosi, and others failed to move it materially.

Erdős conjectured the true answer was n^(1+o(1)) — in technical terms, that no construction could improve meaningfully on the square grid.

What did the OpenAI model prove?

The model constructs, for infinitely many values of n, configurations with at least n^(1+δ) unit-distance pairs, where δ is a fixed constant greater than zero. That exponent locks in place as n grows. Past constructions could only match an exponent that slowly drifted toward 1.

OpenAI's initial proof did not give a numerical value for δ. A forthcoming refinement by Will Sawin, a Princeton mathematics professor, pegs δ at 0.014.

External mathematicians checked the proof, then wrote a companion paper situating the argument in the field. OpenAI published the proof, the companion note, and an abridged version of the model's chain of thought.

What did leading mathematicians say about it?

Fields medalist Tim Gowers, a contributor to the companion paper, called the result "a milestone in AI mathematics." Gowers's framing carries weight: he helped reshape large parts of combinatorics over the last three decades.

Arul Shankar, a leading number theorist at the University of Toronto, said: "In my opinion this paper demonstrates that current AI models go beyond just helpers to human mathematicians – they are capable of having original ingenious ideas, and then carrying them out to fruition."

Thomas Bloom, who wrote the companion note, argued the broader implication: "AI is helping us to more fully explore the cathedral of mathematics we have build over the centuries; what other unseen wonders are waiting in the wings?"

Bloom also addressed what math itself gains. "When assessing the importance and influence of an AI-generated proof, a question I ask myself is: has this taught us something new about the problem? Do we understand discrete geometry better now? I think the answer is a moderated yes: this shows that there is a lot more that number theoretic constructions have to say about these sorts of questions than we suspected."

Why is the result mathematically surprising?

The construction pulls in machinery from algebraic number theory — specifically, infinite class field towers and Golod–Shafarevich theory — to build the required point sets.

Erdős's original lower bound already used Gaussian integers: numbers of the form a+bi where a and b are ordinary integers. The new argument swaps that for richer arithmetic structures whose extra symmetries produce many more unit-length differences.

Those number-theoretic tools were familiar to specialists. What surprised the field was that they would land on an elementary geometric question in the Euclidean plane.

How was the proof obtained?

OpenAI attributed the result to a new general-purpose reasoning model, not a system trained on mathematics or scaffolded to search proof strategies. The company evaluated the model on a collection of Erdős problems as part of a broader test of whether frontier systems can contribute to original research.

OpenAI described mathematics as "a particularly clear testbed for reasoning." Problems are precise, candidate proofs can be checked, and a long chain of logic either holds or fails from start to finish.

Why does this matter beyond mathematics?

The same capabilities that let a model hold a long proof together — coherent reasoning across many steps, connecting distant ideas, surviving expert scrutiny — apply elsewhere. OpenAI pointed to biology, physics, materials science, engineering, and medicine.

The company framed the milestone as part of "our longer-term path toward more automated research: systems that can help scientists and engineers explore more ideas and pursue harder technical questions." It also flagged open questions about alignment, noting "the challenges of aligning very intelligent systems, and the future of human-AI collaboration."

What comes next?

Mathematicians will start looking at sibling problems. Bloom predicted "no doubt many algebraic number theorists will be taking a close look at other open problems in discrete geometry in the coming months." He also projected a wider pattern: "no doubt the coming months and years will see similar successes in many other areas of mathematics, where long-standing open problems are resolved by an AI revealing unexpected connections."

Human mathematicians still set the agenda. OpenAI stressed that "expertise becomes more valuable, not less. AI can help search, suggest, and verify. People choose the problems that matter, interpret the results, and decide what questions to pursue next."