An internal OpenAI model has autonomously disproved one of Paul Erdős’s favourite open problems — the planar unit distance conjecture from 1946 — finding an infinite family of constructions that yield a polynomial improvement over the square grid. Fields medalist Tim Gowers calls it “a milestone in AI mathematics.” It is the first time a prominent open problem central to a subfield has been solved by AI.
🔍 THE BOTTOM LINE
AI didn’t just assist a mathematician — it had an original, ingenious idea and carried it to fruition, using tools from algebraic number theory that nobody expected to apply to a simple geometry question.
What is the unit distance problem?
The planar unit distance problem asks a deceptively simple question: if you place n points in the plane, how many pairs can be exactly distance 1 apart? Erdős posed it in 1946, and the 2005 book Research Problems in Discrete Geometry called it “possibly the best known (and simplest to explain) problem in combinatorial geometry.”
For decades, the prevailing belief was that the rescaled square grid construction was essentially optimal. Erdős conjectured an upper bound of n^(1+o(1)) — meaning no construction could do meaningfully better than linear growth. The best upper bound, O(n^4/3), dates to Spencer, Szemerédi, and Trotter in 1984 and has barely budged since.
What the AI proved
The OpenAI model found an infinite family of configurations achieving at least n^(1+δ) unit-distance pairs for a fixed δ > 0 — a genuine polynomial improvement over anything previously known. Princeton’s Will Sawin refined the proof to give an explicit δ = 0.014.
The previous best construction had been essentially unchanged since Erdős’s original 1946 paper. The upper bound hadn’t moved since 1984. This result breaks open both sides of a problem that had been stuck for generations.
The surprise: algebraic number theory
Here’s what makes this genuinely novel, not just computationally impressive. The proof doesn’t use a clever variant of the grid. It brings in tools from algebraic number theory — infinite class field towers, Golod–Shafarevich theory — to construct number fields with richer symmetries that create far more unit-length differences than the Gaussian integers underlying the square grid.
These ideas were well-known to algebraic number theorists. Nobody expected them to have implications for a simple question about points in the Euclidean plane. As leading number theorist Arul Shankar put it: “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.”
It wasn’t a specialised math system
This might be the most unsettling part. The proof came from a general-purpose reasoning model — not a system trained specifically for mathematics, scaffolded to search through proof strategies, or targeted at the unit distance problem. OpenAI was testing it on a collection of Erdős problems as part of a broader evaluation of whether advanced models can contribute to frontier research. It just… solved one.
What this means
Mathematics has been the canonical testbed for AI reasoning because problems are precise, proofs can be checked, and a long argument only works if the reasoning holds together from beginning to end. This result passes that test with flying colours — external mathematicians verified the proof and wrote a companion paper explaining its significance.
But the deeper implication is that AI is no longer just a calculator or assistant for mathematicians. It can have genuinely original mathematical ideas — ideas that connect distant fields in ways that surprise even experts. Whether that’s exciting or terrifying probably depends on whether you’re a mathematician whose conjecture just got disproved.
❓ Frequently Asked Questions
Q: What does this mean for NZ? NZ universities with strong mathematics departments — Auckland, Victoria, Canterbury — should see this as a signal that AI research tools are becoming capable collaborators, not just productivity aids. Researchers who integrate AI reasoning into their workflow early will have a significant edge.
Q: Was this really autonomous? The model produced the proof without being specifically trained or scaffolded for this problem. It was evaluated on a collection of Erdős problems and solved one. However, Princeton’s Will Sawin refined the explicit constant — so human mathematicians are still adding value in tightening results.
Q: What should I do? If you work in research or mathematics, start experimenting with frontier AI models as reasoning partners, not just writing assistants. The gap between “AI can help me write” and “AI can help me think” just closed significantly.
🔍 THE BOTTOM LINE
An 80-year-old mathematical conjecture fell to a general-purpose AI that connected algebraic number theory to elementary geometry in a way no human had — and nobody saw it coming. The age of AI as a mathematical collaborator has arrived.
