Paul Erdős was known for offering cash prizes for hard problems. He probably didn't expect an AI to collect one 80 years later.
The unit distance problem sounds simple. If you place N dots on a flat plane, what's the maximum number of pairs that can be exactly one unit apart? Erdős posed it in 1946. He guessed the answer was something close to a squared grid, and for almost eight decades, that assumption held. Mathematicians tried tightening the bounds. Nobody managed to disprove it.
Then an OpenAI model did.
The Problem That Defined a Field
The unit distance problem belongs to discrete geometry, a branch of math that deals with arrangements of points, lines, and shapes. Erdős's specific conjecture was that the number of unit-distance pairs grows only slightly faster than the number of points. The conjecture stated that this was n^(1 + o(1)). The grid arrangement was considered basically optimal.
This was not a fringe problem. The upper bound on unit distances had not budged since 1984, when Spencer, Szemerédi, and Trotter published their proof. The Erdős unit distance problem sat on Thomas Bloom's "top 10" list of hardest discrete geometry open problems. He did not expect to see a solution anytime soon.
Bloom posted that list in April 2026. One month later, an OpenAI model cracked it.
How the AI Did It
The breakthrough came from an unexpected direction. Instead of attacking the problem geometrically, the model reached into algebraic number theory - specifically, infinite class field towers and Golod-Shafarevich theory. These are heavy tools usually reserved for abstract algebra, not point configurations on a plane.
The AI constructed vast lattices in high-dimensional space, then collapsed them into two dimensions. The result was a family of point configurations with more unit-distance pairs than any grid could produce. The improvement is small in absolute terms, about 1% more pairs per doubling of point count. But the proof is mathematically definitive. It proves Erdős's lower bound was wrong.
What makes this remarkable is that a human would have needed four things to line up simultaneously, as Bloom later noted: deep time investment in the problem, willingness to bet against Erdős's authority, an intuition to translate geometry into number fields, and specialized knowledge of class field theory. The AI met all four.
It combined what Thomas Bloom called "superhuman levels of patience with familiarity with a vast array of technical machinery."
What the Mathematicians Said
The reaction from the math community has been closer to shock than polite applause.
Misha Rudnev at the University of Bristol called it "absolutely a bomb", a problem he did not expect to see solved in his lifetime. Tim Gowers, a Fields Medalist, was even more direct. In companion remarks published alongside the proof, he wrote that if a human had submitted the paper to the Annals of Mathematics, he "would have recommended acceptance without any hesitation."
Arul Shankar, a leading number theorist, said the 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."
Noga Alon, another titan in combinatorics, confirmed the construction "applies fairly sophisticated tools from algebraic number theory in an elegant and clever way."
This consensus matters because OpenAI has a history of overclaiming on Erdős problems. Seven months ago, a former VP claimed GPT-5 had solved 11 Erdős problems, a claim quickly debunked as the model had merely retrieved existing solutions from literature. This time, the proof is real. The companion paper includes validations from mathematicians who criticized the previous claims. They wouldn't have signed off on something fake.
The Technical Details (for those who care)
The model discovered that by switching to progressively richer algebraic number systems at each step rather than scaling up a single system, you can construct point configurations that consistently beat the grid bound. The core insight involves using number fields with particular ramification properties to generate sets of algebraic integers whose differences yield more unit-length vectors.
The exact construction uses Golod-Shafarevich towers, infinite extensions of number fields with controlled ramification, to generate families of lattices with the right properties. The density gain is approximately n^(1 + δ) where δ ≈ 0.014, a polynomial improvement over the previous best.
The full proof is 40+ pages and freely available. The abridged chain of thought from the model is also published, offering a fascinating window into how an AI reasons about open mathematical problems.
So What
This is the first time an AI has autonomously resolved a prominent open problem at the center of an active mathematical field. That sentence should land harder than it will.
We've seen AI make impressive demo problems look easy. We've seen it generate plausible-sounding but wrong proofs. This is different. This is a real, verified, peer-validated solution to a problem that professional mathematicians have been wrestling with for 80 years. The model was general-purpose, not specialized for math. It had no special harness, no curated training for theorem-proving. It just reasoned its way through.
There is a quieter headline here that I keep coming back to. The AI did not just find an answer. It found an approach that humans had missed because of specialization. Bloom's four criteria for a human solver required expertise in both discrete geometry and algebraic number theory, skills that rarely live in the same researcher. The AI does not have that silo problem. It can hold every relevant technique in its working memory at once and explore connections a human would never think to try.
Terence Tao has warned about "proof indigestion". AI generating proofs faster than humans can digest them. That future may be closer than we think. The proof for this problem took days. The human validation took weeks. At some point, the ratio flips.
But that is tomorrow's problem. Today, an 80-year-old conjecture just got disproved by a machine that does not know what it is proving. That is unsettling and impressive in roughly equal measure.