The Erdős Machine

Scatter a handful of dots on a blank page. Now count the pairs that sit exactly one inch apart. Add more dots, arrange them cleverly, and the count of those "unit distances" climbs — but how high can it go? Paul Erdős asked that question in 1946, and for the better part of a century, the world's mathematicians believed they knew the shape of the best answer: a tidy square grid. In May 2026, a machine that was never taught geometry proved them wrong.

The result came from OpenAI, which announced on May 20 that one of its internal reasoning models had disproved a long-standing prediction tied to the unit distance problem — one of the most famous open questions in discrete geometry. The model did not simply check a human's work or grind through a calculation. It produced a genuinely new mathematical argument, built from tools in a field called algebraic number theory, and arrived at configurations of points that pack in more unit-distance pairs than the grid everyone had assumed was optimal. The proof was valid. The community was, by its own account, stunned.

To understand why mathematicians reacted the way they did, you have to understand what kind of problem this is. The unit distance problem is not a puzzle with a known method waiting to be executed faster. It is the opposite: a problem where the entire difficulty lies in imagining the right construction, in seeing a structure that no one has seen before. That act of seeing has, until now, been considered irreducibly human.

What Erdős Actually Asked

Take n points in the plane. Draw a line between any two of them that happen to be exactly distance one apart. The unit distance problem asks: across all possible arrangements of those n points, what is the maximum number of such unit-distance pairs you can create? Erdős found a construction based on a square grid that produces a certain number of these pairs, and he conjectured bounds on how the count could grow as the number of points increased.

The grid is seductive because it is regular and easy to reason about. Stack points in a lattice and many of them naturally fall a fixed distance apart. For decades the working assumption was that grid-like constructions were essentially the best you could do — that nature had no cleverer trick hiding in the geometry. This is the prediction the AI overturned. Its constructions deliver what mathematicians describe as a polynomial improvement over the grid, and crucially, they do so not for a single lucky value of n but for an infinite family of them.

The choice of weapon is what makes the result so disarming. Rather than tinkering with the geometry directly, the model reached into algebraic number theory — the study of numbers that arise as solutions to polynomial equations — and used its structure to place points with the precision needed to manufacture extra unit distances. It is the kind of cross-disciplinary leap that earns human mathematicians a reputation for genius. No one had pointed the model at that particular toolbox. It went there itself.

Why the Mathematicians Are Paying Attention

Skepticism is the default setting of the mathematical community, and for good reason — the field has seen a long history of grand claims that dissolved under scrutiny. So the reaction to this result is telling. Fields medalist Timothy Gowers, one of the most respected living mathematicians, called it a milestone in AI mathematics. He and Noga Alon, another giant of combinatorics, lent their names as endorsers to a companion note explaining the disproof. These are not people who attach their reputations to hype.

What changed is the type of contribution. AI has been useful in mathematics for years — searching for patterns, suggesting conjectures, formalizing proofs in verification systems like Lean. But those were supporting roles. The machine fetched; the human discovered. The unit distance result inverts that relationship. Here the machine produced the central creative idea, and humans played the supporting role: discussing it, digesting it, polishing the argument into its cleanest form, and exploring what else it implied.

That polishing is worth dwelling on, because it complicates the triumphant headline. OpenAI and the mathematicians involved were candid that the AI's original proof, while valid, was rough — and that human researchers significantly improved it. Within days of the announcement, the mathematician Will Sawin followed the same line of reasoning to an even stronger result. So the honest description is not "AI replaced mathematicians." It is something more interesting: a human–machine collaboration in which the machine supplied the spark.

The Quiet Caveats

It would be easy — and wrong — to read this as the moment mathematics was automated. Several caveats deserve to sit in plain sight. First, this was a disproof, a counterexample. Finding a configuration that beats the conjecture is genuinely hard, but it is a different kind of act from constructing a sweeping proof that something is true for all cases forever. Counterexamples can sometimes be found by sufficiently inventive search; universal proofs generally cannot.

Second, the result has not yet completed the slow, adversarial process of peer review that ultimately certifies mathematical truth. The endorsements of Gowers and Alon carry weight, and the argument has been written up for scrutiny, but the field rightly reserves final judgment until the proof has been picked apart by many eyes over time. Extraordinary claims earn extraordinary checking.

Third, there is the uncomfortable question of how the model did it — and whether it can do it again on demand. A single spectacular result, even a real one, is not the same as a reliable engine of discovery. The history of AI is littered with breakthroughs that turned out to be narrower than they first appeared. The honest position today is that we have one remarkable data point, not yet a trend.

What It Means That a Machine Can Surprise Us

Strip away the caveats and something undeniable remains. A system built to reason in general terms reached into a specialized corner of mathematics, found an idea that eluded the human experts who had stared at the problem for eighty years, and was right. Whatever the limits, that has not happened before. It is a category of event, not merely a point on a curve.

For most of the history of computing, machines extended our reach but not our insight. They were faster, more tireless, more exact — but the ideas were ours. The unit distance result pokes a hole in that comfortable division of labor. It suggests that the part of mathematics we considered most human — the leap, the construction, the seeing — may not be uniquely human after all. And it raises a question that will define the next decade of science: if a machine can surprise the experts in pure mathematics, the most abstract discipline we have, what happens when we point the same capability at physics, at biology, at the unsolved problems that actually shape our lives?

Erdős, who spent his life chasing beautiful problems and famously spoke of an imaginary "Book" in which God keeps the most elegant proofs, would have found this irresistible. He once said a mathematician is a machine for turning coffee into theorems. Eighty years after he posed his question, a different kind of machine has turned electricity into one — and quietly added a page to the Book.