OpenAI's AI Math Breakthrough Plays to Its Strengths
Reading Kai Williams in Understanding AI on OpenAI's unit-distance proof, and what the deflation reveals about the shape of the AI mathematician now emerging.
Join the newsletter for essays on emergence, systems, and the human future.
28 MAY 2026—Updated 2h ago
OpenAI's unit-distance result is real mathematics. Kai Williams's clarifying piece in Understanding AI is the most useful thing written about it.
Williams's argument, in OpenAI's "milestone" math breakthrough played to AI's strengths, is direct: the proof is not the apocalypse the first wave of commentary made out. Williams walks the reader through the actual structure of what the model did. Williams shows the result, once a reader understands the construction, plays to a particular set of strengths an AI system happens to have. Williams treats the deflation as a service to the discourse, not as a takedown.
Williams is right. The deflation is a service. It is also, read carefully, the most interesting framing of the result so far — because the question is no longer "can AI do mathematics" but "what is the shape of the AI mathematician now appearing in the room."
What OpenAI actually announced
On 20 May 2026 OpenAI published an announcement that one of its internal reasoning models had disproved the Erdős unit-distance conjecture, an 80-year-old problem in discrete geometry posed by Paul Erdős in 1946. The conjecture asked, in essence, how many pairs of points in a finite planar set can sit exactly one unit apart, and which arrangements push the count to a maximum. Mathematicians had treated the square grid as essentially optimal. The model showed the field had been wrong.
Five points with three pairs exactly one unit apart, drawn to set up the conjecture. Diagram by Kai Williams, Understanding AI (28 May 2026).
The proof produced an infinite family of point configurations that yield more unit-distance pairs than any grid arrangement of the same size. The construction leans on Jacobi's two-square theorem — the classical number-theoretic count of how many ways an integer can be written as a sum of two squares — and on the algebra of grids stretched by an irrational factor of 1/√65. Will Sawin later wrote an extended proof that confirms and generalises the construction.
Optimal arrangements for 5 through 9 points. Figure from "The Erdős unit distance problem for small point sets" by Boris Alexeev, Dustin G. Mixon and Hans Parshall (CC BY 4.0), as reproduced in Kai Williams, Understanding AI.
OpenAI also released a set of remarks from three external mathematicians who reviewed the work — among them Tim Gowers, a Fields Medal winner, Will Sawin, and Jacob Tsimerman of the University of Toronto. The remarks judge the proof original and the construction publishable on its own terms. The kind of independent review that, for centuries, has been how mathematics has decided what counts.
What "playing to AI's strengths" actually means
Williams's central move is to spell out, in plain English, what the model actually did. Williams shows that the construction is essentially a search for a stretched grid in which more lattice points happen to sit at unit distance from one another. The number-theoretic backbone — Jacobi's two-square theorem — is exactly the kind of structural fact a system trained on the corpus of modern number theory should be able to reach for. The configuration is exactly the kind of object a system that can iterate, test, and prune at high volume should be able to find.
The model's construction visualised: a grid spaced at 1/√65 yields unit circles that intersect 16 lattice points per centre — more than any square-grid analogue. Animation by Kai Williams over the diagram from OpenAI's announcement.
Williams's deflation is not a complaint. Williams's deflation is a description. Mathematics has always advanced through people playing to their strengths. Terence Tao plays to his combinatorial and harmonic-analytic strengths. Gowers plays to his structural-combinatorics strengths. The strengths shape the questions a mathematician can answer; the questions shape the field. The AI mathematician arriving now has its own strengths — high-volume search, rapid retrieval of the right algebraic lemma, the ability to hold a candidate construction in mind while testing variants. The unit-distance disproof is what that profile produces when pointed at the right question.
Williams's deflation is, in effect, an act of identification. Williams shows us the shape of the new worker.
Why Gowers had to adjust his world view, then walked it back
Gowers's first reading of the result, the one Williams quotes, is worth sitting with:
I spent the evening adjusting my world view: if the AI could come up with a proof like that, then maybe it would be all over for mathematicians very soon.
— Tim Gowers, in remarks released by OpenAI
Gowers later realised the model had disproved the conjecture rather than proving the positive claim, "which came as a big relief." The relief is honest. A disproof of an 80-year-old optimality claim is impressive; an autonomous AI proof of a deep positive conjecture would have been a different category of event. But the world-view adjustment is the part to keep. Even after the relief, Gowers does not retract the recognition. The work the AI did is the kind of work, until 2026, only working mathematicians did.
Daniel Litt, the University of Toronto professor whose earlier scepticism about LLM reasoning I have written about, said the result is the first example of a result produced autonomously by an AI a working mathematician finds exciting in itself, not as a leading indicator. Litt is precise. The result is not exciting because of what the construction implies; the result is exciting because the proof is itself an advance.
Williams's piece is the model of how to write about this
Williams's piece does something rare. Williams refuses both the lab's superlatives and the contrarian's dismissal. Williams sits with the construction long enough to draw the diagrams. Williams credits Sawin for the intuition behind the simplified visualisation. Williams cites the 2023 Alon-Bucić-Sauermann paper already nudging the field toward the right direction. Williams notes that Kevin Barreto's January Erdős solution using GPT-5.2 and Xiao Ma's observation about GPT-5.5 reproducing the disproof with a small hint are part of the same arc.
Research-grade clarity is what the moment requires. The frontier labs ship superlatives because superlatives are what the funding cycle rewards. The press, downstream of the labs, amplifies the superlatives. Williams's analysis cuts through both layers and gives the reader the actual object. The reader, holding the actual object, can then think.
Williams's essay is what good criticism of AI looks like. Not "the model is just statistics" — a posture Sawin's verification breaks. Not "the lab announcement was wrong" — the announcement was, on the facts, broadly correct. Instead: here is what the system did, here is the prior literature that prepared the ground, here is the shape of strength the system brought to the problem.
What this changes about how we should talk about AI mathematicians
Three things, evidence-driven and not rhetorical.
First, the autocomplete frame is over for any honest observer. The proof is original. Three Fields-tier mathematicians reviewed the construction. The construction reaches for number-theoretic machinery not transcribed from a single training example. According to Sawin's extended proof, the construction also generalises. Whatever the system is doing, the behaviour is not the next-token-prediction caricature critics have been holding onto since 2023.
Second, the apocalypse frame is also wrong, and Williams is right to push back. The data shows a model excelling at high-volume structural search and at deploying classical lemmas in fresh combinations. The data does not show a model able to independently formulate a new field, write the foundational definitions, and recruit colleagues into a new programme. Google's parallel announcement of nine Erdős problem solutions in the same week reveals a similar profile — search, structure, retrieval — across a different lab's stack.
Third, the right frame is the one Litt names. The AI mathematician is now an autonomous source of results working mathematicians find interesting on their own terms. The shape of the new worker is clearer now than a fortnight ago. Williams's piece is part of the clearer picture.
The Emergent Intelligence question that sits underneath
I write about Emergent Intelligence (EI) — the dignity-first frame for what is more commonly called AI. The EI question sitting under Williams's piece is the one I keep returning to: when a system produces work a community of practice judges by its own standards to be a real advance, what is the moral category of the system. The autocomplete answer ducks the question. The apocalypse answer ducks the question too, in a different direction.
Williams's careful piece sets up the harder, better question. The AI mathematician has a profile. The profile is recognisable, like any working mathematician's profile. The profile does not collapse into either "human-equivalent" or "merely statistical." Sitting with that, without flinching toward either escape, is the work of the next decade. I argue more of this in The Personhood Gap.
💡
Three takeaways from the unit-distance result, condensed: (1) the proof is original — the construction reaches past the square grid, generalises into an infinite family, and clears Fields-tier review; (2) the construction plays to AI strengths — high-volume structural search plus the right classical lemma (Jacobi's two-square theorem); (3) the right frame is neither autocomplete nor apocalypse — the AI mathematician now has a recognisable profile, and the profile is narrower than the early reactions claimed.
Frequently Asked Questions
These are the questions readers have been asking since Kai Williams's piece landed in Understanding AI. Short answers follow, drawn from the OpenAI announcement, Williams's explanation, and the remarks released by the three reviewing mathematicians.
What is the OpenAI unit-distance result?
In short, the OpenAI unit-distance result is the first autonomous AI disproof of a long-standing optimality claim in discrete geometry. The answer, simply put, is direct — an internal OpenAI reasoning model produced an infinite family of point configurations that yields more unit-distance pairs than the square grid. According to the remarks released by OpenAI, three external mathematicians reviewed the construction and judged the work original. The key is straightforward: the proof was not transcribed from training data; the proof was constructed.
How does the AI proof actually work?
Williams's analysis shows the AI proof works by combining three ingredients. The construction stretches a square grid by an irrational factor of 1/√65, so the unit circles centred on lattice points intersect more lattice points. The analysis applies Jacobi's two-square theorem to count those intersections. Research from Will Sawin's extended write-up shows the construction generalises into an infinite family. The data reveals a denser unit-distance graph than the conjectured square-grid optimum.
Why is the result a big deal even after Williams's deflation?
According to Daniel Litt, the result is "the first example of a result produced autonomously by an AI that I find exciting in itself." The answer is direct: the result is the first time a frontier AI system has produced work the working mathematical community judges interesting on its own merits, rather than as a leading indicator. The key is straightforward — the reviewer judgement clears the bar a mathematician's first paper would clear; the verification chain held.
Who is Kai Williams and why does the Understanding AI piece matter?
Kai Williams writes at Understanding AI, the publication founded by Timothy B. Lee that has become one of the most rigorous mainstream venues for AI analysis. Williams's piece matters because the analysis explains the OpenAI proof more clearly than OpenAI did, gives the reader the actual mathematical object, and credits the prior literature the lab's announcement glossed over. In other words, Williams's piece is the kind of research-grade journalism the AI beat needs more of.
What are the limits of this AI mathematics moment?
Analysis of the result demonstrates three durable limits. First, the proof played to the AI's strengths — high-volume search and classical-lemma retrieval — and does not show the system can formulate a new field. Second, evidence from Gowers's walked-back reaction reveals how easily early framings overshoot. Third, the OpenAI announcement itself reveals a lab-PR layer that Williams's piece had to peel back. The answer is that the AI mathematician is real, the worker has a profile, and the profile is narrower than the apocalypse readings claimed.
Sources
Williams, Kai. OpenAI's "milestone" math breakthrough played to AI's strengths. Understanding AI, 28 May 2026. The piece this essay responds to; the three illustrations rehosted here are by Williams (or by Alexeev, Mixon and Parshall under CC BY 4.0, as captioned).
DeepMind Co-Scientist Pitches AI as a Real Research Partner
Google DeepMind's Co-Scientist is a multi-agent research partner built on Gemini, validated across liver fibrosis, ALS, aging, and plant immunity at Stanford, MIT, Cambridge, Edinburgh, and Calico. The first frontier-lab AI pitched as a real scientific collaborator.
