What it means when an AI disproves an 80-year-old maths conjecture

This morning OpenAI announced that one of its internal reasoning models had autonomously disproved Paul Erdős's unit distance conjecture in the plane — a problem that had been open for roughly 80 years. Timothy Gowers, the Fields Medallist, said he would accept the proof for Annals of Mathematics "without hesitation."¹

I want to walk through what actually happened here, because the headline ("AI disproves famous conjecture") hides almost everything that makes the result interesting. The problem itself is one a child can picture. The mechanism by which a language model produced a new geometric construction is not. And the distinction between "competition maths" and "open-problem maths", which the same lab crossed in under a year, is the bit worth understanding if you want to read the next twelve months of AI-in-science news clearly.

The problem, in one paragraph. Put n dots on a flat piece of paper. Count the pairs of dots that are exactly 1 cm apart. How big can that count get, as n grows? Intuitively it shouldn't grow much faster than n itself — each dot only has so much "room" around it at distance 1. Erdős conjectured in 1946 that the count is at most roughly n, give or take a vanishing correction: formally, O(n^{1+ε}) for any ε > 0.¹ That's the near-linear upper bound. The new result says no, you can do better than near-linear. There exist point arrangements where the count of unit-distance pairs grows like n^{1+c/log log n} — which is bigger than near-linear, and disproves the conjecture as stated.¹

Why this problem is hard. It looks like geometry, but the difficulty is combinatorial. You aren't solving an equation; you're constructing an arrangement of points cleverly enough that an unusual number of pairs land exactly 1 cm apart. The best previous construction, by Spencer, Szemerédi and Trotter in 1984, hit n^{1+c/log log n} as a lower bound.² For 40 years that gap — between the n^{1+c/log log n} you could build and the near-linear bound Erdős thought was the ceiling — sat there. Closing it required either a sharper upper-bound proof (no one managed it) or a construction that broke the conjectured ceiling (no one managed that either). The model did the second.

What "the model found a proof" actually means

Here is where I have to be careful, because OpenAI has not released the model, named its version, or published the construction in detail.¹ So I'm going to explain what general-purpose reasoning models do when they tackle problems like this, based on what we know about the o-series family and similar systems, and flag where the public information runs out.

The shape of the search. A modern reasoning model doesn't "think" the way a human mathematician does, but it also isn't doing brute-force search over all possible point arrangements. What it does, roughly, is generate long internal chains of reasoning — sequences of tokens that propose candidate constructions, evaluate them, notice problems, revise, and try again. The reasoning happens in natural language and mathematical notation, the same medium the model was trained on. When OpenAI's o-series achieved gold-medal IMO performance in mid-2025, this was the mechanism: extended chains of thought, often tens of thousands of tokens long, exploring a problem the way a strong student might if they could think for hours without getting tired.²

For the unit distance problem, the model would have been generating candidate point configurations, described mathematically, not visually, and reasoning about how many unit-distance pairs each one produces. Promising configurations get refined. Dead ends get abandoned. The "proof" is the end state: a construction plus an argument that the construction has the claimed number of unit-distance pairs.

Where the metaphor breaks. It's tempting to describe this as "the model searched the space of constructions." That's partly true and partly misleading. The model isn't enumerating; it's generating, which means it's drawing on patterns learned from millions of mathematical papers, textbooks, and proofs. The space it explores is shaped by what mathematicians have done before. That's the source of both its power (it has internalised a vast amount of mathematical taste) and the legitimate concern Hacker News commenters raised: if extremal constructions for this problem appeared anywhere in its training data, the model may be interpolating rather than discovering.³ I haven't seen this concern resolved in the announcement, and I think it's the right question to keep asking until the paper is out.

That expression is the number of unit-distance pairs the model's construction achieves. It's also, suggestively, the same growth rate as the 1984 Spencer-Szemerédi-Trotter lower bound. From the public announcement it isn't yet clear whether the model improved the bound, matched it with a structurally new construction, or did something subtler that resolves the conjecture as stated.² That clarification will come when the proof itself is published.

Why "general-purpose" is the load-bearing word

Last year DeepMind's AlphaProof solved IMO problems by working in Lean, a formal proof language, with a system purpose-built for mathematical theorem-proving. It was a mathematics specialist. OpenAI describes today's model as a general-purpose reasoning model — the same architecture family that handles code, writing, and everyday Q&A.¹

That distinction matters mechanistically. A specialist system like AlphaProof has its search space constrained to formally valid proof steps; every move it makes is type-checked by Lean. It cannot produce mathematical nonsense, because nonsense doesn't compile. A general-purpose reasoning model has no such guard rail. It generates natural-language mathematical argument, and the argument has to be checked by a human (or a separate verifier) afterwards.

The trade-off used to look one-sided: specialists were rigorous but narrow, generalists were broad but unreliable for proofs. Today's announcement is interesting because it suggests the generalists are now strong enough at maths that the lack of formal guard rails isn't stopping them from producing publishable results. The same model that might draft a memo also, with enough thinking time, produced a construction Gowers is willing to put in Annals.

I want to be honest about my uncertainty here. OpenAI hasn't named the model. It's possible the "general-purpose" framing is doing some work — the model may have received heavy mathematics-specific reinforcement learning that isn't disclosed. Without the model card, I can't tell.

The pace, and what it implies

Nine months ago, OpenAI's o-series got IMO gold. IMO problems are hard, but they're a known genre: every problem has a solution discoverable by a strong undergraduate using techniques in the existing toolkit. The problems test whether you can apply known methods cleverly under time pressure.

An open conjecture is a different category. There is no known solution. The techniques required may not exist yet. Resolving one requires producing a mathematical object, in this case, a point configuration, that nobody has produced before. The jump from "applying known techniques to problems with known solutions" to "producing a new mathematical object that resolves a problem with no known solution" is the jump from competition to research.

The same model that might draft a memo also, with enough thinking time, produced a construction Gowers is willing to put in Annals.

Nine months is a short interval for that jump. I don't want to over-claim what it means — one result is one result, the model is unreleased, and the formal peer review hasn't happened. But if you're trying to read the trajectory of AI in scientific research, the relevant data point isn't "AI did maths." It's "the gap between competition-level and research-level maths, for at least one frontier system, was about nine months."

What to watch

Three things will tell us how much of this generalises:

The paper. Until the construction is published and refereed, we have Gowers' word — strong, but not the same as completed peer review. The proof's structure will also tell us whether the model produced something genuinely new or a cleaner version of the 1984 lower bound.

The model. If OpenAI eventually names and characterises the system that produced this, we'll know whether "general-purpose" survives scrutiny, or whether there's a maths-specialised variant hiding behind the framing.

The second result. One conjecture is a result. Two conjectures, from different areas of mathematics, would be a pattern. The next twelve months will tell us which we're in.

Footnotes and links

Further reading

• Spencer, Szemerédi and Trotter, "Unit distances in the Euclidean plane" (1984) — the construction that held the lower-bound record for 40 years.
• Erdős, "On sets of distances of n points" (1946) — the original paper stating the conjecture.
• DeepMind's AlphaProof announcement (2024) — useful contrast for what a maths-specialist system looks like.

¶