The Unit Distance Conjecture and AI for Math

TL;DR

• The result: about two months before the talk, OpenAI announced that its model disproved Erdős's unit distance conjecture, constructing point sets with more than a linear number of unit-distance pairs. The problem, posed by Erdős in 1946, asks whether n distinct points in the Euclidean plane can contain at most roughly linearly many pairs at distance one.
• How it happened matters more: the discovery came from a general reasoning model not trained specifically for math — essentially a version of ChatGPT — during an evaluation on very hard math problems (necessary because most math benchmarks are saturated). There was no harness: one prompt, the model keeps thinking, uses tools as needed, with access to the web and a terminal.
• The construction: Erdős's original was an n×n grid with a carefully chosen scaling. The new one is a grid over the rationals working in the field L(i), where L is a carefully engineered high-degree real number field. The model's own chain of thought explained the intuition: the degree and height of the algebraic objects can be enormous, and that's actually a good thing.
• Where the capability comes from: success probability rises monotonically with test-time compute, and the chain of thought shows the model introspecting — noticing it isn't making progress or has made a mistake, then self-correcting.
• His difficulty estimate: roughly 6 hours of top human time — assuming that human already knows all of mathematics. The proof isn't deep; the hard part is being both a very good algebraic number theorist and a very good discrete geometer, and learning a new area takes about two years. AI's unique advantage is breadth, not depth.

Key Points

Problem and result (~01:06–01:08)

He opens by calling this "more like yesterday's news" — the slides were made a couple of weeks earlier, and in the meantime they've released ten more proofs on major open problems — but the discovery is still worth reviewing.

The unit distance problem is simple to state: does every set of n distinct points in the Euclidean plane contain at most roughly linearly many pairs that are distance one apart? Paul Erdős proposed it in 1946, and it's one of the central questions in discrete geometry. Roughly two months before the talk, they announced that an OpenAI model had disproved the conjecture, exhibiting a construction with a superlinear number of unit-distance pairs.

How the discovery was produced (~01:08–01:09)

This is the part worth writing down. It was achieved by a general reasoning model, not trained specifically for math — completely general, basically a version of ChatGPT. It surfaced during a serious evaluation of the model on very hard math problems, which they need because most math benchmarks are saturated and only hard problems still measure ability. And there was essentially no harness: one prompt, the model keeps thinking, maybe uses some tools, and arrives at a final answer. It had access to the web and a terminal, so it could look up what had been done before.

Old and new constructions (~01:09–01:10)

Erdős's original construction was an n×n grid with a carefully chosen scaling so that many pairs sit at distance one. The new one is more complicated: a grid over the rationals, working in a field L(i) where L is a carefully engineered high-degree real number field.

He doesn't go deep into the details, but quotes the model's own stated intuition from its chain of thought: the degree and height of the algebraic objects can be enormous, but that can actually be a good thing — and the model chose the number field on that basis so the construction goes through.

Test-time compute and self-correction (~01:10–01:11)

After obtaining the solution, they studied how the model's success probability correlated with the test-time compute spent on the problem: the more compute, the higher the success rate. Thinking longer definitely makes it more likely to find a correct solution. Looking at the chain of thought, the model can introspect — realize it isn't making progress in one direction or that it has made a mistake — then self-correct and keep going.

Measuring AI progress (~01:11–01:13)

He borrows METR's plot: years on the x-axis, and on the y-axis the amount of human time a task would take, for tasks the agent can reliably complete at that point in time. The line roughly doubles every 6 months — though he notes METR's graph is mostly in the coding domain.

The math version of the same scale: about two years ago models could solve AIME problems, maybe 20 minutes for a very good high school student. A year later, IMO problems — considerably harder, maybe 90 minutes for top high schoolers. Another year later, a long-standing open problem in discrete geometry.

The six-hour caveat: breadth, not depth (~01:13–01:15)

So how much human time is the unit distance result worth? His claim: roughly 6 hours of top human time — assuming that human already knows all of mathematics.

The reason is that the proof isn't deep. What it requires is taking ideas from a very distinct field and applying them back to discrete geometry. You'd have to be a very good algebraic number theorist and a very good discrete geometer; if you're strong in both, it probably takes a day or two to push through to the conclusion. The challenge is mastering two distinct areas — and learning an area takes something like two years.

That's where AI has a genuinely unique advantage: broad familiarity across all mathematical fields, and decent competence in many subfields. Finding this kind of example doesn't require extreme human time if the human already has the right expertise — but it does if they must first learn the area. Progress on unit distance reflects a multiplied effect of both strengths: thinking longer, and better domain knowledge.

He adds an honest caveat: some mathematical results require years of work from a mathematician pushing deeply on one topic, and it's unclear whether we're at that point yet with AI.

After unit distance they released 5.6, which is very strong at mathematical reasoning, with many public results in the two weeks since; and the day before the talk they announced ten more results across mathematics and theoretical computer science.

Consequences as he sees them (~01:15–01:17)

AI is especially good at connecting distant fields of research — unit distance is itself an example. Mathematicians will be empowered to use ideas from fields less familiar to them, because AI can find the relevant ideas and explain them well.

Mathematics was never just problem solving. He makes a point of saying that the most important human input to the OpenAI unit-distance solution was Paul Erdős, who posed this beautiful question with a simple statement and a very deep solution. AI found a solution, and a human understood it. Since the release, at least five or six follow-up works have appeared from different groups around the world. Because examples are now easy to find, iterating on which deep questions to ask becomes much easier.

His closing analogy: the ability to do precise calculation used to be a big part of math, and calculators and computers empowered mathematicians to do more. Since coding models like Codex and AlphaCode, programmers actually spend more time coding, because there's more they can do once empowered. He expects mathematicians will similarly be empowered to do more math — freed from tedious calculation to focus on high-level ideas, or simply to enjoy the process.

Quotes

"This is achieved by a general reasoning model which is not trained specifically for math. It's completely general. It's basically a version of ChatGPT." (~01:08)

The load-bearing claim of the talk: the breakthrough came from general capability, not a math-specialized system.

"Essentially there's no harness. You just give the model one prompt and then it keeps thinking about things and maybe use some tools and then it arrives at the final answer." (~01:08)

A striking sentence at a summit otherwise full of harnesses and orchestration layers.

"The challenge is really to master two distinct areas. So, in that sense, AI has this very unique advantage here." (~01:14)

The advantage isn't being smarter than a human; it's not needing two years to learn a new field.

"The most important human input to the OpenAI solution to unit distance is actually Paul Erdős, who really posed this beautiful question." (~01:16)

Posing the question is still human work; AI found a solution and a human understood it.