An 80-year-old maths problem is gone. Now what?

In 1946, Paul Erdős posed a deceptively simple question about points and distances on a flat plane. It stayed open for eighty years. Then, earlier this year, an OpenAI reasoning model produced a counterexample that disproved it.

Mathematicians checked the work. It held.

The immediate reaction in research circles split roughly in two. One camp found it exciting. Another found it unsettling. Both reactions are reasonable. What interests me more is the third response, the quieter one: confusion about what exactly happened, and what it means for the people whose job it is to think hard about hard problems.

What the conjecture actually was

The Erdős unit-distance conjecture asked how many pairs of points in a set of n points can be at exactly distance 1 from each other. Erdős believed the answer grew slower than any power of n greater than 1. It sounds technical. The intuition is straightforward: pack points on a plane and try to maximise the number of unit-distance pairs. How far can you push it?

For decades, the best progress came from incremental improvements to upper and lower bounds. Real analysts, working by hand and then with symbolic software, chipping away at the gap. The conjecture was considered one of those problems that would eventually yield to a very clever person with a very clever geometric insight.

It did not yield that way. The model produced a construction, a concrete arrangement of points, that exceeded what Erdős thought was possible. No grand geometric insight. No years of accumulated intuition. A search through a structured space, guided by a reasoning process that no one fully understands, including the people who built it.

The understanding question is the wrong question

Every time a model does something like this, the conversation collapses into the same binary: does it really understand, or is it just pattern-matching?

I think that framing is a trap. Not because the question is unimportant, but because it tends to get asked in a way that lets humans off the hook. If the model is "just" pattern-matching, we can dismiss the result as a statistical accident and go back to assuming that genuine insight belongs to us. If the model "truly understands", we slide into the kind of existential hand-wringing that makes for good conference panels and bad decisions.

The more useful question is narrower. what kind of work can this class of tool now do reliably, and what does that change about how humans should spend their time?

In this case, the answer is specific. Reasoning models are now capable of searching large combinatorial spaces and producing valid mathematical constructions in domains where the verification procedure is clean. That is a real capability. It does not require the model to understand mathematics the way a mathematician does. It requires the model to generate candidates that survive formal checking. Those are different things.

What this looks like from an agency perspective

We work with founders and product teams who are trying to figure out where AI actually fits in their work. The Erdős result is useful precisely because it is so clean. There is no ambiguity about whether the output is correct. The verification is mathematical. That makes it a rare, honest data point.

Most of the work we do with clients is not like that. Design decisions, product strategy, user research synthesis: these domains do not have a proof-checker. The model cannot tell you whether a positioning statement is right. It can tell you whether it is grammatical, whether it resembles other positioning statements, whether it avoids obvious contradictions. That is useful. It is not the same as understanding whether the strategy will work.

The gap between those two things is where agencies earn their keep. Not by resisting AI tools, and not by pretending that a model generating a valid mathematical proof means it can run your go-to-market. The gap is real and it is worth naming clearly.

What the Erdős result does change, at least for us, is the credibility threshold for reasoning models on well-structured problems. If the problem has a clear objective function and a reliable verification step, a reasoning model should be on the table as a primary tool. If it does not, the model is one input among several, not a decision-maker.

The researchers who should be worried, and why it is not who you think

Some mathematicians working on combinatorics and discrete geometry will look at this and feel the ground shift slightly. That is fair. If a model can close an eighty-year-old problem in a domain, the nature of open problems in that domain is changing.

But the researchers who should be paying the closest attention are not the ones whose problems just got solved. They are the ones whose methodology depends on a clear distinction between search and insight.

Mathematics has always had a search component. Trying cases, building examples, testing limits. What changed is the scale and speed at which that search can now happen. A reasoning model does not get bored at case 10,000. It does not make arithmetic errors. It does not have a prior about which approach is elegant. Those properties are genuinely useful in combinatorial search. They are not useful in the same way for problems that require a fundamentally new conceptual frame.

The risk is not that AI replaces mathematicians. The risk is more subtle: that the problems AI is good at solving start to look like the only problems worth working on, because they are the ones that produce results. Publication pressure already distorts research priorities. Adding a tool that is very good at a specific kind of problem does not help that.

What it means for the work we do

At Studio Hyra, we do not treat AI as a single capability. We treat it as a set of tools with different strengths in different contexts. The Erdős result sharpens that picture.

Reasoning models with formal verification are now credible partners on well-bounded technical problems. That is a concrete update. We apply it when we scope research tasks, when we advise clients on where to invest in AI-assisted workflows, and when we pressure-test assumptions about what a model can and cannot be trusted with.

What it does not change is the value of people who can frame the right problem in the first place. Erdős asked a good question in 1946. That question guided eighty years of work. The model answered it. Asking the next good question is still a human job, and probably the most important one on the board right now.

So. worried or relieved? Neither, mostly. Paying attention. That seems like the right posture.