OpenAI Just Solved an 80-Year-Old Math Problem. Should Mathematicians Be Worried?

AI changed the scale of mathematical attention

On Wednesday, OpenAI announced something that sounded almost impossible: one of its latest reasoning models had helped solve an 80-year-old problem in discrete geometry. Not a toy problem. Not a classroom exercise. Not a puzzle dressed up as research. A real mathematical problem connected to Paul Erdős, one of the most prolific and beloved mathematicians of the twentieth century.

Listen to the episode on Apple here or YouTube:

The question is deceptively simple: if you place a bunch of points on a plane, how many pairs of those points can be exactly one unit apart? This is called the planar unit distance problem, and Erdős first posed it in 1946. For decades, mathematicians believed that the familiar square grid was, in some asymptotic sense, the best possible arrangement. It was elegant, intuitive, and stubbornly difficult to beat.

Then OpenAI’s model found a way to beat it.

In this episode of Breaking Math, Noah and I spoke with Daniel Litt, a professor at the University of Toronto who was involved in the OpenAI project, to understand what actually happened. The result was not magic, but it wasn’t trivial. As Daniel explains, the model found an infinite family of examples that improves on the classic grid construction by reaching into algebraic number theory — specifically, ideas involving Gaussian integers, prime splitting, and class field towers.

That last phrase may sound like math wandered into a cathedral and started building scaffolding, but the big idea is this: a geometry problem became a number theory problem. Erdős’ original grid construction was already secretly using the arithmetic of sums of two squares. OpenAI’s model found a more powerful number-theoretic structure that allowed the construction to go further. In other words, the model did not simply “calculate harder.” It recombined existing mathematical ideas in a surprising way.

Daniel was careful not to overhype the result, but also not to downplay it. This is serious mathematics. He described the result as potentially worthy of a top mathematics journal. But the real shock is not merely that AI helped produce one clever construction. It is that a general-purpose reasoning model, not a specialized geometry machine, was able to make progress at the frontier of mathematics by being patient, broad, tireless, and willing to try many problems at once.

Human mathematicians are rare. Their time is expensive because curiosity is deep, but limited by human life, human fatigue, human incentives, and the fact that nobody can seriously attack a thousand open problems in a weekend. Models, on the other hand, do not get bored. They can be pointed at vast collections of problems and occasionally return something meaningful.

But this does not mean human mathematicians are obsolete. In fact, the episode argues almost the opposite. AI can generate plausible arguments, but humans still need to verify them, contextualize them, assign credit, understand their lineage, and decide what they mean. Daniel points out that mathematical proof is not just about getting a correct answer. It is also about human understanding, intellectual history, institutions, training, and the development of mathematical judgment.

That may be the most important takeaway. AI can produce a proof, but it cannot understand something for you.

For students, researchers, teachers, and anyone who works with mathematical ideas, the message is not “panic.” It is “adapt.” Learn to use AI as a tool, its strengths, and failure modes. We need to learn how to check it, challenge it, and use it without surrendering your own thinking. Because the future of mathematics may involve more machine-generated conjectures, proofs, constructions, and counterexamples — but it will still need people who can ask why any of it matters.

The AI mathematician has arrived. But the human mathematician is not leaving.

The job and tools are changing.

And as always in mathematics, the real story begins when someone asks: “Wait… is that actually true?”