An OpenAI model has disproved a central conjecture in discrete geometry

For eight decades, mathematicians have poked at the unit distance problem, a deceptively simple question about how many pairs of points in a plane can sit exactly one unit apart. It traces back to Paul Erdős and has spawned a small industry of partial results, clever bounds, and one stubborn conjecture that many in the field assumed was probably true, even if nobody could prove it. OpenAI says one of its models just tore that assumption apart.

The model didn't just nudge the bound a little. It found a counterexample, a configuration that violates the conjecture outright, which is the kind of result that forces textbooks to get rewritten rather than footnoted. Disproving something is often harder than proving it, because you need one concrete, verifiable object that breaks the pattern everyone believed held. That the object came from a machine rather than a person doing years of chalkboard work is the real headline here.

This isn't the first time an AI system has poked at open math problems, but most of the attention so far has gone to systems that verify known proofs faster or search huge spaces of small cases. Finding a genuine counterexample to a longstanding conjecture is a different tier of contribution. It suggests these models can generate mathematical objects nobody thought to construct, not just recombine existing techniques faster than a grad student with a laptop.

The unit distance problem sits in discrete geometry, a corner of math that's produced surprisingly hard problems from surprisingly simple setups. Erdős himself loved this territory precisely because the questions are easy to state to a teenager and brutal to actually answer. Whatever bound eventually replaces the old conjecture, mathematicians now have to grapple with a new kind of collaborator, one that doesn't get tired, doesn't need tenure, and apparently doesn't respect eighty years of consensus.