OpenAI has officially ended the debate over whether large language models are capable of anything beyond simple next-token prediction. The company’s researchers recently published a counterexample to Paul Erdős’s unit distance conjecture—a landmark problem in discrete geometry that has remained unsolved since 1946. According to the OpenAI preprint, for a fixed δ > 0, the number of point pairs in a plane at a unit distance can exceed n to the power of 1 + δ for an infinite set of n. This directly contradicts the classical limit proposed by Erdős nearly 80 years ago.
The technical execution represents a case of algorithmic dominance over human intuition. To achieve this result, OpenAI researchers constructed an infinite unramified class field tower with 3-step Galois groups of increasing degree. The method is rooted in Golod–Shafarevich theory, which guarantees the existence of such a tower even after factorization that nullifies specific Frobenius classes. Essentially, the researchers replaced standard Gaussian integers with a complex field extension where the base field degree tends toward infinity. Using multidimensional lattices and Dirichlet’s principles for class groups, the model identified elements that maintain an absolute value of one under any complex embedding.
This breakthrough is significant not just for mathematicians, but for every business investing in R&D. We are witnessing a transition from the mere interpolation of training data to the generation of fundamentally new knowledge in abstract structures where the human brain often falters. Reasoning models are evolving from sophisticated autocompleters into autonomous tools capable of finding counterexamples and breaking through theoretical deadlocks. This isn't just process optimization; it is a signal that AI is ready to act as a full-fledged scientific researcher in high-abstraction tasks. The ability to solve problems of this complexity proves that the potential of LLMs is no longer tethered to the boundaries of their datasets.