What made OpenAI's unit distance counterexample succeed?

The remarks present a "short, digested, human-verified version" of an OpenAI internal model's counterexample to Erdős's unit distance conjecture, the conjecture that n points in the plane determine at most n^{1+o(1)} unit distances. Their Theorem 1.1 gives point sets with at least |P_i|^{1+ε} unit distances for some fixed ε > 0, which refutes the conjecture. The abstract says the argument "relies crucially on ideas that may, at least in retrospect, be attributed to" Ellenberg–Venkatesh, Golod–Shafarevich, and Hajir–Maire–Ramakrishna. The remarks locate the novelty more narrowly: the construction of Golod–Shafarevich towers with infinitely many split primes "already appears in the literature", and "a novel ingredient of the AI argument is to take [K : Q] → ∞."

The mechanism runs through two lemmas. Lemma 2.1 turns a lattice with many unit-modulus points in its polydisc into a planar point set with many unit distances, provided the count of those points grows fast enough relative to the lattice's skewness. Lemma 2.2 supplies the unit-modulus points by a pigeonhole argument in a CM field. Golod–Shafarevich towers keep the root discriminant bounded as [K : Q] → ∞, so a fixed split prime "drowns out the main enemies, the class number h(K) and discriminant Disc K." The AI's chain of thought, quoted in the introduction, points the same way: "Maybe that enormous degree is not just an annoyance but a source of possible counterexamples. Number fields deserve a closer look."

The excerpt keeps two operations apart. The construction is "human-verified", so the authors checked the AI's argument; the remarks are also "digested", a somewhat simplified and generalized version a human can follow. The excerpt does not say how well humans followed the original AI proof, and its own need for simplification is the only evidence about that. Where the reflections turn to understanding, they concern why the search succeeded. Noga Alon writes that "the AI was able to do here what lots of excellent human researchers tried and failed to do," and holds this "with or without a full agreement with these reasons" that colleagues have offered. Thomas Bloom calls the result "both surprising and impressive." On this excerpt, the gap between correct and understood sits in the explanation of the success more than in the proof.

Set against the nearest notes, this is a different case from Can identical outputs hide broken internal representations?: there a benchmark can hide broken internals, while here a checkable proof confirms correctness without explaining why the search worked. It shares with Do foundation models learn world models or task-specific shortcuts? the pattern of right outputs without the general mechanism, but the output here is an argument people can audit. It also differs in what counts as validation: the Can AI systems improve themselves through trial and error? note describes replacing proofs with benchmarks, and this result returns to a human-checked proof.

The excerpt does not establish several things. It gives no value for ε in Theorem 1.1, and Lemma 2.1 leaves u, v, and δ unquantified beyond the condition u > 36v/π, so the power-law gap is shown in principle, not measured. It omits the AI's transcript and describes no verification procedure beyond the authors' own label "human-verified." It also leaves open the explanation Alon himself hedges. At the strength the evidence allows, the implication is narrow: a human-checkable argument for a fixed power-law gap exists in these sections, its new move is the growing field degree, and whether this kind of AI success generalizes to other open problems is a separate question the source does not answer.