What came out of the transcript
The claim was that a reasoning model belonging to OpenAI produced a counterexample to the unit distance conjecture, a discrete geometry problem Paul Erdős posed in the 1940s that had stood unresolved since. The counterexample did not arrive as a submitted proof. According to Gary Marcus's account of the episode, professional mathematicians read through a long transcript of the model's chain of thought, located the counterexample inside it, extracted the relevant portion, and rewrote it as a more succinct proof themselves. [2]
A Hacker News commenter identifying as a mathematics postdoctoral researcher, posting as trostaft, called the result "novel and exciting" and described its adjustments to prior literature as genuinely non-trivial rather than a restatement of known work. That assessment came with a specific complaint: the model's quoted final response did not cite the prior results it built on, so the attribution had to be reconstructed by the humans reviewing the transcript rather than being present in what the model produced. [3]
How much of the work was the model's
Marcus frames the actual contribution narrowly. The model did not hand over a citable piece of mathematics; it applied what he calls superhuman patience to explore lines of reasoning that a human mathematician would likely have abandoned as unproductive, and a person still had to find the useful part and put it into a form a mathematician would recognize as a proof. He compares this to the way computer tools let architects attempt more daring designs without the tool itself being the architect. [2]
Atlas interpretation: That framing matters because of what was withheld rather than what was shown. Marcus points out that OpenAI disclosed no information about how the model was trained, how many unsuccessful attempts preceded this one result, or how the approach performs on other problems: a numerator with no denominator. A single extracted success inside an unpublished transcript, curated afterward by mathematicians who already knew what they were looking for, supports a narrower claim than a model that independently produces citable proofs. The distinction between the two is exactly what a chain-of-thought transcript, read after the fact, cannot settle on its own. [2][3]
Reception and cost
Reaction on Hacker News split along familiar lines. Commenter runarberg compared the approach to a mathematician's ordinary effort by describing it as "kind of just throwing money at the problem," pointing to the compute cost of running a reasoning model long enough to produce the transcript. Commenter isotypic raised a different worry, that treating mathematics as something already sitting in a model's "phantom textbook" waiting to be retrieved would drain the point out of doing the discovery yourself. Commenter shalmanese pushed back that mathematics is a social activity, so a result only counts once the field absorbs it into its working knowledge, regardless of where it originated. [3]
Atlas interpretation: Marcus adds a business observation that cuts against the framing of the announcement as a capability showcase: pure academic mathematics is one of the smaller, less commercially valuable fields a model could be good at. Read that way, a result like this says less about what a frontier reasoning model can do for OpenAI's paying customers than about where its most legible advances currently happen to land, which is math and code rather than the commercially larger markets the company has been building products for. [2]
Sources
- An OpenAI model has disproved a central conjecture in discrete geometry
OpenAI · May 20, 2026
- Checking the math behind OpenAI and Anthropic's latest headlines
Marcus on AI · May 21, 2026
- An OpenAI model has disproved a central conjecture in discrete geometry
Hacker News · May 21, 2026