"Maybe that enormous degree is not just an annoyance but a source of possible counterexamples. Number fields deserve a closer look."
That sentence, lifted from the chain-of-thought of an internal OpenAI model, is the closest thing we have to the moment a machine changed its mind about an eighty-year-old problem. And it is worth sitting with, because it is not a flash of genius. It is a shrug. An annoyance, reconsidered.
On May 20, 2026, OpenAI announced that this model had disproved Paul Erdős's 1946 unit distance conjecture — the first time a prominent open problem central to a subfield had been settled autonomously by AI. Scientific American reportedly declared it the first AI proof that would merit publication in math's top journal even if a human had done it alone. Tim Gowers called it "a milestone in AI mathematics."
Last week in these pages we tracked the capability debate through agentic hacking and inflated demos — machines doing things that looked impressive until you checked. This week the claim is different and heavier: a piece of genuinely new mathematical knowledge, generated in one shot, that no human had produced.
And here is the strange part. The mathematicians brought in to verify the triumph kept circling back to what it wasn't.
Sébastien Bubeck of OpenAI conceded it plainly: "this proof isn't exactly the spark of genius that we see sometimes in mathematics." Thomas Bloom, who maintains the canonical list of Erdős problems, went further. A proof that Erdős was right, he reportedly noted, "would have been truly incredible" — because being right about that upper bound would have demanded totally new insight. The disproof, by contrast, relied on perseverance.
Read that gap carefully, because it is the whole story. The community got its landmark and then spent its remarks explaining that the landmark lacked the one quality mathematicians most prize.
I want to argue that this is not a disappointment to be explained away. It is precisely why the result matters — and why it breaks something.
Consider what actually happened inside the proof. The model brought algebraic number theory to a plane-geometry question, a connection most experts never saw because most experts believed Erdős was right and spent their effort trying to prove him so.
The path it took is, in the companion paper's own accounting, treacherous. It builds a large set of magnitude-one algebraic numbers inside a carefully chosen number field, assumes that field is a CM field for a technical convenience about absolute values, then reaches for an infinite class field tower of Golod-Shafarevich type to keep the root discriminant bounded as the degree runs to infinity.
You do not need to follow that to feel its texture. This is a route down which the "enormous degree" — the thing a human would file under nuisance — becomes the engine of the counterexample.
Jacob Tsimerman put the asymmetry exactly. "AIs have an edge," he said. "They can play for longer and in more treacherous waters than mathematicians without getting overwhelmed."
That is the milestone, honestly stated. Not insight, but stamina — the willingness to walk a tedious, unpromising path with no enticing hint of success, the kind of path a human, largely convinced Erdős was correct, would abandon.
The milestone, honestly stated: not insight, but stamina — the willingness to walk a tedious, unpromising path a convinced human would abandon.
Melanie Matchett Wood noted she wasn't even sure it was a breakthrough for AI; the currently available public models, she thought, might have produced it. One researcher reportedly reproduced the proof with a public model. The genius, if there is one, lies in the inhuman patience, not the intelligence.
So the disruption is not really to insight. It is to something more foundational, and here we arrive at the thing this piece is about.
Look at how the result reached us. Nine mathematicians — Alon, Bloom, Gowers, Litt, Sawin, Shankar, Tsimerman, Wang, Wood — reportedly wrote the companion paper, "a short, digested, human-verified version" of the machine's output.
They reportedly did not see the raw output. As Scientific American recorded in a parenthesis that deserves a chapter: "No external experts have seen the AI's original output — just an edited version of its train of thought."
The proof, the paper tells us in a footnote, was "first mathematically generated in one shot by an internal model," then "expositionally refined through human interactions with Codex." What the world reads is a digestion. A human-readable rendering of something no human read in the original.
And so the nine become, functionally, the proof's authors — the byline on arXiv, the h-indexes attached, the persons the discipline can point to and hold responsible. They are, more precisely, its editors. The author is elsewhere, and has no name.
This is where the credit economy of mathematics starts to buckle, and Wood is the one who names the fracture.
"LLMs have read ALL the papers," she said. "They have read all the commentary and notes, and everything that's online. It's not clear that there's a way for [AI] to reasonably attribute the source of the ideas."
Sit with what that dissolves. Mathematics is not merely a stock of true statements; it is a lineage. Every proof traces its debts — this lemma from Ellenberg and Venkatesh, that tower construction already in the literature, this idea "attributed to Golod-Shafarevich."
The companion paper does this attribution work beautifully, in retrospect, precisely because the machine could not. The reflections credit Ellenberg-Venkatesh, Golod-Shafarevich, Hajir-Maire-Ramakrishna — human names threaded back through decades of human papers.
But the model that actually made the connection cannot say where any of it came from. It read everything, and so it can cite nothing. Its debts are total and therefore untraceable.
This matters because attribution is not bookkeeping. It is what makes mathematics a discipline of persons rather than a warehouse of answers.
The Erdős number — degrees of separation through co-authorship, Einstein sitting at two — is a joke that is also a theology. It says the field is a network of people who handed ideas to each other, and that you can trace any result back through the hands it passed through. Erdős himself paid out cash bounties, sometimes docking the reward when he didn't like the proof: credit as a personal, embodied transaction.
A result whose author read all the papers and can attribute none of them arrives from outside that network entirely. It has no place in the genealogy. It has no hands to trace back through.
So the familiar frames fail. Collaborator implies a shared table and mutual credit; there was no table. Tool implies a human wielding it toward a human's idea; the idea was the machine's, and the humans came after to make it legible. Rival implies a peer in the same game, playing for the same stakes of reputation and citation; the model wants nothing, cites no one, joins no seminar.
What we have instead is an author with no biography. A source of a genuine, checkable, publishable idea that cannot be located anywhere in the human story of how the idea came to be — and the names in the byline are its editors.
Last issue we asked whether "merit" and "the seminar of equals" still hold their meaning under machine acceleration. Here is the test case: a landmark result whose author has no seminar and no peers, and by construction can name neither its teachers nor itself.