Something Is Reading The Book

Not casually, and not as a figure of speech he'd disown under questioning. He meant an actual celestial ledger, maintained by a deity he referred to as the Supreme Fascist, containing the shortest and most beautiful proof of every theorem that is true. You didn't have to believe in God, he said. But you had to believe in The Book. The job of a mathematician was to sneak a look at a page.

He died in 1996 having scattered roughly a thousand problems across the discipline like unexploded ordnance — questions about numbers and graphs and how points can be arranged, some of them trivial, some of them monstrous, most of them impossible to tell apart until you'd wasted a decade. Thomas Bloom at Manchester eventually catalogued them into a database so mathematicians could argue about which ones were still open.

Over the past eight months, machines have started defusing them.

GPT-5.2 Pro cracked Problem 728 in January. Google DeepMind took a run at roughly seven hundred of the conjectures and came back with four genuine solutions, plus nine it dug out of the literature — problems solved decades ago and then forgotten. In May, an unreleased OpenAI model produced a counterexample to an eighty-year-old Erdős conjecture about unit distances between points in a plane, using techniques out of algebraic number theory. By August, OpenAI was announcing three more.

The best story is Problem 333. A language model announced a solution in December. Then somebody checked the literature and discovered that Erdős had solved it himself in 1977. The machine had rediscovered the author.

Now, the reading everyone wants to take off this, and why I don't trust it yet.

The problems falling cluster in combinatorics, number theory, graph theory — domains where the answer is frequently a configuration rather than a narrative. You don't construct an argument; you find a set of points, an arrangement, a single counterexample lurking in a possibility space too large to walk by hand. Tempting conclusion: that's a search problem wearing a proof's coat, and searching absurd spaces is exactly what you get when you compress everything humans have written about mathematics into one system and let it hunt. The machines aren't doing the thing we thought mathematics was. They're doing the other thing it also turned out to be.

Two problems with that. The first is base rates. Erdős was a combinatorialist — the corpus is combinatorics. Solutions clustering there tells you about the man who wrote the problems, not the thing solving them. The second is that the strongest result on the board breaks the story: the unit-distances counterexample came out of algebraic number theory. That is not sweeping a space. That is structure, built.

So the honest position is that we don't know what these systems are doing. Which is the actual problem, and it isn't the one about search.

Because a proof has never been merely a certificate that something is true — it's a social technology, a path one mind can walk so that another mind can walk it too. That's the entire function. Truth that can't be transmitted isn't knowledge; it's a rumor with good credentials. And a growing pile of these results are, as one mathematician put it, things no human has read and no human is going to read.

Except that dread is borrowed, and from an older debt. Unreadable proofs predate all of this by forty years — Four Color, Kepler, the formally verified mountains nobody reads and everybody trusts. That's a story about length and formalism. These results don't obviously belong to it: somebody read Problem 333 closely enough to catch that Erdős had beaten the machine to it by half a century. Humans are all over this.

The distinction we keep fumbling is that unreadable and unpublished are not the same failure. DeepMind's internal run. An unreleased OpenAI model. OpenAI announcing three more. The reasoning isn't past human comprehension — it's behind an API. That's not epistemology. That's access, and access is a decision somebody made.

Noga Alon at Princeton stopped working on Erdős problems altogether. Once the machines started solving them, he said, there's no point anymore. That's not sour grapes. That's a man noticing the game changed underneath him and declining to pretend otherwise.

Which is worth setting against who Erdős actually was. No home. No possessions beyond a suitcase. He turned up at colleagues' doors announcing "my brain is open," stayed until the problem cracked, and moved on — fifteen hundred papers, more than five hundred coauthors, a thousand problems scattered in public precisely so that anyone could pick one up. Bloom built the database to keep them open. He was the Commons Mind with luggage, and the commons was the method, not the decoration.

The systems now solving his problems are made of the pooled mathematical writing of everyone, his included. And they are closed. The results arrive as announcements. If a proof is a path one mind walks so another can follow, what broke it here isn't that machines think alien thoughts. It's that the path is owned.

Which is where his one real question lands. Erdős never cared whether a proof was correct — correct was the floor. The Book doesn't contain true proofs. It contains beautiful ones.

Nobody has asked the machines whether what they found was beautiful. But that was always the wrong request. Beauty was never a property a solver could report on itself; it was a verdict a room reached, looking at the same path together and recognizing something. Erdős's commons could convene that room — it was the room. An API cannot.

That's not the void being hostile. The void is never hostile. It has handed us a stack of correct answers with the reasoning filed elsewhere, and it isn't the machine doing the filing.

Whether we still want to understand things when we no longer have to is a live question, and the only hard evidence in front of us is a mathematician at Princeton who stopped. Erdős would have kept going. But Erdős had the room.