I keep going back to this paper. OpenAI dropped it on the 20th, and I’ve been chewing on it for a week without writing it up because I wanted to make sure I understood the proof before I tried to.
The setup, in one sentence: an internal OpenAI reasoning model autonomously disproved Paul Erdős’s 1946 unit-distance conjecture by exhibiting an infinite family of point configurations using infinite class field towers, beating the long-standing square-grid construction by a polynomial factor.
That sentence contains roughly four separate things that have never happened before in machine-generated mathematics. Let me take them in order.
What Erdős asked
Place n points on a plane. How many pairs of those points can be exactly distance 1 apart from each other?
Sounds simple. The “obvious” construction is a √n × √n square grid, which gives you roughly n × (some logarithmic factor) unit-distance pairs. For eighty years, the best known constructions were variants of the square grid, all giving the same essential shape.
The conjecture, in its strong form: the square grid is essentially optimal — you can’t asymptotically do meaningfully better.
That’s what the model disproved.
What the model produced
A construction that exceeds the square grid by a polynomial factor, asymptotically. Not a constant-factor improvement. A genuinely better growth rate.
The route it took to get there is the headline:
- Instead of staying in geometry, the model re-cast the problem in algebraic number theory.
- It used infinite class field towers — specifically the Golod–Shafarevich machinery from the 1960s.
- It produced an infinite family of configurations, parametrised by the structure of those number-field towers.
- The proof has been checked by external mathematicians, with a companion paper by Tim Gowers giving the human-facing exposition.
This is not a model that ran an exhaustive search and found a counterexample. It is a model that noticed a mathematical bridge — between a discrete geometry problem and a deep result in algebraic number theory — that human mathematicians had not crossed in eighty years of effort on this conjecture.
Why “AI mathematics arrived” is not hyperbole
There have been a lot of AI-mathematics announcements. AlphaProof, AlphaGeometry, the OpenAI IMO results, the various Lean-formalisation efforts. Most of them are extraordinary engineering. Almost none of them have produced a result that working mathematicians would cite for its mathematical content, independent of the fact that an AI produced it.
This is the first one that does.
Three reasons:
1. The conjecture was open. Not a re-derivation, not a new proof of a known result. The conjecture was unsolved. The result is new mathematics.
2. The technique is non-obvious to humans. The Golod–Shafarevich connection is not the obvious tool. It’s the kind of connection that working number theorists know about but discrete geometers typically don’t. The model crossed a sub-disciplinary boundary that most human researchers would not have. Tim Gowers’s companion paper essentially says: I would not have looked there.
3. Gowers signed off. Fields medallists do not put their names to AI-mathematics papers casually. The “milestone in AI mathematics” line is doing a lot of work — it’s a statement of validation from someone who has been deeply sceptical of AI-mathematics overclaiming for a decade.
What the proof tells us about the model
We don’t have weights or architecture details — OpenAI is being characteristically opaque about which internal model produced this — but the shape of the result tells us a few things.
It’s a reasoning model, run for a long time. This is not a one-shot inference. The proof requires sustained mathematical reasoning across multiple sub-fields, and the kind of multi-day compute budget that only the internal-tier reasoning systems currently have.
It has access to mathematical tools. Some combination of formal verification (Lean is the obvious candidate), symbolic manipulation, and probably literature retrieval. The Golod–Shafarevich theory isn’t something you’d reconstruct from first principles in a reasonable time budget.
It’s cross-domain. The single hardest thing about machine mathematics has been bridging sub-disciplines. The model demonstrably did that.
The slightly uncomfortable implication
If this generalises — and it’s one result, the base rate on “single result that generalises” is low — then the floor on machine-assisted research has just moved. The picture changes from “AI helps mathematicians with bookkeeping and search” to “AI proposes proof strategies that humans then verify.” That’s a different research workflow. It’s the workflow Gowers’s companion paper is implicitly endorsing.
The careful read: this is one open problem, in one corner of combinatorial geometry, where one specific algebraic-number-theory connection happened to apply. It’s not yet evidence that arbitrary open problems are now tractable.
The less careful read, and the one I find harder to dismiss the more I sit with the paper: this is the first credible existence proof that a frontier LLM, given enough thinking time and the right tools, can produce mathematics that working mathematicians find genuinely novel. The base rate of that happening, before this paper, was zero.
It is no longer zero.
What I’m watching
Three things, in priority order:
-
Replication. Does the same model — or another lab’s model — produce a second result on a different open problem, on a similar timescale? One result is an existence proof. Two is a trend.
-
The mathematicians. How do the working number theorists and discrete geometers respond over the next six months? The Polymath-style follow-ups, the simplifications, the generalisations. If the result catalyses further human work, the floor has moved durably.
-
The methodology paper. OpenAI will, eventually, publish on the reasoning system that produced this. The architecture and prompting details are the artefact I want most. If the answer is “we let a reasoning model think for a week with tool access,” it changes how every lab budgets for the next twelve months.
I think this paper will be cited in the AI history books. Not because of the unit-distance problem specifically — most readers will never need to care about that problem — but because of where the technique came from and what it implies. The math is the lever. The lever is what just moved.