OpenAI Just Disproved a Geometry Conjecture. Both Reactions Are Wrong.
A model produced a counterexample to a long-standing open problem. Tech Twitter is split between "the singularity is here" and "it's just brute search." Both are missing what actually changed.
This week, OpenAI announced that one of their models produced a counterexample to a long-standing open problem in discrete geometry. The reaction online is split, and I think both sides are reading it wrong.
I have seven IEEE-published papers. I know what it feels like to read a peer's proof and realize the argument is genuinely new. So I want to take this one seriously — both more seriously than the hype crowd is taking it, and more seriously than the dismissive crowd is taking it.
Two narratives running through tech Twitter
Narrative 1: "AI did original mathematics. The singularity is here."
The framing is intoxicating. A machine produced a result the field hadn't produced. Headlines write themselves: AI Does What Humans Couldn't. Mathematicians Reckon With Obsolescence.
But the model didn't propose the conjecture. It didn't ask whether the conjecture was true. It didn't construct the conceptual scaffolding that made a counterexample worth pursuing. It produced — within a search space carved out by humans, with an objective function specified by humans, with verification done by humans — a solution.
That's still extraordinary. But it's the same shape of extraordinary as AlphaGo beating Lee Sedol. The thing the machine did is hard. The thing the humans did before and after the machine ran is also hard. Both stories matter.
Narrative 2: "It's just brute search. Nothing new is happening."
The dismissive camp says: this is just better search. The counterexample lives in a high-dimensional but finite combinatorial space. A sufficiently large compute budget plus a smart enough objective function was always going to find one eventually.
This is also too easy. The reason the problem stayed open isn't that nobody had compute. It's that nobody knew what to search for. The model plus the team around it didn't just throw FLOPs at the wall. They reframed the problem.
Three honest observations
1. The class of solvable problems just got bigger.
Some open problems are open because we lack the conceptual tools to even attempt them. Others are open because the search space is genuinely large and our heuristics are weak. The second category is now smaller than it was last month. That's significant. It's not the same as the singularity.
2. The bottleneck moves to problem selection.
If finding counterexamples becomes cheaper, the scarce resource becomes knowing which conjectures to point a model at. That's a senior-mathematician skill, not an LLM skill. And it just got more valuable, not less.
3. Verification is the new chokepoint.
A counterexample is interesting only after it's been verified. The peer-review apparatus is now going to be increasingly important — and increasingly stretched. We should be watching how the field absorbs that load.
The honest read
What just happened isn't the end of mathematics. It's also not nothing. It's a real expansion of a real capability, in a real corner of a real field.
The honest read sits between the two loud narratives. Don't let either tribe tell you otherwise.
What's your take?