AI did real maths — here's what that does and doesn't mean

Start with what happened. AI from OpenAI and Google DeepMind produced original results on Erdős problems — open questions that are easy to state and brutal to solve, with no memorised answer to crib. No lookup table exists for these, because no human had solved them first. When you crack one, it means actual reasoning. The 'just autocomplete' dismissal dies on contact with the Erdős testbed.

The strongest results lean on Lean — a proof checker that verifies every logical step against mathematical axioms. This is exactly why they're trustworthy. It's also a tell about the shape of the achievement. Lean-verified results come from a generate-then-verify loop: the AI proposes a step; the checker either accepts it or throws an error and feeds it back. The model writes in Lean's formal language; a flawed step is simply rejected. The AI isn't free-soloing genius; it's operating under strict mechanical constraints. That's the innovation — and the limitation, both at once.

Put the two May 2026 results next to each other and the 'who won' framing collapses:

The 'nine to one' scoreboard that went viral is a category error: a machine-certified proof is a different kind of object from a human-checked, peer-review-pending construction. They're not playing the same game, so counting across them is noise.

Here's the part that actually matters long-term: the loop, not the result. A reliable pipeline where AI proposes and a formal system certifies works because, in mathematics, 'correct' is precisely and mechanically checkable. Lean can accept or reject a proof step with no ambiguity. Any flaw is rejected; the model gets the error back and tries again. That tight feedback cycle is what makes the results hold. Most domains don't offer that — strategy, policy, design, the messy middle of science all lack a Lean equivalent. So the achievement is sharply bounded: extraordinary inside a mechanically checkable box, untested and probably far weaker outside it. That narrowness isn't a gotcha; it's the structural cost of the trustworthiness. The certificate comes from the constraint. You can't have both the freedom and the proof.

Hassabis said the system is 'still not AGI'. That's the vendor actively capping the hype on his own headline result. When the person who built it is more cautious than the people tweeting about it, you update toward the builder. AI is now a real instrument for parts of maths — narrow parts, with a formal checker holding the pen — and pretending otherwise just sets up the next disappointment cycle.

So file it correctly: a real research milestone, a narrow one, and a preview of how AI gets useful in serious domains — not by being trusted, but by being checked. The generate-then-verify template is the interesting part. If it spreads to other domains where claims can be formalised, that pipeline — not any individual proof — is the result worth screenshotting. The headline 'AI solves famous maths' is true. The implied 'AI is now a mathematician' is not. The machinery in between is the only part worth your attention.