🔮 What is left to do

OpenAI has put out a striking batch of mathematics from an unreleased frontier model: 722 manuscripts spread across 372 families, cutting across number theory, complexity theory, and mathematical physics. The average result took the equivalent of three hours of ChatGPT Pro thinking. That is not a small demo. It is a pile of work that would have taken human teams a very long time to assemble, let alone verify.

Derya Unutmaz says the set amounts to 81% of the major math discoveries made in the past three years. That claim lands hard, even before the caveat. Many of the results have been checked in Lean, but not all of them have. Still, even if only half of the output holds up, the picture is unusual: problems that resisted some of the best human minds for decades were apparently pushed over by one afternoon of compute.

The bigger issue is what happens next. The source points to a future with millions of superhuman research agents working in parallel, while formal systems handle proof checking. Humans may end up with the proofs but not the mental map. The machine can build the web of concepts; people may only see the summary layer on top.

That is where the idea of two mathematicses comes in. One would be machine mathematics: huge, verified, and mostly for other AIs to consume. The other would be human mathematics: a compressed effective theory of the frontier, small enough for us to actually grasp. That split would change more than the research pipeline. It would change who mathematics is for.