Definition
Plain language
A strong automated math-proving system used as a comparison point for cost and accuracy.
As stated in the literature
An automated theorem-proving system benchmarked on competition math; cited as a high-cost baseline (roughly a hundred sixty-eight dollars and ten GPU-days per problem) against cheaper coding-agent-based autoformalization.
Why it matters: It serves as a high-cost reference point, highlighting how much cheaper newer approaches can be while still solving hard math.
For example, Seed-Prover can crack competition math problems but at a steep cost of many days of computing per problem.
Heard on the show
“One strong system, Seed-Prover, solves eighty-six-and-a-half percent at around a hundred and sixty-eight dollars a problem — ten GPU-days each.”Episode 188 — A Coding Agent Found a Hole in a Peer-Reviewed STOC Proof for Five Dollars