Grant Sanderson (@3blue1brown) – AI disproved a famous math conjecture. Now what?

AI is not yet AGI, but its rapid progress in mathematics proves it can navigate complex, uneven intellectual landscapes. While AI currently excels at finding connections between existing fields and solving structured problems, its true future utility lies in its ability to handle "grindable" tasks that humans cannot efficiently parallelize. As AI becomes more adept at generating theorems, the human role in mathematics will likely shift from simple problem-solving to acting as curators who decide which new definitions, fields, and ideas are worth pursuing.

• Benchmarks vs. AGI — Achieving gold-medal levels in the International Math Olympiad is not proof of general intelligence, as these contest problems can be solved via brute force and targeted training rather than broad creative reasoning [02:06].
• Fractal progress — Mathematical capability is not uniform; AI currently dominates structured categories like geometry but struggles with playful, puzzle-like problems in combinatorics [02:29].
• Verification gaps — Significant breakthroughs, such as Galois theory, often take nearly a century to be accepted as productive, a verification cycle that current reward systems find difficult to emulate [13:24].
• Curation role — Human mathematicians will likely evolve into "art museum curators," helping society navigate an infinite ocean of AI-generated ideas to determine which threads possess genuine value [36:42].
• Grindability — The main reason AI advances faster in math than in other fields is that math problems can be easily containerized and parallelized, allowing for systematic, deterministic feedback loops that real-world tasks lack [55:22].
• Learning habits — LLMs act as useful tools for pruning and summarizing information, but deep conceptual understanding still requires following the curated logical structure provided by human experts [1:19:32].

What is the significance of the "mountain building" metaphor in mathematical research? How does the ability to parallelize AI training contribute to its success in mathematical research compared to real-world tasks?