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

Summary

Grant Sanderson discusses why AI progress has been fastest in mathematics and what that reveals about AI's spiky frontier. The conversation explores whether AI will move from theorem proving to conjecture generation, definition creation, and cross-field lightning-bolt connections. They also analyze why math and code are uniquely grindable and verifiable, while writing and computer use remain harder. Market implications are mostly indirect: continued AI progress may create broad economic value, but the episode offers few directly tradeable security calls.

• AI is already gold-level at IMO except on some combinatorics problems; geometry is solved almost instantly.
• The next frontier after proof-solving is generating important conjectures and definitions, which is hard to benchmark.
• The history of Galois theory and Abel's work shows that conceptual breakthroughs can have very long verification loops.
• LLMs may become supercharged connectors across fields over the next few years, aided by parallelization and context reset.
• Math and code progress is driven by verifiability plus grindability; Lean formalization may enable autonomous exploration.
• LLMs are strong explainers and distillers but weak at original writing and at reframing a student's actual confusion.
• Economic leakage from AI math is likely to be incremental in engineering and PDE applications rather than immediately transformative.