Grant Sanderson

Creator of 3Blue1Brown, the YouTube channel that teaches mathematics through custom-built animation (rendered with his own open-source library, Manim). Sanderson has made visual, motivation-first exposition of topics from linear algebra and calculus to the Riemann zeta function into a widely-followed art form, and is now producing a documentary-style series interviewing mathematicians about AI’s progress in the field.

Core positions

Sanderson treats mathematics as the leading indicator for AI: because progress is fastest there, maths shows ‘very concretely’ what AI progress in other fields will look like. He distinguishes sharply between the kinds of mathematical advance an AI might make — a lightning bolt connecting two known fields (human-parsable, and squarely what LLMs’ breadth should enable) versus mountain building, the invention of genuinely new theory that demands a different order of intelligence — and argues the deepest work, generating good conjectures and definitions, is precisely what resists benchmarking and easy training. His century-long Galois-to-quarks example is a standing argument that a real conceptual breakthrough may pass no verifiable reward for a hundred years, so any training signal should reward compression and elegance, not just solved problems.

On what survives automation, his view has shifted: he no longer expects explanation to be the human redoubt (the models will explain well too), but curation — the mathematician as art-museum curator, and teaching as a relational, coaching role — because motivation to care about ideas is a social phenomenon. His advice to students is characteristically practical: ‘who matters more than what’ when choosing courses and books, use LLMs as a souped-up search for the right human-made resource, and understand where your salary actually comes from and what value you add.

In the wiki

• Grant Sanderson on AI and the Future of Math, Conceptual Breakthroughs, and Human Curation — Dwarkesh Podcast: the anatomy of a mathematical breakthrough, why maths is uniquely grindable, the century-long verification loop, Lean and the endless Mathlib, and curation as the durable human role

See also

• Dwarkesh Patel — the host
• Large Language Models — the technology whose reasoning style (autoregression, RL, verifiability) he dissects