Grant Sanderson (@3blue1brown) – AI and the future of math
TL;DR
AI math progress is fractally spiky: AI cold-solves IMO geometry in 19 seconds via brute force but fails at combinatorics, showing that even within a "spike" like math, capability is uneven and benchmark headlines mask significant gaps.
Three ways AI could solve the Riemann hypothesis: finding lightning-bolt connections between fields (human-parsable), building entirely new theoretical mountains (requires digestion), or raw computational hustle (potentially uninterpretable), each with different implications for economic automation.
The Galois theory lesson: Group theory took roughly 100 years from Lagrange's initial insight to modern recognition, illustrating that the verification loop for truly novel mathematical concepts can span generations, making RL-style training fundamentally difficult.
Grindability matters as much as verifiability: Math and coding advance rapidly not just because answers are checkable, but because you can run thousands of parallel rollups in containerized environments, unlike real-world tasks like computer use that hit bot detectors and friction.
Mathematicians may become curators: In an AI-abundant world, the human role shifts toward deciding what ideas are worth pursuing, plus teaching, which remains stable because it's inherently relational rather than just explanatory.
Lean enables infinite exploration: While possibly overrated for current progress, formal proof systems could let AI endlessly extend repositories like Mathlib without human oversight, exploring axiom spaces and potentially discovering novel algebraic structures.
Summary
Grant Sanderson explains why AI's gold-medal IMO performance wasn't an AGI moment: progress is "spiky," with geometry solved in seconds but combinatorics remaining a struggle. The next real benchmark isn't theorem-proving but conjecture-generation and definition-creation, skills that resist easy measurement and took historical insights like Galois theory a century to be recognized as valuable.
Was This Useful?
Share
Keep Reading
Make Alcreon Yours
Tune your feedFive quick questions, and the feed ranks what matters to you first.Or just get notified
The weekly Echo. Signal worth keeping in your inbox.
Every new piece, announced on X.
Read Next
See all
Playbook
The Retirement Email Isn't a Warning
Model retirements now arrive every few weeks; the config-eval-rehearsal loop turns each deprecation email from a fire drill into an afternoon swap.
Playbook
The Cheapest Model That Passes
OpenRouter lists 400 models behind one API. The fix for choosing isn't a better leaderboard, it's a four-step protocol that ends in a real eval.
Playbook
Cheap Models, Hard Tasks
Most agent workflows route every step to the frontier model by default. The bill scales with how chatty the agent gets, even when most steps don't need that brain.