Why Some AI Skills Compound and Others Stall: Grant Sanderson's 'Grindability' Thesis

Math, code, and board games keep getting easier for AI systems to master. Computer use, robotics, and open-ended knowledge work do not. That split, increasingly visible across the industry's benchmark charts, is the starting point for an argument made by Grant Sanderson, the mathematician and educator behind the YouTube channel 3Blue1Brown, during a recent appearance on Dwarkesh Patel's podcast.

Sanderson's claim, highlighted in a recent opinion piece from The Sequence newsletter, is that verifiability alone doesn't explain why AI improves quickly in some domains and crawls in others. The missing variable, he argues, is "grindability": how cheaply and repeatedly a model can practice a task, fail, get a clean signal back, and try again at scale. Chess, Go, competitive math, and programming share this property. A move is legal or illegal, a proof holds or it doesn't, code compiles and passes tests or it fails. Not only can these outcomes be checked automatically, they can be checked millions of times per hour without a human in the loop and without the world pushing back in unpredictable ways.

Robotics and computer-use agents look superficially similar, since success or failure is often just as checkable in principle. But the practical cost of generating that feedback is far higher. A robot arm interacting with physical objects can't be reset and retried at the same throughput as a chess engine playing itself. A computer-use agent navigating a live interface runs into brittle UIs, non-deterministic states, and consequences that are harder to simulate cheaply. Open-ended knowledge work compounds the problem further: even when an answer is "correct," there's often no crisp, low-cost signal confirming that it is, which makes reinforcement-style training loops far less efficient to run.

This reframes a debate that has largely centered on whether a task's outputs can be verified as right or wrong. Sanderson's point is that verifiability is necessary but not sufficient. A task can be perfectly verifiable and still resist rapid AI progress if the feedback loop is slow, expensive, or hard to simulate at scale. That distinction helps explain why coding assistants have advanced so fast relative to physical or judgment-heavy tasks that seem, on paper, no less structured.

The observation surfaced in The Sequence, a newsletter that has covered AI research and industry trends for more than five years and has operated without sponsorships for over two years, a detail the publication notes as relevant context for its independence in evaluating such arguments.

For teams building AI products, the implication is practical rather than philosophical. Betting on rapid capability gains in a given domain means asking not just "can we verify success here?" but "can we generate that verification cheaply and repeatedly, at the volume training requires?" Domains that fail the second test, however verifiable they look on paper, may keep disappointing regardless of how much compute gets thrown at them.