Overview
Sutton argues that AI methods able to exploit growing computation eventually outperform approaches built around human domain knowledge. Falling computing costs change which research strategies succeed; optimizing only for present resources can favor methods that later plateau.
His examples span chess, Go, speech, and vision. Deep search drove chess breakthroughs; Go combined search with self-play learning. Statistical speech models and later deep networks displaced explicit linguistic machinery. Vision similarly moved from designed features toward learned representations.
Research incentives reinforce the problem: adding knowledge can produce immediate, satisfying gains, while creating complexity and commitments that obstruct later scaling. Search and learning offer general mechanisms for using additional computation.
The prescription is to build discovery processes that learn complex representations and useful approximations, rather than manually specifying their contents. His vision example still retains architectural structure such as convolution; engineering does not disappear.
Read this as a historical research argument, not a controlled scaling experiment. It does not establish the best method under a fixed budget or prove that computation alone solves every task.