Gwern — Anonymous writer who predicted AI trajectory on $12K/year salary

In a Dwarkesh Patel interview, Gwern Branwen, an anonymous researcher and writer, gives a compact theory of intelligence: it is search over Turing machines, the formal model of any computation, of many lengths, with no master algorithm and no special intelligence fluid. The video uses an avatar to protect the anonymity, and the words are Gwern's. On this view a large trained network is a gigantic ensemble of small specialized solutions, which Gwern says is why a small model for one task can be pulled out of a large one, and why no single place in the brain removes fluid intelligence when damaged.

Gwern uses the same idea to explain why human-level intelligence is rare: where a small program would do, evolution can hard-code it in genes, which beats a slow, expensive, glitchy general learner, so intelligence pays only in some niches. Differences between people become differences in compute available for search. The interviewer pushed back that clever friends seem to have general horsepower. This is Gwern's interpretation, not a test result.