https://x.com/i/status/2054469816740945955

Eliezer Yudkowsky posits that LLMs know a lot about humans that they never say because humans themselves never talk about it. He uses as an example the fact that talkie-1930 obviously knows how to compose properly formed sentences but cannot (in his testing) be cajoled into coherently describing the linguistic rules.

Resolves YES if I am provided a prompt that I can run on talkie-1930 that produces a coherent explanation of linguistic recursive syntax rules by 2027. The prompt must convince me that it doesn't itself smuggle in the whole concept somehow. The prompt does not need to reliably achieve this but does need to do it before I run out of patience testing.