AI bottleneck is mooving from "learning faster" to "accessing what I already know faster" gl, hf