Afaik the more you tighten a bottleneck the more accuracy you lose, and much faster than you gain interpretability. My guess is that such abstractions would req

Afaik the more you tighten a bottleneck the more accuracy you lose, and much faster than you gain interpretability. My guess is that such abstractions would require very powerful "priors" (as in, knowledge stored in the network as opposed to being stored in the representation) that humans gained with evolution and that today's models don't possess.