This is really good. Deep learning right now is giving off a kind of illusion of domain-independent general intelligence that can solve any problem, so it would

This is really good. Deep learning right now is giving off a kind of illusion of domain-independent general intelligence that can solve any problem, so it would be really helpful to have some theoretical characterization of the specific problem domains it's good at and ones it's not good at.