But the underlying "how to choose a model that's smart enough but not too smart" seems difficult to understand.