Sorry this is going to be pretty crude, but my intuition is that finding the weights that generate some set of (lowest energy possible) activations across some set of inputs is harder than verifying that it is in fact the lowest energy possible. The fundamental problem is most problems aren't useful