> I mean a theoretically absurdly large/wide network can approximate almost any function Yes. But chances are that the network size scales exponentially with t

> I mean a theoretically absurdly large/wide network can approximate almost any function Yes. But chances are that the network size scales exponentially with the number of e.g. input bits. Good luck building/training/storing a model with 2^128 neurons in the hidden layer.