Statements like "X is can be used to represent any program/function, therefore it is all we are every going to need" are pretty silly.
First of all, almost any program-representation medium (except particularly trivial ones) can be used to represent any program, so that property doesn't tell you anything about the merits and limitations of a particular approach.
Second, it is not relevant whether a program *can be represented* (that's almost always true), one care whether a program *can be developed* (not just in theory, but in practice) using a particular approach.
So statements like "NNs are universal approximators (or, 'there are Turing-complete RNN architectures'), therefore DL has no limits" is like saying "you can use a pen & some paper to produce the complete works of Shakespeare". Technically true, but pretty silly at the same time.
