How can we understand the fact that automata behaviour cannot be completely predicted by automata?

This is indeed the realm of Turing, von Neumann, McCulloch and Pitts and others, with implications for current discussions on artificial intelligence and computational limits.