programming for anything than the Von Neumann architecture is very hard.
Generally true.
But neural networks are non-Von Neumann, and we 'program' them using backprop. This can also be applied to cellular automata.