I'd say that gradient descent and deep learning have made obsolete whatever cellular automata stuff Wolfram has come up with. Neural nets can approximate any function. Cellular automata has a much more limited range that only includes the generation of some interesting pictures and not much else of interest, IMHO.

To be charitable, if he had some way to mathematically slide between and combine networks of cellular automata, he might have something, but it would probably just end up reinventing deep learning and neural nets on a slightly different set of primitives.