One problem is that, even though it is turing-complete, many practical operations are very difficult. Patterns tend towards chaos and they tend towards fading out, which are not good properties for useful computation. Simply moving information from one part of the grid to another requires complex structures like spaceships.

You might have better luck with other variants. Reversible cellular automata have a sort of 'conservation of mass' where cells act more like particles. Continuous cellular automata (like Lenia) have less chaotic behavior. Neural cellular automata can be trained with gradient descent.