neural nets are non-grid CA with locally-variable neighborhoods (edge weights)
CA evaluations can be made hyperefficient through adjacency matrix compositions
by leveraging the equivalence posited earlier, you can optimize model evaluation
training is still expensive tho