CAs are interesting computational systems but it is really difficult to follow the multiway hypergraph rewriting idea and its connection to physics. It sounds reasonable but I am missing a few inductive steps along the way. That may be a me thing... but I also don't know the physics to say one way or another that the outputs he finds match the physical result.
The mutational stuff for parameter search is known already via genetic algorithms which have been employed across the board in optimization, neural networks, machine learning, and even to CAs.
Exploring the space of all 'CAs' (or even the ruliad) by enumerative exhaustion is somewhat interesting but is it similar to defining a FSA by the exposition of all strings accepted by some FSA? If so that seems to be a waste of time. It seems to be a CS version of reading the tea leaves (tasseography). But I may be missing something here as well.
I think this is why he's moved on to "evolution" i.e. mutation with selective constraints (evolution). That seems to be the only way to find something somewhat automatically but it also seems that there may be an infinite number of paths... however, the domain of these functions would be limited by the size of the 'genome' (in classical genetic algorithms) or in the functional inputs of the CA. CA are generally restricted to neighbors and all information is propagated via local neighbors. This limits your rule search space and defines your state space but you could easily have CAs that depend on all neighbors (global state). NNs carry global state and early single layer perceptron's didn't scale until they were made deep (and given additional explicit structure like convolution or transformation)... yes they are all computationally equivalent but one set aren't very useful (even for us trying to understand/reason about) and others result in ChatGPT...
The mutational stuff for parameter search is known already via genetic algorithms which have been employed across the board in optimization, neural networks, machine learning, and even to CAs.
Exploring the space of all 'CAs' (or even the ruliad) by enumerative exhaustion is somewhat interesting but is it similar to defining a FSA by the exposition of all strings accepted by some FSA? If so that seems to be a waste of time. It seems to be a CS version of reading the tea leaves (tasseography). But I may be missing something here as well.
I think this is why he's moved on to "evolution" i.e. mutation with selective constraints (evolution). That seems to be the only way to find something somewhat automatically but it also seems that there may be an infinite number of paths... however, the domain of these functions would be limited by the size of the 'genome' (in classical genetic algorithms) or in the functional inputs of the CA. CA are generally restricted to neighbors and all information is propagated via local neighbors. This limits your rule search space and defines your state space but you could easily have CAs that depend on all neighbors (global state). NNs carry global state and early single layer perceptron's didn't scale until they were made deep (and given additional explicit structure like convolution or transformation)... yes they are all computationally equivalent but one set aren't very useful (even for us trying to understand/reason about) and others result in ChatGPT...