Super-cool. I haven't looked at the source, but I bet it's massively parallel, at the pixel level? (I do see you say it's "not a cellular automaton", though...

Super-cool. I haven't looked at the source, but I bet it's massively parallel, at the pixel level? (I do see you say it's "not a cellular automaton", though...) If so, I wonder how difficult it would be to port this to GPU, either writing CUDA yourself, or maybe even using PyTorch. If PyTorch, you might be able to make it differentiable (by changing ANDs to MULs and such). What would be achieved by making the simulation differentiable I don't know. But surely you'd come up with something interesting. Given how much effort has gone into developing architectures whose gradients don't explode, my guess is you'd run into difficulties, but maybe you could even learn programs that model various sequences. The main point would just be to run it on the GPU, though.