Strange, for you to say this in the age of transformers. They are obviously not there yet, but it seems inevitable to me that e.g. gpt 'understands' things better than a markov chain, and that future systems will understand more.
I know you're responding to someone else, but funny you say that. I suspect modern NN's are just iteratively solving some really large instances of SAT (similar to WalkSAT), to find a representation of a circuit that best matches a function defined on training examples (which are the constraints). So they're, unwittingly, kinda the SotA SAT solvers.
js8|1 year ago