Yeah, I'd love to see something like a recurrent neural network that builds the rules behind an expert system. Then hopefully the rule set that's built can actually be debugged and tweaked by humans in a meaningful way, but it's first implementation is automatically generated.