I think what he's saying is that currently we can get one of these expert systems to create 1000 examples with the encoded rules, which a NN might then generali

I think what he's saying is that currently we can get one of these expert systems to create 1000 examples with the encoded rules, which a NN might then generalize correctly or it may come up with a completely batshit insane ruleset that overfits on that data. There's no way for us to actually feed the NN the rules directly, like one would give them to a human brain.