Machine learning is incredibly good at brute forcing better-than-human outcomes in path-dependent situations.
The software can be set up to run more iterations in a week than you could in a lifetime. E.g. it can play literally millions of games of Chess, Go etc.
Conversation
This does not magically translate into fluid learning unconstrained by human limits in other environments.
E.g. grammar is a complete and utter clusterfuck, and we are still terrible at teaching machines natural language.
1
2
This doesn't necessarily reflect a mistake in methodology or a hardware restriction per se, sort of some sci-fi uberquantumcomputer.
It simply isn't an easy thing to model, nor to simulate in a way that provides reliable feedback.
1
1
Sooner or later, someone is going to figure out ways to get around these constraints - or to make other technological innovations that obviate them.
But it's not a given that it's happening now, or soon.