It's interesting to see the path towards increasingly sophisticated question in --> answer out. The way people do this is with dialogue, but that doesn't fit as

It's interesting to see the path towards increasingly sophisticated question in --> answer out. The way people do this is with dialogue, but that doesn't fit as easily into standard supervised learning with an easy-to-collect dataset. If you asked me "What's a good restaurant nearby?" I wouldn't answer with a list of restaurants, I'd say "What kind of food do you feel like?" Seems like we aren't even working our way in that direction. Maybe the sample complexity of RL and language modeling needs to come down a ton first.