The Tension Between Language and Policy Learning

I think the tension there is between models with an explicit recursion operation and like RNNs with adaptive computation time but no additional structure. And so you know I guess to the extent that those are kind of the two the two problems that have always been really useful for me just like sharpening my intuition about models can do and can't do right now instruction following and question answering. Yeah yeah I see. So maybe to go on to the these policy sketches I guess like one thing to maybe start on is that this now brings in reinforcement learning or policy learning. It's important to be very clear that we're not talking about natural language here we're talking about sort of synthetic highly structured highly