The Challenges in Grounding the Language Model in Intensive Plans

There's limits to the kinds of grounding that we can do. We're not able to control the dialogue model for things like the style of communication or long term strategy. It usually gets it right, but sometimes it will say too much or like we're going to attack somebody. If you have a bunch of candidate messages, one sends a very generic message being like, "Oh, it's been great working with you so far"