I think he's talking about how during conversations, there are different cognitive levels to a conversation. You are basically having a conversation with yourself about what to say and remembering things to talk about, while at the same time considering the context of the situation, such as the environment or activity.

So he's saying for a model like this, would it be possible to tune the model so that it is able to give better answers in a conversation.