Imagine doing this with one AI: It reformulates each question, then it gets the new prior, then it talks to the human. Ignore that it has to do N work in the first step. That might make this easier to see: Why do you think bringing the questions into a form that allows for easy memorization by humans has anything to do with understanding? It could just do the neural-net equivalent of zip compression of a hashmap from reformulated questions to probabilities.
Imagine doing this with one AI: It reformulates each question, then it gets the new prior, then it talks to the human. Ignore that it has to do N work in the first step. That might make this easier to see: Why do you think bringing the questions into a form that allows for easy memorization by humans has anything to do with understanding? It could just do the neural-net equivalent of zip compression of a hashmap from reformulated questions to probabilities.
But that hashmap has to run on a human mind, and understanding helps us run things like that.