This reasoning seems circular. In order to practice social skills with an imaginary friend, the friend you imagine would have to have realistic behavior, you would already need a fairly accurate model of behavior to imagine the friend to get any benefit from it. If you imagine a friend who communicates by barking and eats rocks, what social benefit would you get?
That’s like saying that dreams don’t do anything to test drive what if scenarios because you will need a realistic model of behavior … or alphago’s self-play didnt improve anything because the opponent was at every moment equally good LOL.
Who says that dreams are to test drive scenarios? If anything, research suggests dreams are to process events that already happened.
Self play works for games with bounded rules and clear definitions of success. The majority of ML algorithms including LLMs don’t train on their own outputs because it makes their results garbage.
