If GPT-4 is talking to another instance of itself vs 3.5 are the results similar? Or is it only good at fooling a less capable version?