But the Tweet said GPT 4 is the first model to explain the joke correctly. So in some area there must have been an improvement, even if it's just access to existing information. Or it's just a fluke.