Ye ye fair enough. My point is that LLMs can double bluff if need be and in the end the only thing we’ll have to identify them by is trying to like stress-test their tokenization.