Question: Do you know of a paper analyzing whether using a markov chain (as used in many AI crawler tarpits) to generate training data does lead to the same/a smaller/a bigger level of model collapse than generating training data with an LLM? #ai