It wouldn’t be a Markov chain but I’d imagine you could make something functionally close to an LLM with creative uses of word frequency counts and probabilities over the massive amounts of text.