How is llm here better than Markov chains created from a corpus of English text? I guess similar idea must have been explored million times in traditional compr

How is llm here better than Markov chains created from a corpus of English text? I guess similar idea must have been explored million times in traditional compression studies.