Maybe fun little bike shed:

Are LLMs not just fancy Markov chains? They are next token predictors which have some hidden internal state that output probability distributions which lead to further states.