The point being made is equivalent to pointing out that by strict mathematical definition there aren't physically reliazable distinguishably-non-Markovian processes, because computers are finite objects. Yes, you can talk about Markov chains as referring to everything up to and including actual human brains, but the kind of Markov chains people actually refer to as Markovian in practice are much more specific than that, typically those systems with structurally simple, indexical nodes, but sometimes extending to more complex ideas like RNNs where the states are at least of finite bandwidth. An LLM not only has continuous-valued nodes, it has nodes whose description length grows proportionally to and theoretically-unboundedly with respect to the size of its history.