By "understanding" he simply means it builds a model of higher order abstractions over the text in order to model the meaning and context of each token in relat

By "understanding" he simply means it builds a model of higher order abstractions over the text in order to model the meaning and context of each token in relation to other tokens. This is more sophisticated than a Markov process.