The mental model is pretty easy. It's effectively a markov chain text predictor, only big.

Your takeaways are good and fit that model