After finishing exercise 3 I don’t really understand the core logic. Why do we train a model only to build another right after that?
They explain that in the text of the notebook. You’re training that LSTM cell using the first model and then it’s used as a component of the actual music generation (synthesis) model.