Basically, each new character (letter) is generated by picking a character based on what is most likely given all the previous text in the setlist and what the model has learned from seeing the history of all Phish setlists.
Outside of seeing every Phish setlist ever, the model knows nothing about the English language or grammar. So it can be ridiculous at times, especially when it decides to generate "show notes".
The slider controls the "temperature" of the sampling. So the higher it is, the stranger and more unlikely the choices it makes in terms of what character comes next.