One way of getting good random samples is to give model a random starting points. For example: "write a short story about PP doing GG in XX". Here PP, GG and XX

One way of getting good random samples is to give model a random starting points. For example: "write a short story about PP doing GG in XX". Here PP, GG and XX are filled algorithmically from lists of persons, actions and locations. The problem is model's randomly generated output from the same prompt isn't actually that random. Changing the temperature parameter doesn't help much. But in general it's a big secret because the training data and techniques are the only difference between models as architecture is more or less settled.