They're not manually generated or anything, it's just a setting. Too few and the model doesn't have enough flexibility to capture complex patterns. Too many and the model can just memorize the data you train it on rather than capturing the patterns driving it.