For a “large” language model, the diversity of outputs is pretty small.