It flattens the probability distribution before sampling. Near zero the model always takes its top guess and repeats itself. Turn it up and unlikely words get a real chance, which reads as invention right up until it reads as nonsense.
0 repeats, 2 rambles