How wild the pick is

Think of temperature as a knob that reshapes the list of next-word percents before a pick. A low setting piles almost all the percent on the top word, so that word comes up almost every time. A high setting spreads the percent out, so second-place, third-place, or even 999th-place words get written more often.

Another way to see it: a low setting makes the reply consistent, and less surprising. The top-weighted word comes up more often, or even always. A high setting lets words with smaller percents come up more often — more variety, and more stray words.

Neither setting is good or bad on its own. A very low setting is a good fit for “what is 2+2,” and a poor fit for “tell me a story.” A high setting can write a livelier nursery rhyme. On “what is 3×1?” it can also write 3… or 8… or penguin.

Turn the temperature down: the top token’s share grows. Canberra, as a possible next word for “What is the capital of Australia?”, might go from 62% to 90%. In that case the next word would almost always be Canberra, whether the model is using sampling or greedy decoding.

Turn the temperature up: the top share shrinks. Sydney and other words get more of the 100%. Sampling then writes more variety — but also allows more stray words.

What the number is doing

Chat tools often have a default temperature setting near 0.7. Canberra still leads, and a runner-up can still appear. At 0.1, almost the whole 100% sits on Canberra, so sampling looks like the usual pick. Past 1.0, the slices get closer in size, so a far-down word — kangaroo, or Wolverine — gets more of a chance. Turning the knob does not retrain the model. It only changes how the next word is taken off this list.