Temperature

Temperature is the single knob that decides how adventurous a language model is allowed to be when it picks each next word. Before a model like a GPT commits to a token, it doesn’t have one answer in hand, it has a whole spread of candidates with odds attached: maybe 60% “the”, 20% “a”, 5% “an”, and a long tail of unlikely stragglers. Temperature reshapes that spread right before the dice are thrown. Turn it down and the odds get sharper, the favorite gets even more favored, and the model plays it safe. Turn it up and the odds flatten out, the longshots get a real seat at the table, and the model gets loose, surprising, eventually unhinged.

Mechanically it’s one division. The model’s raw scores (the “logits”) get divided by the temperature value before they’re turned into probabilities. Divide by a small number and you exaggerate the gaps between candidates; divide by a big one and you squash them toward equal. The usual dial runs from 0 to about 2. At temperature 0 there’s no gamble left at all: the model takes the single most likely token every time, which makes it deterministic and repeatable (same prompt in, same answer out). Around 0.7 to 1.0 is the everyday range, enough spice to sound human without falling apart. Push past 1.5 and coherence starts to dissolve into word salad.

One thing temperature is not: it’s not the random seed. The seed fixes which roll of the dice you get; temperature decides how loaded the dice are in the first place. You can hold one steady and move the other. A clean example lives in the deep-dive On the Machine We Switched Off , where a 1930-trained model is interviewed twice with the temperature pinned at 0.8 and the seed pinned at 1930, so the only thing changing between runs is the language of the questions. Freezing temperature is what makes that an experiment instead of an anecdote: it takes randomness off the table as a suspect.