What is Temperature in AI?

Temperature is a setting that controls how predictable an AI model's word choices are. Low values make it pick the most likely next word almost every time; higher values let it gamble on less likely ones. It tunes variety, not intelligence.

Before writing each token, the model holds a ranked list of candidates with probabilities attached, and temperature reshapes that list. Near 0, the top choice wins almost every time, so the same prompt gives nearly the same answer twice. Around 1, the model samples in proportion to the original probabilities, and above that the gaps shrink until longshot words get real chances. Typical products run between 0.7 and 1.0 for chat, while code tools and extraction pipelines sit near 0, because a creative answer to a formatting task is just a wrong answer. Higher settings raise variety and with it the odds of drifting into nonsense, since every unlikely word makes the next one likelier to be unlikely too. The name comes from physics, where heat is what shakes particles out of their most stable positions.

An office lunch order. Play it safe and everyone just gets the restaurant's bestseller, identical and reliable every single day. Turn the dial up and the intern picks by throwing darts at the menu: some days a discovery, some days deep-fried regret.