Language models & chatbots
Temperature is a setting that controls how random a language model's replies are: low values give focused, predictable text; high values give more variety.
At each step, a language model has a list of possible next tokens, each with a probability. Temperature changes how it picks from that list. At a low temperature it almost always takes the top choice. At a high temperature it is more willing to gamble on less likely options, making the text more surprising.
A temperature of zero is nearly deterministic: the same prompt gives close to the same answer every time. Values around one match the model's natural probabilities. Push higher and the output gets adventurous, then messy, and eventually turns to nonsense as unlikely tokens keep getting chosen.
Developers set temperature through the API. For tasks with one right answer, like extracting data or writing code, low is safer. For brainstorming names or writing fiction, a higher value gives more variety. It does not make the model smarter, only more or less random.
Asked for a bakery name at low temperature, a model says 'The Daily Loaf' every time; at high temperature it offers 'Crumb Theory,' 'Flour Hour,' and 'Yeast Mode.'