Temperature
Temperature is a setting that controls how random a model's word choices are. Turn it low and the model plays it safe, picking the most likely next word almost every time. Turn it high and it gets loose and creative, which is either what you wanted or a stream of surprises you did not.
Why it matters
Leave temperature high on a task that needs one right answer, and the same question returns a different reply each run, with some drifting off the rails. That wrecks anything you test or bill against: your golden set passes on Monday and fails on Tuesday with no code change. Low temperature buys you repeatability, which is what most support bots, extractors, and classifiers need. High temperature earns its keep when you actually want variety, like brainstorming or draft copy.
How it works
Temperature is usually a number from 0 to 2. It reshapes the probability the model assigns to each possible next word before it picks one. Near 0, the top choice dominates and output barely changes between runs. Higher values flatten those odds so less likely words get a real chance. It pairs with top-p, another dial that trims the candidate pool. Even at temperature 0 the output is not fully guaranteed identical, so it reduces randomness rather than erasing it.
A refund bot runs at temperature 1.0 and answers the same policy question three ways across three customers, one of them subtly wrong. Drop it to 0.2 and the bot returns the same clean, on-policy answer every time. Save the high setting for the marketing tool that writes ten different subject lines on purpose.