Temperature

There's a single dial in every AI playground that decides whether your AI behaves like a buttoned-up accountant or a jazz musician. Most people never touch it, because nobody told them what it does.

Remember tokens, the little bricks the AI builds with? The AI works by predicting the next brick. Out of all the words that could come next, it holds a ranking of how likely each one is. Temperature is the dial on how adventurously it picks from that ranking.

The chef's recipe knob. Imagine a chef with a slider that runs from "follow the recipe exactly" to "improvise."

• Turn it down (toward zero) and the chef follows the recipe to the gram, always grabbing the single most likely next word. Predictable, consistent, a little boring. Ask the same question twice and you get nearly the same answer twice.
• Turn it up and the chef starts improvising, sometimes reaching past the obvious word for a less likely one. More surprising, more creative, more variety, and more likely to wander off and say something strange.

So temperature is a creativity-versus-consistency slider. It's not making the AI smarter or dumber. It's changing how boldly it picks from what it already knows.

When to turn it which way, the actually useful part:

• Low for anything where you want the same right answer every time: pulling data, classifying, extracting a number from a document, writing code, or anything a machine reads next as strict JSON. You do not want your invoice-reader feeling inspired.
• High for anything where you want variety and surprise: brainstorming names, drafting ten different hooks, riffing on angles, creative copy. Cranking the dial is how you stop getting the same safe answer on every reroll.

The trap most people fall in: they blame "the AI" for being repetitive when the fix was to turn it up, or blame it for going off the rails and inventing things when the fix was to turn it down. It was the dial the whole time. And here's the tie back to hallucination. A high temperature makes hallucinations more likely, because "reach for the less obvious word" is exactly the moment it drifts from the truth.

This is the second thing you'll keep running into in an API playground. A rate limit is something you obey. Temperature is something you tune. Most chat apps hide it and quietly pick a sensible middle for you. The moment you're building something yourself, it's the first knob worth understanding.

Temperature is the improvise dial. Down for the same right answer every time. Up for ten different answers. Match the dial to the job.