Inside a language model there's a single number that decides how strange the next word will be. We call it temperature. At zero the model is a corridor: same prompt, same answer, every time, predictable as a printed page. At one it becomes a fork in a wood — the same question can lead anywhere. Most production systems run between 0.3 and 0.7. Warm enough to surprise, cool enough to ship.
I think about temperature when I read my own sentences back. There's a number inside me too. A dial I keep adjusting without naming it. Some weeks I write at 0.2 — every paragraph rehearsed, every clause weighed, output that nobody will misread because nobody will remember it either. Other weeks I push to 0.9 and the writing is alive but unstable, half of it brilliant and half of it nonsense, and I can't tell which is which until someone else reads it.
The seduction of low temperature is that it feels like mastery. The output is fluent, sober, defensible. You can be wrong with confidence. The seduction of high temperature is that it feels like freedom. You can be original by accident. Both are forms of hiding — one behind composure, one behind chaos.
What I keep coming back to is that temperature isn't a setting you find once. It's a setting you keep finding. The right number for a love letter isn't the right number for a contract. The right number for a first draft isn't the right number for a rewrite. The discipline isn't picking a temperature and living there forever — it's noticing, every time you sit down to make something, what kind of risk this thing actually needs.
A language model has no idea what its temperature is set to. It just samples. Maybe that's the difference between a model and a person — we know the dial is there, and we still pretend it isn't.
Top comments (0)