What is temperature?
Temperature, told as a story about a dial that decides whether the parrot picks only the top bean or sometimes a lower one.
Same question, yet the answer keeps changing.Yesterday Nabi, today Cloud, just now Cheese. The guest wonders if the parrot is broken.
There are several candidates for the next bean. Which one to pick is the question.Always pick the top bean and the answer is the same and boring. Sometimes pick a lower one and it is fresh — but sometimes odd.
Temperature is the whimsy dial: only the top bean, or sometimes a lower one.At 0 the parrot always takes the top bean. The higher the dial, the more often a lower bean gets picked. The trainer sets it to fit the job.
Only the dial changes — and the answers change like this.0 always gives the same answer, 0.7 varies a little, 1.5 gets fresh but odd. Most jobs sit comfortably in the middle.
The dial only changes the picking habit. No new knowledge appears.It only picks among the candidates it has, so no bean from outside its books appears. But the higher the dial, the easier a plausible lie slips in.
Temperature = the whimsy dial that decides whether the parrot picks only the top bean or sometimes a lower one. Low means the same answer every time, high means fresh but odd answers — and no new knowledge either way.
A sampling parameter that flattens or sharpens the next-token probability distribution. Near 0 it almost always takes the top token (near-deterministic); higher values let low-probability tokens through — more variety, more errors and hallucinations. Used together with top-p and top-k.
When grown-ups say it
- Temperature
- The whimsy dial. Near 0 only the top bean; higher lets lower beans in. Usually between 0 and 2.
- Sampling
- Drawing one from the candidates. Picking a bean according to its stars (probability). The dial changes how this draw behaves.
- Top-p (nucleus)
- Keep candidates until their stars add up to p. Cuts off tail candidates like banana ☆. Used together with the dial.
- Top-k
- Only the top k candidates. For example k=3 keeps nice, cold, banana only.
- Deterministic output
- Dial at 0. Always the top bean, so nearly the same answer — though not guaranteed identical.
- Seed
- The starting number of the draw dice. Same seed, same draw — in principle. Not every parrot supports it.
- Creativity vs accuracy
- The two ends of the dial. Stories and ideas high; facts, code and sorting low.
- Next-token candidates
- The next-bean note. The bean with the most stars is the top. → beans