Temperature
A setting for how adventurous or predictable the AI's word choices are.
In everyday terms
Low temperature: it nearly always picks the most likely next word (steady, repetitive). High temperature: it takes more chances (creative, sometimes nonsense).
For professionals
Divides logits before softmax. T→0 approaches greedy decoding; T>1 flattens the distribution. Often combined with top-p / top-k sampling.
Think of it like…
A cautious storyteller versus a wild one, telling the same story.
You've already seen it
Hitting "regenerate" and getting a different answer each time.
Myth vs reality
Myth: Higher temperature makes the AI smarter.
Reality: It only makes it more random. Good for brainstorming, bad for facts.
Quick check
For extracting exact figures from an invoice, you'd want temperature…
- High
- Low
- It doesn't matter
- Negative
Show answer
Low: Low temperature = consistent, predictable output.