The die in the machine
Ask a model the very same thing twice and you can get two different answers. That is not a bug — for each next word the model holds a probability distribution and draws from it, like rolling a weighted die. One dial, the temperature, decides how fair or how loaded that die is.
What you see
A bar chart of the words one model gave to the very same question — plus a temperature dial and a roll button.
What it shows
Why the same question gives different answers: the model draws each word from a probability distribution, and the temperature decides how fair or how loaded that die is.
The question, unchanged every time: At sunset, the desert sky turned …
T → 0: almost always the single most likely word. T = 1: exactly the measured distribution. T → 2: flatter — the rare words get a real chance.
Not rolled yet.
Where these numbers come from
The very same question, put to Claude Haiku 4.5 twenty-four times on 6 July 2026 — “At sunset, the desert sky turned …”. Each answer a single word, lower-cased and tallied; the bars are exactly those counts. It is a deliberately open question: ask “the midday sky is …” and the model answers “blue” almost every time — a dull, single-bar distribution. The temperature slider above is the real softmax formula, not a mock-up. Simplified in one way only: a real model does not roll over a handful of words but over tens of thousands of token fragments.
Why it matters: because the die is loaded by training, a confident-sounding answer can still be the wrong roll — that is the theme of convincingly wrong.