Temperature Explorer

Run the Temperature Explorer MicroSim Fullscreen

About This MicroSim

"Temperature controls creativity" is the explanation everyone gives, and it is not wrong, but it is not an explanation. This MicroSim shows what temperature actually does: it rescales the probability distribution over candidate next tokens before one is sampled.

Slide temperature to 0 and the tallest bar swallows everything — the same word wins every time. Slide it to 2 and the bars flatten until bureaucratic is nearly as likely as young. Between those extremes is where useful text lives.

Each bar shows two numbers: the raw probability the model assigned, and the temperature-adjusted probability actually used for sampling. Watching the second number move while the first stays put is the whole lesson.

How to Use

Pick a prompt stem from the dropdown. Each one has its own hand-built token distributions.
Move the temperature slider and watch the bars redistribute. Nothing is sampled yet — this is purely the shape of the distribution.
Press "Generate Next Token" to sample one token. The chosen bar is outlined in red and the word is appended to the sentence below.
Press "Generate 10 Tokens" to run the rest of the sequence at the current temperature.
Press "Reset", change the temperature, and generate again. Run the same stem at 0.0, at 0.7, and at 1.8, and compare the three sentences.

About the model

The probability distributions here are hand-authored to be illustrative, not scraped from a live model. The softmax-with-temperature maths is exactly what a real model uses; the numbers it operates on are teaching data.

Iframe Embed Code

You can add this MicroSim to any web page by adding this to your HTML:

<iframe src="https://dmccreary.github.io/prompt-class/sims/temperature-explorer/main.html"
        height="567px"
        width="100%"
        scrolling="no"></iframe>

Lesson Plan

Learning Objective

Students will be able to demonstrate how the temperature parameter changes the probability distribution of token selection, and predict the effect on output creativity.

Bloom's Level: Apply (L3) — demonstrate, practice

Grade Level

High school through adult learners. No mathematics beyond reading a percentage is required.

Duration

15 minutes

Prerequisites

Students should know what a token is and that models generate one token at a time (Chapter 1).

Activities

Predict before you sample (5 min): With temperature at 0, students predict which word will be chosen, then press Generate Next Token to check. Repeat three times and confirm the result never changes.
Find the breaking point (5 min): Students raise the temperature in steps of 0.2, running the full sequence at each setting, until the sentence stops making sense. They record the temperature where that happened. Compare answers across the class — they will differ, which is itself worth discussing.
Match the task (5 min): Give students three tasks — extracting dates from an invoice, brainstorming product names, drafting a customer email — and have them argue for a temperature setting for each.

Discussion Questions

At temperature 0, the raw and adjusted numbers look very different, but the winner is always the same. Why bother rescaling at all?
Why is temperature 0 usually the right choice for data extraction and usually the wrong choice for creative writing?
The word purple is in several of these distributions with a low raw probability. At what temperature does it start winning, and what does that tell you about high-temperature output?
If a model gives you a different answer each time you ask, is that a bug?

Assessment

Can the student predict the output at temperature 0 before running it?
Can the student explain, using the two numbers on a bar, what temperature changed?
Can the student recommend a temperature for an unfamiliar task and justify it?