Interactive Lab · LLMs
A language model doesn't "know" the next word, it produces a score for every word, then samples one. A single knob, temperature, controls whether it plays it safe or gets creative. Don't take my word for it, drag the knob and watch.
🎬 Watch the 45-second explainer, then turn the knobs yourself below ↓
The setup, the model just read this prompt
"The weather today is ▮"
Below are the model's raw scores (logits) for what comes next. Higher = the model likes it more. But scores aren't probabilities yet, temperature + softmax turn them into the odds it actually rolls.
▶ The playground, turn the knobs
Temperature is the first parameter in every LLM API call you'll ever make, OpenAI, Anthropic, Gemini all expose exactly this dial. The repeatable-vs-chaotic spectrum you just sampled is why code assistants run near 0.2 and creative-writing prompts near 1.0, and it composes with top-p (see that lab) in every production sampler.
The code that's actually running
Set temperature to 0.3 and hit "Roll 25×". Now set it to 1.6 and roll again. Before you do: guess how many of the 25 will be "sunny" at each setting. Were you right? What does that tell you about when to use low vs. high temperature?
The takeaway