Temperature in plain words: why the same prompt gives different answers

A 5-beat faceless video outline for “Temperature in plain words: why the same prompt gives different answers” in the AI tools explained niche, derived from caption timing of 2 reference videos.

Angles

• Cold-open on the experiment: same question 20 times — one setting gives one answer 20 times, another gives 18 different answers
• Explain the pipeline in two steps: the model scores next tokens, then a random draw picks one in proportion — that draw is sampling
• One sentence for temperature: it sharpens or flattens the chances, nothing more (softmax analogy from physics as an optional aside)
• Close on "zero means most repeatable, not most right" and a which-setting-for-which-job card

Beat-by-beat structure

1. Hook (0:00–0:30): two screenshots of the same prompt with different answers.
2. Next-token scores (0:30–2:00): a bar chart of candidate words; why the model never "knows" one answer.
3. Temperature dial (2:00–4:00): low vs high on the same chart; extreme ends (always top word vs near-random).
4. Tail cutters (4:00–5:30): top-k / top-p in one line each; defaults differ by app.
5. Use card (5:30–7:00): facts/code/extraction low; brainstorming higher; temp 0 still may vary in the cloud. Original narration only.

Reference videos

• LLM Temperature, Top-P, Top-K Explained: I Ran It 650 Times hID83fOkh40
• Softmax - What is the Temperature of an AI?? YjVuJjmgclU

Method

yt-dlp VTT captions (manual/uploader only) → timed beat heuristics → editorial topic synthesis (no ASR). Structure only; write original narration.