The Video Assistant · Lesson 5 — The Strict Referee ← Course

The strict referee — temperature.

Show the same shot twice and you might get two different calls. A single dial — temperature — decides how often that happens.

From Lesson 1, the model never really sees one certain answer. It sees the odds: a spread of possible calls, each with a probability. But it has to commit to just one. How it picks is the question this lesson answers.

It picks by drawing from those odds, and a setting called temperature controls the draw. Turn temperature down and the model almost always takes the single most likely call — so it is steady and repeatable. Turn it up and the less-likely calls get a real chance — so the answers vary and surprise.

Think of a strict referee who calls everything by the book: the same shot earns the same verdict, every single time. A more relaxed referee might see that same incident and rule differently on a different day. Temperature is the dial between those two referees.

Low temperature when you want consistency — a verdict, a number, a fixed format. High temperature when you want range — ideas, phrasings, variety. Same model, same shot; the dial decides how adventurous the answer is.

Set the temperature, then review the same shot a few times. Low and steady, or high and surprising — watch the calls.

Set a temperature, then review the same shot a few times.
temperature
T0.7
balanced
→ the odds for this one shot
goal0%
no goal0%
stands0%
→ calls drawn from those odds
no calls yet — press review

One setting on the call.

Temperature is a single argument. Low for answers you want to be the same every time; high when you want the model to range around.


      

The dial between steady and surprising.

The model turns a shot into odds. Temperature reshapes those odds before a call is drawn: low, and the top one towers over the rest — the strict referee, the same verdict each time; high, and the odds flatten so the long shots start landing — varied, fresher, less predictable.

It changes how the answer is chosen — not how much the model knows, and not how confident it is: the underlying scores are identical either way. Turn it down for consistency and format; turn it up for range and ideas — and remember that more variety is not the same as more correct.

Next: the other dials →
Go deeper — what temperature really does optional

It reshapes the odds

Under the hood, temperature divides the raw scores before they become probabilities — the softmax step from Lesson 1's maths. A low temperature stretches the gaps apart, so the top option towers over the rest; a high temperature squeezes them together, so the options even out and the draw gets a real choice.

Greedy at zero, random at the top

Pushed all the way to zero, temperature stops being a draw at all — the model simply takes the single most likely call every time, which is fully predictable. Pushed very high, the odds flatten toward an almost random pick. Everyday chat sits in between, so replies feel natural rather than robotic.

Variety is not quality

A higher temperature is not “more creative thinking” — it is only a willingness to pick lower-probability words. That can find a fresher phrasing, but it can just as easily wander into a mistake. If an answer must be correct and consistent, keep the temperature low.

Not the only dial

Temperature is not the only dial on the draw — the next page has two more.