How random is random?

Cutting off the tail is one way to control what gets drawn. There is another, and it works on the whole spread rather than its bottom end.

Softmax pulls the gaps between scores apart, and how hard it pulls is set by a single number called temperature. It is not a creativity dial bolted onto the model. It is one setting on the arithmetic you just watched.

Take the same four scores and turn it.

Turn the temperature down and the gaps stretch further apart. At a low setting, those four tokens come out as jar 91 percent, lid 8, box under 1, ievable effectively zero. There is barely a decision left. Ask a hundred times and you get jar almost every time.

Turn the temperature up and the gaps compress. At a high setting the same four become jar 38 percent, lid 30, box 22, and ievable 11. The order has not changed, but the model is now a coin toss between several options, and a fragment that makes no sense is winning one draw in nine.

Nothing about the model changed. Same weights, same raw scores. The only thing that moved is how sharply those scores were turned into chances. Temperature changes variety, not truth, and a low setting gives you the same answer every time rather than the right one.

Which explains something people notice constantly. Ask a model the same question twice and you can get two different answers. It did not change its mind. A token was drawn from a spread, and the draw landed somewhere else the second time.

It also explains the usual advice. Keep it low for a factual lookup, where you want the model's best guess and want it consistently. Raise it for writing, where the second and third choices are often the more interesting ones. Raise it too far and you get the eleven percent.

The next slide hands you the dial.