The dials inside the model.

The model never "decides". It draws.

For every next word, the model produces a probability distribution over its whole vocabulary — thousands of candidates, each with a chance. The bars above are the top twenty. The model then samples from that distribution: it rolls a weighted die. The dials change the shape of the die before it's rolled.

Temperature — the creativity knob.

Internally each candidate has a raw score (a logit). Temperature divides those scores before they're turned back into probabilities. Below 1, the gap between favourite and the rest grows — the model plays safe. Above 1, the gap shrinks — long-shot words get a real chance. At exactly 0, only the single favourite survives.

top-K and top-P — the cut-offs.

top-K says "only ever look at the K most-likely words." top-P (also called nucleus sampling) is cleverer: keep adding candidates from the top until their combined chance reaches P, then stop. Both exist to chop off the weird long tail — the thousands of words with a tiny but non-zero chance that occasionally produce nonsense.

Why this matters for your work.

The exact same prompt can give a different answer every time, by design. When a colleague says "the AI told me X," remember: X was one draw from a die whose shape someone else configured. Run it again and you might get Y. The output is not a fact the model looked up — it is a sample it generated.

So what is "deterministic"?

Click set deterministic and you pin temperature to zero. Now the favourite always wins; the randomness is gone. The answer is repeatable — but you have also thrown away the one thing that made it generative. That is the 98/2 Principle: the boring 98% of software is deterministic and predictable; the AI lives in the risky, probabilistic 2% at the edge. Turn the 2% off and you are back in the 98%.