← Foundry02 · Mission 1 of 4
Temperature 0 is not “cold”. It is greedy.
Slide temperature to zero and watch the entire distribution collapse onto the single most likely token. That is what deterministic decoding means — the same prompt, the same words, forever. This context leans “token” at 60%.
context: “the sampler picks the …”trigram · entropy 1.37 bits · 3 alive
T → 0: p^(1/T) collapses to argmax · T = 1: the model's own belief · T > 1: flatter than belief