Temperature is the one dial that decides how wild the electron wanders. Cold, and the odds spike to a single word — the model walks a straight line, safe and dull. Hot, and the blizzard flattens until every word is nearly equal — chaos, surprise, nonsense. Slide the temperature, watch the odds reshape and the storm thicken, then tap the storm and see which word the electron finds.
Real softmax with temperature: the next-word probabilities are p_i = softmax(logit_i / T), recomputed live as you slide T. Cooling (T→0) sharpens toward the single top word (greedy, entropy→0); heating (T→∞) flattens toward the uniform blizzard (max entropy = log₂(10) ≈ 3.32 bits). The ENTROPY readout — how many bits of surprise the next step carries — is exact, and the toddler's landing TALLY converges to these same odds (TOP's fairness check: tap enough and frequency → probability). The exact sampling mathematics of every language model.
Ten hand-set logits for one toy prompt stand in for a real model's vocabulary of ~50,000 (no model runs here); the words and their scores are illustrative. The softmax-temperature relationship, the entropy, and the sample-frequency convergence are all exact.