Old computers followed rules: same question, same answer, every time. But in 1948 Claude Shannon showed how to make English without a single rule: open a book at random, find a word, copy the word that follows it, turn to another page. Out came "attack on an English writer that the character of this". Just odds.

Today's language models do Shannon's trick with far more text: for every next word they hold odds, and they draw one. The harness decides the draw. Temperature reshapes the odds, top-p keeps the likeliest words until they cover 90 percent and drops the long tail, and training on the answers people rated tilts the odds. Ask twice and you get two draws. Even at temperature zero, one test asked the same question 1,000 times and got 80 different answers.

The odds on screen for "Name a famous scientist" are an illustration, not a real model's numbers.

What would you ask twice?

Sources: Claude Shannon, "A Mathematical Theory of Communication", Bell System Technical Journal (1948); Holtzman et al., "The Curious Case of Neural Text Degeneration", ICLR (2020); Ouyang et al., "Training language models to follow instructions with human feedback" (2022); Horace He and Thinking Machines Lab, "Defeating Nondeterminism in LLM Inference" (2025).
Sounds: Freesound, CC0 (swordofkings128, B.Mitchley, deraj, Hano_van_Dalen, ramsamba); our own, made in code (earshot, CC0).