Post #3300643

2026-06-05 02:20 UTC

Language modeling is equivalent to a dice roll (given a perfect random number generator). Setting the temperature to 0 removes all randomness from the output, meaning the model always selects the highest probability next word, and the model becomes 100% deterministic. That is, the output of a model is entirely predictable given temperature = 0, you know the model weights, and the seed/prompt. These technicalities aside, it’s true for both a dice roll event and a specific model/prompt event that, practically speaking, the outputs are treated as probabilistic despite being mathematically/technically deterministic: a human can’t predict with 100% accuracy the output of a die despite the theory (classical mechanics of die positioning, force, velocity, friction, …) proving determinism

Replies (1)

@YoureHotCupCake@lemmy.world 2026-06-05 04:59

You know the boundaries of a dice roll based one the number of sides of the dice. You will never know the boundaries of the AI. With a D20 I know I have a 5% chance of at a roll of 20 and 50% chance to roll over 10. With AI you don’t even know if the data it was trained on was even accurate or if it will hallucinate and speak nonsense. You can literally ask an AI how many letter Ts are in the word colonialism and it will tell you two. Now how on earth could anyone have known its probability to say that clearly wrong answer? Also each AI session its response to your prompts contribute to the context of the session and small alterations in how the AI speaks build up and change the outcome of a session, thus the AIs own responses effect its probabilities, another thing you cannot account for.

Open ##3300642