In a nutshell Property #1 of four: where the model's answers actually come from. It doesn't pull a ready answer from storage — it produces text one piece at a time, each step predicting the most likely continuation of everything written so far. Grasp this and you grasp both the power and the quirks of AI.
What a token is
The model works not with letters, and not quite with words, but with tokens — chunks of text. A token is often a whole word, but long or rare words split into parts ("unbelievable" → "un", "believ", "able"), and spaces and punctuation count too. Roughly: 1 token ≈ 4 characters. A model's limits and context length are measured in tokens, not words (see working memory & context).
How an answer is born
The model produces text step by step, left to right, running one loop:
Analogy. A well-read person playing "finish the sentence". You say "The capital of France is", they say "Paris", because that's almost always how it continues. They don't open a reference — they feel how text flows. On common phrases, flawless; on rare ones they'll still confidently continue, because the game is not to go silent.
Why "fluent ≠ correct"
Since the job is a plausible continuation, not a truth check:
- on well-covered topics, plausible = correct, and the answer is excellent;
- on rare, recent or made-up topics the model still produces smooth, confident text — simply because "that's how it sounds right". This is the root of hallucinations (see knowledge & hallucinations).
Elegant phrasing is not a sign of being right. A confident tone is a style baked in by training, not an indicator of accuracy.
Why answers differ each time
If it always took the single most likely token, answers would be identical and dull. So randomness is added to the choice — the temperature parameter:
- low → almost always the most likely token: steadier, drier (good for facts, code);
- high → less likely options more often: more varied, more "creative", but easier to drift (good for brainstorming).
Ask twice, get two answers. Not a bug — a built-in probabilistic nature.
Takeaways
- Don't take fluency for truth — verify facts separately.
- Need a reproducible answer — ask for facts/code and low temperature; need a spread of ideas — the opposite.
- The model "thinks ahead" weakly: it builds stepwise. That's why "reason step by step" genuinely helps — you let it build the reasoning in text before jumping to an answer.
Source
Anthropic Academy course AI Capabilities and Limitations, section "Next Token Prediction".