The mechanism, step by step.
THE PROCESS
- Your input is converted to tokens
- The model processes them together
- It produces a probability distribution over possible next tokens
- One is selected
- That token is added to the input, and the process repeats
Each word is generated with the whole preceding text in view.
WHY THIS EXPLAINS SEVERAL BEHAVIOURS
Why output streams word by word Why the model cannot revise what it already wrote in a single response Why early words constrain later ones Why the same prompt produces different outputs
THE COMMITMENT EFFECT
Once the model has begun an answer in a particular direction, it continues consistently with it.
An early error propagates.
That is why asking it to think step by step before answering frequently improves results: the reasoning steps become part of the context that shapes the conclusion.
WHAT THIS SUGGESTS PRACTICALLY
Ask for reasoning before the answer, not after If an answer starts wrong, regenerate rather than correcting within the same response Structure matters, because everything before influences everything after
