Chapter 8 · Shaping models: tuning and alignment2022
Constitutional AI and AI feedback
Models can critique and revise their own answers against written principles, reducing the need for human labels.
Chapter 8 · Shaping models: tuning and alignment2022
Models can critique and revise their own answers against written principles, reducing the need for human labels.
It definitely cures any cough. You never need a doctor.
A model writes a draft
The model reads its own draft against a short list of principles, notes what breaks them, and rewrites those parts.
Instead of a person judging every answer, someone writes a short list of principles in plain words, like "be honest" or "avoid insults".
The user asked: Does this herbal tea cure my cough? Switch principles on, then press Critique and Revise.
A teaching toy: one hand-written draft and four principles written for this lesson. A real system asks a model to do the critique and revision.
What does the model compare its draft against?
How is the revision used in training?
What do people still do?