🛡️Constitutional AI
Also known as CAI, AI Constitution, RLAIF
Training a model with a written list of rules. It checks and fixes its own answers.
Think of it like
It is like a kid who checks their own work with a list of rules. Then they fix it.
Example
Anthropic made this method public in 2022 and used it to train Claude. One kind of rule asks for the answer that is the least harmful.
Why it matters
The rules are written down where anyone can read them. That makes it clearer why a model acts the way it does.
Definition
Common mistake
Key takeaways
- 1A written list. Of principles.
- 2Self-critique. Check and rewrite.
- 3AI feedback. In place of many raters.