Why it matters

Constitutional AI scales alignment beyond human labels. Anthropic's approach.

Advertisement

The architecture

Principle set (constitution).

AI critiques + revises.

RLAIF replaces RLHF.

Constitutional AIConstitutionprinciple setAI critique + reviseSL stageRLAIFAI feedback RLBai et al. 2022 Anthropic; RLAIF scales without labeler bottleneck
Constitutional AI.
Advertisement

How it works end to end

Bai 2022 Anthropic.

RLAIF scales.

Principle set.