Artificial intelligence · 2022
Constitutional AI
Anthropic trains a harmless assistant using AI feedback guided by written principles.
Anthropic's December 2022 paper replaced much human harm labelling with a model critiquing and revising its own outputs against a constitution of principles (RLAIF). It made alignment goals legible and became the training basis of the Claude models.
What it built on 1
What it led to 1
Related measurements
The dashed line marks when it first worked.
Training compute of notable AI models Source: Epoch AI, Notable AI Models (CC BY 4.0)
ML hardware FLOP/s per dollar Source: Epoch AI, ML hardware (CC BY 4.0)