AI trains itself to be harmless using only rules

Original: Constitutional Ai Harmlessness From Ai Feedback

Source: anthropic.com ↗