This work and CAI both observe the same basic phenomenon: if language models are sufficiently large and we add enough RLHF to make them helpful, we can more effectively get them to abide by high-level ethical principles expressed in natural language.

→ View original post on X — @anthropicai,