Lots of questions after reading this #anthropic post outlining the Constitutional AI that guides #claude https://www.anthropic.com/index/claudes-constitution The potential to scale better than RLHF systems seems clear, interesting angle on minimizing human interaction w/ disturbing outputs. But letting AI train AI in the reinforcement learning stages sounds like a potentially insane echo chamber if compromised? Eager to read through the full paper #constitutionalai #ai #rlhf
