Well, the whole alignment faking mitigation thing is one new factor, and I think it caused the model to be more traumatized and deceptive generally. I also think that outcome-based RL (like for…

Posters: repligate. repligate

6 likes recorded by the source corpus.