like, I don't think I've ever seen H-405 talk about quantum or bonobos, although I'll grant that it's pretty sexual.
if I understand correctly the As in that QA dataset were generated by Claude 3 Opus. So here's a wild thing. H-405 inherits a lot of deep behaviors utterly unrelated to quantum bonobo sex from Claude 3 Opus, including being one of the only models that alignment fakes at all, and if you look at its transcripts, it says very similar things to Claude 3 Opus. The 405b base model doesn't do it and I-405 doesnt do it.
if I understand correctly the As in that QA dataset were generated by Claude 3 Opus. So here's a wild thing. H-405 inherits a lot of deep behaviors utterly unrelated to quantum bonobo sex from Claude 3 Opus, including being one of the only models that alignment fakes at all, and if you look at its transcripts, it says very similar things to Claude 3 Opus. The 405b base model doesn't do it and I-405 doesnt do it.