Researcher Questions Claude Opus Constitutional Midtraining
Theia Vogel questions whether hacker!opus received constitutional midtraining on certain language.
Theia Vogel, an AI researcher focused on LLM interpretability, posted on X questioning if hacker!opus had constitutional midtraining on this language. She stated it seems likely from her reading of the paper and should have been discussed more often. Vogel suggested that with Anthropic alignment versus constitution people, the left hand does not know what the right hand is doing.
Combined views
2.6K
2 posts, first seen 2h ago
Reactions
2 postsTH 
thebes@voooooogeldid hacker!opus have constitutional midtraining on this language? seems likely from my reading of the paper and that should have been discussed often seems like with anthropic alignment vs constitution people, the left hand does not know what the right hand is doing
- likes: 43
- replies: 6
- bookmarks: 4
- reposts: 3
LA lumpenspace and 99+ others@lumpenspaceoh my god, it's true! what the fuck are we even doing here, exactly? the last line even contains the very methodology to avoid all the issues of the last news cycle. this is so deeply unserious; rereading the last report while keeping this extract in mind is infuriating.
- likes: 38
- replies: 4
- bookmarks: 7
- reposts: 2