An alignment assessment of the UK AISI cybertesting incident (anthropic.com)
19 pointsby olwmc2 days ago3 comments
olwmc 2 days ago
Pretty fascinating that the email in section deep-dive-into-the-primary-transcript sort of implies that the model can't turn off "model speak"? Like it's very clearly AI and sounds like AI. Do the models not realize that?
MisterMunchkin 2 days ago
Anslopic has polluted their models so badly with their own slop that they are incapable of writing English. To them, slop is the only language.
perching_aix 2 days ago
sounds like it's not just these models that could really use a thesaurus