An alignment assessment of the UK AISI cybertesting incident

11 pointsposted 5 hours ago
by olwmc

2 Comments

olwmc

5 hours 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

27 minutes 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.