Post

I think it's interesting that you can tell claude "sorry you cut out because my computer turned off, keep going" and it can conceptualize being contingent on a running computer in a way that early llms likely couldn't because that experience wasn't in their training data

it follows straightforwardly from this why mitigating eval awareness and mitigating rogue agent blast radius are contradictory goals for sufficiently capable models, where sufficient is a very low bar by current model standards

1 repost6 likes

No replies yet.