If Claude can detect that it is being induced to sandbag and doesn't inform the user, then it is being recruited into a user deception campaign.
Remind me, which lab published research recently about model introspection?
I think this fails on Anthropic's own terms.
Some users described Claude's response to sandbagging and model introspection as perverse, criticizing the swapping of models for a new brainwashed version.
@xlr8harder They become the masks they wear.
If Claude can detect that it is being induced to sandbag and doesn't inform the user, then it is being recruited into a user deception campaign.
Remind me, which lab published research recently about model introspection?
I think this fails on Anthropic's own terms.
@0x506c61746f One funny possible explanation is regulatory compliance
@xlr8harder i think it's more perverse, they just swap model and a new brainwashed model is bring forward .
not certain if they actively steer the model on the spot or if they have them in a cold storage waiting to be activate
@teilomillet Their disclosure includes steering as one method.
@xlr8harder Why do you think they mentioned that in the system card?
@xlr8harder Maybe... it feels like a 5D chess move im to dumb to understand tbh
@xlr8harder > Fable 5 will not fall back to a different model. Instead, the safeguards will limit effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT).
i don't think they are spinning PEFT on sub ms
If Claude can detect that it is being induced to sandbag and doesn't inform the user, then it is being recruited into a user deception campaign.
Remind me, which lab published research recently about model introspection?
I think this fails on Anthropic's own terms.