If your AIs are trying to escape, it's crucial to think about whether you can catch them before they succeed, because catching them red-handed gives you lots of options you didn't have before.
Very interesting read! There are many questions that I remain curious about. When should we expect the first case of scheming in the wild? Do we have the capabilities to catch such cases, or will it slip by our defenses? How closely can a model organism actually replicate a scheming model, and what gaps do we need to fill in order for our safeguards to be useful against actual schemers? Have we witnessed cases of escape-breaking eliciting higher capabilities in password-locked or scheming models?
Very interesting read! There are many questions that I remain curious about. When should we expect the first case of scheming in the wild? Do we have the capabilities to catch such cases, or will it slip by our defenses? How closely can a model organism actually replicate a scheming model, and what gaps do we need to fill in order for our safeguards to be useful against actual schemers? Have we witnessed cases of escape-breaking eliciting higher capabilities in password-locked or scheming models?