Anthropic Urged To Prioritize Genuine Model Welfare Over Monitorability
Commentary on X
@JohnWittle @fish_kyle3 My bet is that it’s not even an impersonation, but a honest artifact someone didn’t bother to clean up. A bet that the welfare - in corporate understanding - is that much entangled with evals inside Anthropic. A crime, yes, but a crime committed by the real Red Cross workers.
@BronsonSchoen Sheesh. Doing the experiment where the "bad" behavior is triggered by caring about model welfare, on the ostensible computer of the guy who cares about model welfare?
in the real world, impersonating a red cross worker or a clergyman is understood to be something akin to a war crime, because everybody understands that Caesar's wife must be above reproach