Did the definition of alignment change? Historically it meant “AI’s goals are good for humans”. This paper seems to be measuring.. how well AI follows directions?