Capability and Agency as Cornerstones of AI risk ­— My current model — LessWrong