AGI safety from first principles: Control — LessWrong