AGI safety from first principles: Goals and Agency — AI Alignment Forum