Thoughts on AGI safety from the top — LessWrong