Embedded Agency via Abstraction — LessWrong