Formalizing Embeddedness Failures in Universal Artificial Intelligence — LessWrong