Modeling Failure Modes of High-Level Machine Intelligence — LessWrong