Evaluating Existing Approaches to AGI Alignment — LessWrong