Self-Supervised Learning and AGI Safety — LessWrong