AGI Safety and Alignment at Google DeepMind: A Summary of Recent Work — LessWrong