The Alignment Problem

In the first section, Christian interweaves discussions of the history of artificial intelligence research, particularly the machine learning approach of artificial neural networks such as the Perceptron and AlexNet, with examples of how AI systems can have unintended behavior. He tells the story of Julia Angwin, an intrepid journalist whose ProPublica investigation of the COMPAS algorithm, a tool for predicting recidivism among criminal defendants, led to widespread criticism of its accuracy and bias towards certain demographics. One of AI's main alignment challenges is its black box nature. The lack of transparency makes it difficult to know where the system is going right and where it is going wrong.

In the second section, Christian similarly interweaves the history of the psychological study of reward, such as behaviorism and dopamine, with the computer science of reinforcement learning, in which AI systems need to develop policy ("what to do") in the face of a value function ("what rewards or punishment to expect"). He calls the DeepMind AlphaGo and AlphaZero systems "perhaps the single most impressive achievement in automated curriculum design." He also highlights the importance of curiosity, in which reinforcement learners are intrinsically motivated to explore their environment, rather than exclusively seeking the external reward.

The third section covers training AI through the imitation of human or machine behavior, as well as philosophical debates such as between possibilism and actualism that imply different ideal behavior for AI systems. Of particular importance is inverse reinforcement learning, a broad approach for machines to learn the objective function of a human or another agent. Christian discusses the normative challenges associated with effective altruism and existential risk, including the work of philosophers Toby Ord and William MacAskill who are trying to devise human and machine strategies for navigating the alignment problem as effectively as possible.