“Cooperative inverse reinforcement learning is the closest existing work to a formal theory of AI alignment, but it fails to address key difficulties like tasks the principal cannot understand and multi-agent scenarios.”