The rapid rise of computational power allows ever more capable AI agents to be trained in simulation. A simulator, of course, does not fully reflect reality nor human preferences. How can AI agents learn useful, human-aligned behaviors in simulation that transfer to new settings and people?

I have considered this question from the lens of generalization, human-AI collaboration, and open-ended learning at Facebook AI Research, Google DeepMind, and Meta Superintelligence Labs. Recently, I have left the frontier labs to approach these problems from a different perspective.