“Edouard Harris's research found that in two-agent simulations, if the AI's and human's goals are totally uncorrelated, the agents consistently compete over instrumental sub-goals, also known as power.”