“Edouard Harris's recent AI safety research provides what seems to be the first experimental evidence suggesting that highly advanced AI systems should be expected to seek power by default.”