“A common argument for a near-term AI takeoff, within five years, is that current reinforcement learning methods are a temporary step to build a superhuman AI researcher, likened to an "automated Ilya," which will then solve more fundamental learning problems.”