My dumb guy take on it: suppose we build a human-level AGI and it turns out to be limited by compute and memory. Those being limiting factors don’t seem at all far-fetched to me; it seems unlikely that the first real-time AGI will be mostly idling its CPUs. So then wait 18 months and run that same program on a machine that’s this year’s model plus a Moore’s Law doubling. You’ve probably got ASI. Right?