How is AGI takeoff scenario compatible with the world energy consumption? While the latter in 2023 was of the order of 30 thousands of TWh per year, in 2023 and 2024 many researchers established the growing energy footprint of artificial intelligence. Even the GPT-3 interactions apparently require an electricity consumption of approximately 3 Wh, meaning that the world’s energy consumption is unlikely to sustain more than interactions a year. If someone were to allow as many people as possible to interact with GPT-3 a hundred times a day or even 300 hundred times a year (which is actually a little smaller!), then the world’s energy consumption would be barely enough to sustain interactors, which is less than 1.5 OOM higher than the population of China alone. What about smarter models like Claude or GPT4o? Am I mistaken somewhere? Or does it mean that AGI might replace just, say, half of the scientists, IT specialists and high-level art creators while letting the majority of humans live as normal?
P.S. There was an article on LessWrong that tried to promote a similar point: the human brains are already fairly well optimized, and a potential AGI is unlikely to exceed the same constraints, just like grey goo may be unlikely to be more efficient than life.
P.P.S. ChatGPT was installed over a hundred of millions of times, which is just 2.5 OOM away from the aforementioned estimate of . Does anyone know how often people actually use it?
Take into account exponential compute scaling and exponential efficiency scaling.
I mentioned increases in compute scaling in my other blogpost, which I messed up. The question is how efficiency is worth it and how many OOMs is the recent o3 model away from the AGI.