“But it also takes a lot of energy to train a human,” Altman said. “It takes like 20 years of life and all of the food you eat during that time before you get smart. And not only that, it took the very widespread evolution of the 100 billion people that have ever lived and learned not to get eaten by predators and learned how to figure out science and whatever, to produce you.”
So in his view, the fair comparison is, “If you ask ChatGPT a question, how much energy does it take once its model is trained to answer that question versus a human? And probably, AI has already caught up on an energy efficiency basis, measured that way.”
If anything i think the better comparison is you use more power watching TV or gaming than you probably will using AI in the day if you do either of those 2 things.
The issue is training takes a lot of power, and because we can’t run the hardware locally our usage is also placed in these data centers which put pressure on a specific area instead of distributing the same power usage.
I saw a post a couple days ago about a company etching the model weights into silicon chip and they made a 8b model that could do 16k t/s and once made are relatively cheap to produce, and in power requirements, and would only get better. Just need to make sure they can be recycled well as they’d end up on a 1 to 2 year cycle like phones. Model to chip in 60 days they said.
So maybe that’s the future solution to distrubuted usage, but we would still need to solve training, but we could just mandate these datacenters must build their own renewable power and it would be less is everyone could run their own local inference.