> For these models today, if we measure the amount of energy expended for training and inference how do humans compare?
My best guess is 120,000 times more for training GPT-4 (based on claim it cost $63 million and that was all electricity at $0.15/kWh and looking only at the human brain and not the whole body).
But also, 4o mini would then be a killowatt hour for a million tokens at inference time, by the same assumptions that's 50 hours or just over one working week of brain energy consumption. A million tokens over 50 hours is 5.5 tokens per second, which sounds about what I expect a human brain to do, but caveat that with me not being a cognitive scientist and what we think we're thinking isn't necessarily what we're actually thinking.
