You’re absolutely right! According to the research you cited, the energy use is actually much LOWER than I stated in my comment.
Your source shows that an efficient AI model (Qwen 7B) used only 0.058 watt-hours (Wh) per query.
Based on that, my entire 3-prompt chat only used about 0.17 Wh. That’s actually less energy than a single Google search (~0.3 Wh). Thanks for sharing the source and correcting me.
If one assumes a 1/3 correctness is sufficient and the provider is using a 7B model, it is a safe assumption that it was energy efficient and better than a traditional search. However, on the other end of the spectrum, if one assumes the most efficient reasoning model, which consumes ~400x more energy and still only amounts to 4/5 accurate responses, the entire discussion is flipped on its head.
It is however comical to see one jump to an irreproducible edge case to prove one’s point, it does really exemplify how weak the position was from the beginning. Intellectual dishonesty galore.
edit: the supplied reply is also highly unlikely to have come from a non reasoning model given the structuring of the text.
I’d be curious to have the exact 3 prompts to input into Qwen7B and get that exact response.
This is not correct and can easily be disproven, even if one assumes less than 480g/Kwh.
And that is ignoring the infrastructure necessary to perform a search vs AI query.
You’re absolutely right! According to the research you cited, the energy use is actually much LOWER than I stated in my comment.
Your source shows that an efficient AI model (Qwen 7B) used only 0.058 watt-hours (Wh) per query.
Based on that, my entire 3-prompt chat only used about 0.17 Wh. That’s actually less energy than a single Google search (~0.3 Wh). Thanks for sharing the source and correcting me.
If one assumes a 1/3 correctness is sufficient and the provider is using a 7B model, it is a safe assumption that it was energy efficient and better than a traditional search. However, on the other end of the spectrum, if one assumes the most efficient reasoning model, which consumes ~400x more energy and still only amounts to 4/5 accurate responses, the entire discussion is flipped on its head.
It is however comical to see one jump to an irreproducible edge case to prove one’s point, it does really exemplify how weak the position was from the beginning. Intellectual dishonesty galore.
edit: the supplied reply is also highly unlikely to have come from a non reasoning model given the structuring of the text.
I’d be curious to have the exact 3 prompts to input into Qwen7B and get that exact response.