AI already uses as much energy as a small country. It's only the beginning(vox.com)
vox.com
AI already uses as much energy as a small country. It's only the beginning
https://www.vox.com/climate/2024/3/28/24111721/ai-uses-a-lot-of-energy-experts-expect-it-to-double-in-just-a-few-years
13 comments
It’s not like the workers that got their jobs replaced by an LLM cease to exist. Unless LLMs caused people to work more from home, or have fewer kids, I can’t see how they affected the environment positively.
Small countries: Blockchains were at the level of The Netherlands at some point. So we’re not talking city state, but real, industrialised country.
Small countries: Blockchains were at the level of The Netherlands at some point. So we’re not talking city state, but real, industrialised country.
> It’s not like the workers that got their jobs replaced by an LLM cease to exist. Unless LLMs caused people to work more from home, or have fewer kids, I can’t see how they affected the environment positively.
Say you're improving computer vision detection of material defects. If you replace an old inefficient CPU-based algorithm running on a server drawing 500 watts with a detector network on an NVIDIA Jetson drawing 5 watts - in isolation you have drastically improved energy efficiency.
But what if, instead of being decommissioned, that server's CPU cycles are reallocated to some other task that still needs them? Was what you did still an improvement?
I don't have a definite answer for how we should think about it, but my gut feeling is yes. You've taken the energy requirements of a task from 500W to 5W, and thus decreased energy required per abstract unit of overall productivity. It happened to be realized as an increase in productivity (old job + new job now take roughly same energy as old job alone) rather than a decrease in energy usage, but even if only considering the environmental impact I think that's now a better position to be making progress towards green solutions.
Say you're improving computer vision detection of material defects. If you replace an old inefficient CPU-based algorithm running on a server drawing 500 watts with a detector network on an NVIDIA Jetson drawing 5 watts - in isolation you have drastically improved energy efficiency.
But what if, instead of being decommissioned, that server's CPU cycles are reallocated to some other task that still needs them? Was what you did still an improvement?
I don't have a definite answer for how we should think about it, but my gut feeling is yes. You've taken the energy requirements of a task from 500W to 5W, and thus decreased energy required per abstract unit of overall productivity. It happened to be realized as an increase in productivity (old job + new job now take roughly same energy as old job alone) rather than a decrease in energy usage, but even if only considering the environmental impact I think that's now a better position to be making progress towards green solutions.
Yes, but replacing a CPU with a Jetson is not the same as replacing a human worker with a GPU.
A CPU can be decommissioned and turned off, or get repurposed, while a human still generates CO2 at an equivalent rate as before being let go.
My intuition tells me that most “… and AI!” initiatives just add more energy use where there was none. Which is the point this article makes, too.
Maybe we’re happier and more prosperous for it. Maybe it does not justify contributing to global warming any more than derpcoins.
A CPU can be decommissioned and turned off, or get repurposed, while a human still generates CO2 at an equivalent rate as before being let go.
My intuition tells me that most “… and AI!” initiatives just add more energy use where there was none. Which is the point this article makes, too.
Maybe we’re happier and more prosperous for it. Maybe it does not justify contributing to global warming any more than derpcoins.
> A CPU can be decommissioned and turned off
My analogy starts with the case where this is true and the energy efficiency gain is clear, but only as a comparison point from which to consider the case where "instead of being decommissioned" the server's CPU cycles are used for something else (corresponding to humans choosing different jobs, or even just recreation). I argue that even in this case it's an environmentally positive change.
My analogy starts with the case where this is true and the energy efficiency gain is clear, but only as a comparison point from which to consider the case where "instead of being decommissioned" the server's CPU cycles are used for something else (corresponding to humans choosing different jobs, or even just recreation). I argue that even in this case it's an environmentally positive change.
hmm I'm having a hard time thinking AI is going away.
It feels like pandora's box has opened a bit.
So I guess should invest in energy companies ?
Human bodies also use A LOT of energy, in scale of the whole humanity. For starters, that would be around 2kWh/person/day for each adult.
How much less energy does a query against an 8B model take, vs a 70B model? Can we get more clever about using smaller, more specialized models?
Replace "AI" with "post-industrial capitalism" and suddenly all this fear mongering seems insignificant.
Definitely Crypto 2.0.
What is the implication here, ban anything that uses more energy than a small country? No more eating, driving, traveling, or building for humanity.
> Included for the first time were projections for electricity consumption associated with data centers, cryptocurrency, and artificial intelligence.
Also no mention of any energy usage that might be displaced by AI. Pretty sure firing off a bunch of LLM API calls is less energy intensive than hiring workers who will drive to the office, etc.
I also think that there’s a vast number of things that qualify as “using as much energy as a small country”. What’s the energy usage of the Vatican?