Because of the explosive growth of data centers, Dominion Energy utility in Virginia, USA, predicts that in 25 years, its carbon emissions will be 65% higher than it was expecting just two years ago.
https://www.businessinsider.com/ai-data-centers-land-grab-google-meta-openai-amazon-2023-12
It will be interesting to see if those projections turn out real (hopefully not). It looks to me like projections made now based on the needs of models like GPT-4 will be overestimates because developments in AI are heading away from gigantic foundational models and toward a multitude of specialized models.
On the other hand, even if the tech gets more efficient there may be way more demand and use-cases than expected, so AI resource use could skyrocket anyway.
@HumanServitor Even if there are efficiency gains, these historically have never lead to lower total energy and material use, because overall demand explodes.
Let's say today an AI search is 10 times more energy intense than a traditional search. Even if they have its energy intensity and claim "efficiency", it's still 5 times more wasteful than the old way, which was already using huge quantities of energy.
@gerrymcgovern I think that goes back to the idea of properly pricing energy to represent its real costs.
If that 10 times more intense AI search cost too much in electricity then it wouldn't be done, promoting lower total energy use.
In other words, since we've always underpriced electricity the incentives have never been there to really focus on lowering total energy use, and we could change that.
Demand will not explode if prices surge to reflect costs.
