The Hated One has been pretty solid in the past regarding privacy/security, imho. I found this video of his rather enlightening and concerning.
- LLMs and their training consume a LOT of power, which consumes a lot of water.
- Power generation and data centers also consume a lot of water.
- We don't have a lot of fresh water on this planet.
- Big Tech and other megacorps are already trying to push for privatizing water as it becomes more scarce for humans and agriculture.
---personal opinion---
This is why I personally think federated computing like Lemmy or PeerTube to be the only logical way forward. Spreading out the internet across infrastructure nodes that can be cooled by fans in smaller data centers or even home server labs is much more efficient than monstrous, monolithic datacenters that are stealing all our H2O.
Of course, then the 'Net would be back to serving humanity instead of stock-serving megacultists. . .
Also, I can't even imagine how many resources image-generating AIs take up, especially when it's all based around "refining prompts" over and over and over....
I've been limiting my AI use ever more because of that, and I found out that AI isn't as necessary as they want us to believe. In fact, the downsides may be more than that small productivity boost we're gaining, in terms of rights to clear and verifiable informations and in terms of manipulative tactics and uses made possible by that.
It's a lot less than playing a video game. My fans on my GPU spin up harder and for more sustained time whenever I'm playing.
I think the training part is not to be neglected and might be what is at play here. Facebook has a 350k GPU cluster which is being setup to train AI models. Typical state of the art models have required training for months on end. Imagine the power consumption. Its not about on person running a small quantized model at home.
Such training can be done in places where there's plenty of water to spare. Like so many of these "we're running out of X!" Fears, basic economics will start putting the brakes on long before crashing into a wall.
That's not what I replied to though.
You said running an imagine generating AI on your GPU is less demanding than a video game. While possibly true, the topic of water scarcity and energy demands are not about what one person runs on one GPU, hence my response.
The person I replied to was only commenting on people how much it cost for people to "refine prompts" which isn't where the problems lie. People at home on consumer hardware can't be the ones causing issues at scale, which is what I pointed out. We weren't talking about training costs at all.
Besides, the GPU clusters models are trained on are far outnumbered by non-training datacenters, which also use water for cooling. It seems weird to bring that up as an issue while not talking about the whole cloud computing industry. I've never seen any numbers on how much these GPU clusters spend versus conventional use, If you have any I'd like to see them.