1mo ago

Chinese AI tech could cut projected spike in US electricity demand.

floodlightnews.org Utilities to spend billions to power AI. Are they jumping the gun?

Chinese AI tech DeepSeek is forcing a second look at proposals to build a record number of new fossil fuel power plants

cross-posted from: https://ponder.cat/post/1486028

A new artificial intelligence model from China not only upended stock markets this week, it also called into question whether the rush to build new, mostly fossil-fueled power plants to run data centers is premature.

Technology @lemmy.zip

Cat @ponder.cat

1mo ago

Chinese AI tech could cut projected spike in US electricity demand.

floodlightnews.org /chinese-ai-tech-could-cut-projected-spike-in-us-electricity-demand/

16 0

You're viewing a single thread.

15 comments

I'd say if AI gets better and more efficient, it's likely to be used by more people and applied to more tasks. That would nullify hypothetical energy savings.
- More of the energy cost is in training rather than use.
  
  Sure. But more AI adoption also means more AI training or fine tuning, doesn't it? I mean the only option where training stops is if it doesn't make progress any more. So yes, if it's just one company and there isn't much competition I'm sure less training will happen. But competition kind of incentivises them to do more. Plus there are other tasks that could be done if we don't need to all focus on inefficient training of singular big chatbot models... Like multimodality, robotics, specific tasks... Making them more efficient for mobile use, or just more intelligent... I'm sure people will find other uses for the training farms. Or it'll just become more people. But yeah, we'll have to see.
- Exactly. There is no end to this and it‘s hardly uplifting news when it isn’t much of news but speculation instead.
- It's not supposed to be about reducing the price of energy, it's about reducing the nonmonetary costs that come with dedicating electricity to Language Learning Models, even if only on a per-person basis. Would you rather everyone use AI that requires 10 Watt-hours per person per day (times 8+ billion) or 2000 Watt-hours per person per day? Per capita energy usage reduction is still overall energy usage reduction.
  
  Sure. What I'm trying to convey is, I don't think it's a given fact that per capita energy consumption will be reduced. Take TVs for example. They've become way cheaper and more efficient since the 90s... But the dynamics are: People will now buy way larger TVs. And they're affordable to more people. So, I haven't looked at the numbers, but I'd be surprised if our TVs use less electricity now. Despite them having become more efficient.
  But ultimately I don't really care about the numbers or energy usage. It'd be nice not to wreck the climate and the planet. That's what I'm concerned with. So as long as we have the energy to spare, and it doesn't pollute the environment and harm the environment and us... I'm fine with AI using energy. But yeah, being efficient is a good thing. We shouldn't waste things if we can avoid that.
  
  Neither do I think that or disagree with you then. Fortunately, renewables are proving very cost-effective. The primary reason for these plants being fossil-fueled is because old money and new money are in bed with each other and want profit/kickbacks from making money off of oil and coal. AI is, of course, usually funded by said money and so they try to use mineral fuels to preserve the market value of mineral wealth.
  However, considering one can now run LLMs on a PC with no internet connection, I imagine most heavy computer users will encourage the public to make custom-built LLMs just like they currently do with desktop PCs.
  
  Yeah, I hope we someday manage to transition to renewables and get cheap and relatively clean energy. I'm living in a country which isn't sitting on huge oil reserves, so I'd say it'd be clever if we made an effort... And we kind of do. But it's probably a bit uncoordinated. And there are people lobbying for the opposite... (And seems it's a big undertaking.)
  I hope AI is going to get a bit more democratized in the future. And as you said, more efficient. It'll probably be a combination of factors. More efficient hardware, custom-built LLMs tailored to specific use-cases, scientific progress... I'd like more affordable hardware to run LLMs at home. I think something like Apple processors with their "unified memory" might be promising. I heard LLMs run pretty well on modern MacBooks, without any seperate gaming graphics card.
  And I'm not even sure how it'll turn out. Sure, the AI companies predict a high demand for AI, and they're building datacenters and need new power plants to power all of that. But I'm not totally convinced. Maybe that's part of the big AI hype, and it'll turn out there is far less demand than what they tell their investors. Or they're unable to keep up the pace and it'll take longer until AI is intelligent enough to do all the things they envisioned. AI will be some part of the world's electricity bill, though.

15 comments