cthos 🐱Jun 18, 2025

So, re: LLMs and the "jury is out on energy usage", I just want to point out a couple of historical things.

When studies came out linking smoking to increased risk of cancer, the Tobacco industry paid for and produced studies that said "nuh uh".

When Obesity and heart disease rates increased in the United States and the problem was linked to sugar intake, the sugar industry produced studies saying it was fats.

https://pmc.ncbi.nlm.nih.gov/articles/PMC3490543/#sec4

Inventing Conflicts of Interest: A History of Tobacco Industry Tactics

Confronted by compelling peer-reviewed scientific evidence of the harms of smoking, the tobacco industry, beginning in the 1950s, used sophisticated public relations approaches to undermine and distort the emerging science. The industry campaign ...

PubMed Central (PMC)
Show threadcthos 🐱Jun 18, 2025

The fact that the GenAI industry isn't providing actual data about their energy usage should make you _very_ skeptical about their claims of "it's fine".

And I'm a little disappointed in seeing folks I generally respect being like "oh no, it's not as much as you think" - y'all need to show me hard data.

Also, I would fucking hope a text extruder uses less energy than serving 4k video, the fact that it presently *isn't* is a travesty - nevermind its other travesties.

Show threadAsta [AMP]

@cthos

The fact that the GenAI industry isn't providing actual data about their energy usage should make you very skeptical about their claims of "it's fine".my brain is just screaming "this this this", because the fact that they don't provide data while simultaneously pursuing access to larger and larger power sources is a huge red flag. If their claims were true, the data would be published, proudly. They would brag about it.

This isn't taking into account their partially externalized power costs for training and such, either; scrapers are massively increasing compute demands on other servers, which translates to higher power costs that they have no way of quantifying (and are definitely not interested in doing so).

Jun 18, 2025 at 4:18pmWeb
Show threadcthos 🐱Jun 18, 2025
@aud "Yeah, we totally need nuclear power but we pinkie swear we're making stuff more efficient. Oh look, new more power-hungry Nvidia cards!"
Show threadAndreas KJun 18, 2025

@cthos @aud Well, considering that nuclear power is one of the most expensive power sources, that kind of makes no sense.

Literally, even with massive state subsidies, they are still the most expensive per MWh. Without massive state subsidies, nobody would even consider building these crazy things. And if you factor in the costs of getting rid of the nuclear waste properly, the correct costs raise off the graph, even with a log scale.

So not exactly sure what the angle here is.

Show threadcthos 🐱Jun 18, 2025

@yacc143 @aud Correct, their claims about energy usage (when they bother to make claims) don't make sense. ^_^

To be accurate though I've only seen them trying to recommission existing reactors rather than building new ones so I was being *slightly* hyperbolic above but, wouldn't put it past them.

And if we look at Musk's xAI plant I think they just intend to pollute until someone makes them stop, so.....

Show threadAndreas KJun 18, 2025

@cthos @aud Ah, funny that, they probably want these new "advanced nuclear plants", that are supposed to be cheaper, but the only existing such plant, Plant Vogle, GA, 3&4 is estimated at LCOE $100-180/MWh while the advertised “prospectus” values for advanced nuclear plants are in $80-$85/MWh range.

Note that "analysts" (who partially admit that they don't consider all the financing costs) are on the lower end of the bracket, while the Princeton researcher gets a whopping $178/MWh.

Show threadAndreas KJun 18, 2025
For comparison, wind/PV are in the $30/MWh range, and you can locate data centres in consistent weather zones. With that energy usage, batteries, and reserve capacities literally pay for themselves (Hint, hint: Texas, the guys with their private grid, and thus grid balancing issues from hell, have been literally subsidising crypto miners to have a huge electricity user, that they can basically turn off instantly, if they need to balance the grid.)
Show threadAndreas KJun 18, 2025

With commercial financing btw, old style nuclear plants, when newly built have LCOEs slightly under $200/MWh

Even in the USA, these things are literally the most expensive form of generating electricity.

Show threadAndreas KJun 18, 2025

Ah, I'm more relaxed. At some point, somewhere the costs will have to meet the usefulness.

Yes, these things have a certain usefulness (especially when harnessed in higher level algorithms that keep them on track.)

But the question is, will the providers of these model be able to charge what it costs them (and make a somewhat relevant profit, capitalism lives for the profit).

And almost all applications that produce at least somewhat useful results use big contexts, which implies costs.