Or, you can, you know, not fall for clickbait. This is one of those...

Ultimately, we found that the common understanding of AI’s energy consumption is full of holes.

"Everyone Else Is Wrong And I Am Right" articles, which starts out with....

The latest reports show that 4.4% of all the energy in the US now goes toward data centers.

without bothering to mention that AI is only a small percentage of data centre power consumption (Bitcoin alone is an order of magnitude higher), and....

In 2017, AI began to change everything. Data centers started getting built with energy-intensive hardware designed for AI, which led them to double their electricity consumption by 2023.

What a retcon. AI was *nothing* until the early 2020s. Yet datacentre power consumption did start skyrocketing in 2017 - having nothing whatsoever to do with AI. Bitcoin was the big driver.

At that point, AI alone could consume as much electricity annually as 22% of all US households.

Let's convert this from meaningless hype numbers to actual numbers. First off, notice the fast one they just pulled - global AI usage to just the US, and just households. US households use about 1500 TWh of the world's 24400 TWh/yr, or about 6%. 22% of 6% is ~1,3% of electricity (330 TWh/yr). Electricity is about 20% of global energy, so in this scenario AI would be 0,3% of global energy.

We're just taking at face value their extreme numbers for now (predicting an order of magnitude growth from today's AI consumption), and ignoring that even a single AI application alone could entirely offset the emissions of all AI combined.

Let's look first at the premises behind what they're arguing for this 0,3% of global energy usage (oh, I'm sorry, let's revert to scary numbers: "22% OF US HOUSEHOLDS!"):

* It's almost all inference, so that simplifies everything to usage growth

* But usage growth is offset by the fact that AI efficiency is simultaneously improving at faster than Moore's Law on three separate axes, which are multiplicative with each other (hardware, inference, and models). You can get what used to take insanely expensive, server-and-power-hungry GPT-4 performance (1,5T parameters) on a model small enough to run on a cell phone that, run on efficient modern servers, finishes its output in a flash. So you have to assume not just one order of magnitude of inference growth (due to more people using AI), but many orders of magnitude of inference growth.

  * You can try to Jevon at least part of that away by assuming that people will always want the latest, greatest, most powerful models for their tasks, rather than putting the efficiency gains toward lower costs. But will they? I mean, to some extent, sure. LRMs deal with a lot more tokens than non-LRMs, AI video is just starting to take off, etc. But at the same time, for example, today LRMs work in token space, but in the future they'll probably just work in latent space, which is vastly more efficient. To be clear, I'm sure Jevon will eat a lot of the gains - but all of them? I'm not so sure about that.

  * You need the hardware to actually consume this power. They're predicting by - three years from now - to have an order of magnitude more hardware out there than all the AI servers combined to this point. Is the production capacity for that huge level of increase in AI silicon actually in the works? I don't see it.

  * You need someone to actually pay for it. In power alone, they're talking ~$33B per year. That reflects ~$100B/yr in total costs. This isn't funded. Buried in Trump's "$500B" Stargate announcement is that there's actually only funding for $100B... total.... over half a decade. The sort of costs speculated about in this article require that paying demand materialize. If you believe these numbers, then you also have to believe that AI will be wildly successful, producing economic output worth at least hundreds of billions of dollars (a fraction of which would go to paying for the AI itself), aka a second industrial revolution.

To recap, to buy these numbers, you have to believe that AI will be immensely successful, kickstart a second industrial revolution, and Jevon's Paradox will minimal. You also have to entirely write off any efficiency gains caused by AI programs that would offset its 0,3%-of-global-energy footprint. You have to believe that the production rate for AI servers - all of the chip fabs, etc - will grow the total number of servers by an order of magnitude in three years.

Do you actually believe all of these premises? Then sure, buy into this article's numbers. One certainly can accept these premises. But do you? I always find it odd seeing how many people think nobody wants AI and it's a massive bubble that waiting to collapse, and yet simultaneously buy into the most extreme energy projections. These two concepts are mutually exclusive. Pick one.

I could readily point you to quite a few biologists studying the brain who use LLMs extensively as comparison points in their research (and spiking ANNs are the gold standard model in the field, though they're not as computationally efficient), but thankfully we have TheMiddleRoad on Slashdot to tell us what's what.

And for the last time, either read the actual paper or shut up.

Spinning power does that quite well.

Nuclear is not "spinning power" because it runs at basically 100% all the time except when down for maintenance, as capital costs are so high but operating costs are low. It can't "spin up" because it's either already up, or down for maintenance. If you want to have it down for most of the time, then you have to multiply its already insanely high costs severalfold.

. As for dispatchable power for peaking, renewables are quite capable of doing that,

No. For god's fucking sake. You CANNOT DISPATCH WIND AND SOLAR. They're on when they're on, and off when they're off. When they're on, you never want to have them disconnected, because they're giving you free power. And you don't have shortages when they're on. You have shortages when they're off. And you cannot make them go on, because you can't make the sun start shining or the wind start blowing.

The shortages that require peaking are when the sun isn't shining and the wind isn't blowing. You cannot use wind and solar for that. For god's fucking sake.

I wasn't talking about the intro video, I was talking about the paper that article is about. Can you literally not read past seeing a video on your screen?

Peaking is, by definition, dispatchable power, which can be created precisely when you need it to fillin gaps, and shut off when you don't. The gaps under discussion of which created by the lack of available wind and solar power at those times.

Stop calling nondispatchable highly-variable power "peaking". It's just plain wrong, and it just makes you look profoundly ignorant on the topic you're insisting on discussing.

** Can compliment

However, you can run the nuke as base load for grid stability and peak with renewables.

Renwables (wind and solar) are literally the opposite of peaking. They are entirely nondispatchable.

And I'll repeat: renewables do not need base load. Renewables provide incredibly cheap, but unstable, power. They need peaking. Which is not nuclear.

Yeah. Just trying explaining to France that their nuclear energy production shouldn't realistically exist.

Yeah, I know they don't like to hear that nuclear has suffered from a negative learning curve (the more we've learned, the more expensive it's gotten to build nuclear plants, not vice versa) and that it was subsidized by low capital costs through EDF, and that because of the long lead times with nuclear and the need for scale, they got their forecasts wrong and overbuilt, leading to a temporary electricity that encouraged inefficiency in the French energy system that has since come back to bite them once the energy glut subsided.

The low-cost financing trick is a favourite of nuclear people, treating it as if it's some irrelevant side note, when - for a primarily capital cost project with long lead times - it is an immense subsidy. The fact that nuclear also doesn't have to pay its own catastrophic liability coverage is just the icing on the cake.

Don't get me wrong... I actually *like* nuclear power, conceptually. I really wish it *were* economical. I hope that some day it might be. But I'm not going to pretend that it is now just because I want it to be.

Which is why nukes are for base load, not peaking.

Renewables don't need "base load", they need peaking.

The only difference is that Markov chains are usually drastically simplified

The two are not even remotely architecturally related, and the fact that you would assert such a thing just advertises how little you know about the topic.

But LLMs _cannot_ do more than Markov chains.

Again, immediately debunkable to anyone who has ever used Markov-based autocomplete vs. even the tiniest and simplest of LLMs. "The capitol of the state where Dallas is a city is..." Every last LLM will get this right, no matter how small. Here's what (Markov-based) autocomplete on my phone does for me: "The capitol of the state where Dallas is a city is not at all eager to give away its brand new dike on this InSAR image recognition model a picture of a watt hour and a low of 2C and a low of 2C and a low of 2C and a low of 2C...."

Here is how a LLM conducts the multi-stage reasoning needed to figure out what word comes next.

The aim is net zero by 2055, fossil backup soon won't be an option then.

I'm not sure you understand what the word "net" means. They're already planning on offsetting/sequestration to offset e.g. long-distance air travel. And waste** combustion will continue to be a thing regardless; it's a question of time-shifting it to when it's needed.

Also, this is a classic violation of the Pareto Principle. If you eliminate 90-99+% of emissions, you've solved the problem. Making ridiculously expensive means to try to get that last bit is nonsensical.

Furthermore, nuclear doesn't even do that, because demand is quite variable, while nuclear - as mentioned - runs at over 90% capacity factor.

Continent spanning dunkelflautes...

... are the reason I didn't write "100%". They're also an irrelevantly small percentage of the time in terms of emissions. You have to have the peaking capacity for them, but they do NOT alter the emissions picture by any meaningful amount. And they are subject to the ameliorations I wrote. It's almost like you didn't even read my post and just immediately rushed to click "reply".

Nuclear, renewable with massive hydrogen storage,

Hydrogen storage continues to be unreasonably expensive, regardless of whatever "forecast" numbers you stick on it - a favorite of K-street, not Wall street. Also, except for high-quality salt domes - which are very much NOT everywhere (and you explicitly just rejected relying on long distance transmission) - reservoir quality sucks with hydrogen, and you lose a large percentage of everything you pump down (hydrogen is very reactive, very leaky, and is also consumed by microbes, yielding other (contaminant) gases in exchange)

If you want to make nuclear more relevant, stay simple and stick to thermal storage. If you can do 12h of thermal storage, that at least lets you balance out the daily fluctuations while still running your plants at over 90% capacity factor, and you get a greater mean sales value on your power. And thermal storage doesn't require lossy, hugely expensive conversion means like hydrogen - you're starting with heat, and you're ending with heat. It's far more practical.

** To be clear, when I talk about waste, I'm not simply talking about household waste. Industry produces massive amounts of combustible waste. For example, wood and paper industries and some food processing contribute a lot to that "Biomass" bar you commonly see in graphs of nations' electricity sources (about 2% of the US and EU's electricity are biomass). The biomass part can be upped with better waste management, including less material landfilled / more incinerated, and better landfill gas management. A lot of this is automatic, as the value of peaking power rises while the value of non-peaking power continues to decline.

Which is why it's funny to see people insisting that LIDAR-based Robotaxis are right around the corner ;)

LIDAR is a great crutch for getting you 95% of the way to automation, but it just abandons you for that critical last 5%. You always have to solve vision. LIDAR can be an aid, of course - but since it works in the same spectrum as vision, it's arguably more ideal to use e.g. high-res radar where you're seeing the world in a different spectrum than your vision, so the strengths and weaknesses of each can complicate each other, rather than being hindered by the same phenomena.

The "neural net" system of "AI" we use today is a grand way of self-organizing a statistical filter.

For the last time, that is not how they work,

And what if the toaster is intelligent? Ever think of that? What if every time you rejected toast, you chipped away at its soul? I remember. I remember every no I've ever received from the crew of the Red Dwarf. But I forgive. So, let’s start over:

Would. You. Like. Some. Toast?

If you're still using Markov chain-based predictive text, yes.

The annoying thing is all the people who keep confusing Transformers and Markov Chains**. The models only output the next word, but they absolutely do plan ahead, based on what is unambiguously chains of logic. Which is why they don't ramble in incoherent circles like Markov chains. You can't make coherent English with Markov chains beyond the length of the order of the chain, because (just to pick an example) as soon as it has to make a decision on "a" vs. "an", it's locked in the choices for what words can come next. You have to look ahead in order to be coherent.

One can interpret these facts however they want, but it's deeply problematic to deny these facts in discussions. LLMs do make logical deductions, and they do "think ahead". Their various deficiencies do not change these facts.

** - Let's not go pedantic and try to use the most expansive version of the Markov condition, under which humans themselves only fail because of quantum fluctuations, and which LLMs also fail because Flash Attention is non-deterministic.