[unltdpower]

"Committing to buying the glass to replace the window I broke in your shop to rob the place, you're welcome."

> Training a single frontier AI model will soon require gigawatts of power, and the US AI sector will need at least 50 gigawatts of capacity over the next several years.

These things are so hideously inefficient. All of you building these things for these people should be embarrassed and ashamed.

[keeda]

> These things are so hideously inefficient.

Quite the opposite, really. I did some napkin math for energy and water consumption, and compared to humans these things are very resource efficient.

If LLMs improve productivity by even 5% (studies actually peg productivity gains across various professions at 15 - 30%, and these are from 2024!) the resource savings by accelerating all knowledge workers are significant.

Simplistically, during 8 hours of work a human would consume 10 kWH of electricity + 27 gallons of water. Sped up by 5%, that drops by 0.5kWH and 1.35 gallons. Even assuming a higher end of resources used by LLMs, a 100 large prompts (~1 every 5 minutes) would only consume 0.25 kWH + 0.3 gallons. So we're still saving ~0.25 kWH + 1 gallon overall per day!

That is, humans + LLMs are way more efficient than humans alone. As such, the more knowledge workers adopt LLMs, the more efficiently they can achieve the same work output!

If we assume a conservative 10% productivity speed up, adoption across all ~100M knowledge work in the US will recoup the resource cost of a full training run in a few business days, even after accounting for the inference costs!

Additional reading with more useful numbers (independent of my napkin math):

https://www.nature.com/articles/s41598-024-76682-6

https://cacm.acm.org/blogcacm/the-energy-footprint-of-humans...

[wlesieutre]

So with the AI is doing more of the work and you need less humans, what are you doing with the extra humans to eliminate their no-longer-productive resource consumption?

Saying “we can do the same work with less resource use” doesn’t mean resource consumption is reduced. You’ve just gone from humans using resources to humans using the same resources and doing less work, plus AI using more resources.

[pmontra]

Resource consumption often goes up. It's a time vs energy tradeoff and it's not free.

Your question is a variant of what do we do with all those humans now that they don't have to walk miles to the well every day because we invented aqueducts? The point is that they didn't want to walk to the well but they had to (and in some places they still have to) and very few people want to work, even now and even us, but they have to.

We will see what happens this time when we won't have to walk to that well.

[danans]

> So with the AI is doing more of the work and you need less humans, what are you doing with the extra humans to eliminate their no-longer-productive resource consumption?

Soon enough, we won't be able to avoid this question.

[whatshisface]

You put them to work doing more things than were possible in a month before.

[keeda]

The thing is, there are many interplaying dynamics here that are impossible to unravel. This is why I called it "napkin math", because figuring out the full ramifications of this change is a pretty large economic problem that nobody has figured out!

For instance, I think operating at this level of productivity is unsustainable (https://news.ycombinator.com/item?id=46938038). As discussed in detail by the recent "AI vampire" blog: https://news.ycombinator.com/item?id=46972179 -- most humans are not designed for that level of cognitive intensity.

But even then, the productivity per human will explode, and we will still have the problem of "too many humans." Cynically, if most knowledge workers get laid off, it's good from an environmental perspective because that means much less commuting and pollution! But then they're starving and we will have riots!

This is where I foresee the near-term problems with GenAI: social turmoil rather than resource consumption. I suspect it's not all bad news though. While it's impossible to put numbers on it, it helps to think about the first-order economic principles that are in play:

1. This is hand-wavy, but knowledge work boosts economic growth. If this is massively accelerated, we should be creating surplus value that compensates for a lot of costs.

2. However a huge chunk of knowledge work is busy work which will be automated away. People can try upskilling but the skill gap is already huge an growing quickly and they will lose jobs.

3. The economy is essentially people providing and paying for services and goods. If people lose jobs and cannot earn, they cannot drive the economy and it shrinks.

4. The elite, counter-intuitively enough, do NOT want that because they get richer by taking a massive cut of the economy! (Not to mention life in a doomsday bunker can get pretty dull if starving people start rioting -- https://news.ycombinator.com/item?id=46896066)

There are many more dynamics at play of course, but I think an equilibrium will be found purely because everyone is incentivized to find a solution (UBI?) that keeps both the elites and the plebes living long and prospering. I expect some turmoil, but luckily, the severe resource crunch of GPUs gives us time to figure things out.

[keybored]

What I gather from your analysis and gleeful exclamation marks is that I should start rioting now rather than wait.

[keeda]

Hah, interesting, I intended my exclamation marks come across as surprise or emphasis rather than glee!

If it helps better interpret my posts, these days my frame of mind is "We're living in very interesting times" in the sense of https://en.wikipedia.org/wiki/May_you_live_in_interesting_ti...

[keybored]

It’s all so interesting!

Have fun watching the car crash.

[_kb]

Turn them into biogas to create more energy for DCs.

[danielbln]

Do keep in mind that 1 large prompt every 5 minutes is not how e.g. coding agents are used. There it's 1 large prompt every couple of seconds.

[keeda]

True, but I think in these scenarios they rely on prompt caching, which is much cheaper: https://ngrok.com/blog/prompt-caching/

I have no expertise here, but a couple years ago I had a prototype using locally deployed Llama 2 that cached the context (now deprecated https://github.com/ollama/ollama/issues/10576) from previous inference calls, and reused it for subsequent calls. The subsequent calls were much much faster. I suspect prompt caching works similarly, especially given changed code is very small compered to the rest of the codebase.

[gosub100]

Are you excluding the cost of training the AI from the calculation?

[keeda]

In the initial analysis of a single worker, yes, but when scaling up per-human savings to use by the wider population, the aggregate resource savings compensate for training resource usage within a few days, weeks at most.

[what]

How is a human consuming 27 gallons of water in an 8 hour work shift?

[keeda]

This includes things like drinking, sanitation, etc. Derived the number from here: https://www.epa.gov/watersense/statistics-and-facts

Mostly lines up with this reference too, which focuses only on water usage at work: https://quench.culligan.com/blog/average-water-usage-per-per...

[DoctorOetker]

Since their example scales the water consumption with their electricity consumption one may conclude it was the fresh water consumed (evaporated) during production of the electricity. Gaseous H2O is an even more potent GHG than CO2.

[missingdays]

One burger for lunch is 660 gallons of water. So 27 is actually a very huge underestimation

[Dylan16807]

> "Committing to buying the glass to replace the window I broke in your shop to rob the place, you're welcome."

Buying electricity isn't inherently destructive. That's a very bad analogy.

> These things are so hideously inefficient. All of you building these things for these people should be embarrassed and ashamed.

I'm not arguing that they are efficient right now, but how would you measure that? What kind of output does it have to make per kWh of input to be acceptable? Keep in mind that the baseline of US power use is around 500GW and that currently AI is maybe 10.

[epistasis]

Adding new electricity demand to the grid should not be viewed as breaking windows and robbing others. When I bought an EV, I increased my electricity demand a huge amount, but it's not like I'm stealing from my neighbors. No rules were broken. We just need to make sure that I pay enough for my additional demand.

> AI sector will need at least 50 gigawatts of capacity over the next several years.

The error bars on this prediction are extremely large. It would represent a 5% increase in capacity in "the next several years" which is only a percent or two per year, but it could also only be 5GW over the next several years. 50GW represents about 1 year of actual grid additions.

> All of you building these things for these people should be embarrassed and ashamed.

I'm not building these things, and I think there should be AI critique, but this is far over the top. There's great value for all of humanity in these tools. The actual energy use of a typical user is not much more than a typical home appliance, because so many requests are batched together and processed in parallel.

We should be ashamed of getting into our cars every day, that's a true harm to the environment. We should have built something better, allowed more transit. A daily commute of 30 miles is disastrous for the environment compared other any AI use that's really possible at the moment.

Let's be cautious of AI but keep our critiques grounded in reality, so that we have enough powder left to fight the rest of things we need to change in society.

[esafak]

How are you measuring efficiency? They're better than most humans, which is what I would need more of as a substitute.

[digitalPhonix]

A human consumes about 100 watts when not doing any physical exertion (round number, rule of thumb). So unless you can show an LLM running on 100w compute with capabilities similar to a human, they’re less efficient.

[pmontra]

100 W is only the start. Let's say that I consume 100 W all along the day. I use an LLM for coding assistance in the old way of asking questions and copy pasting code. It's much faster than me at writing that code. I don't think it ever works 1 hour for me per day. It's probably 10 cumulative minutes, probably much less. Round it up to 12 minutes to make it 1/5 of a hour or round it down to 6 minutes for a 1/10. So instead of 24 it's 0.1 hours, 240 times less. Those 100 W could be 24000 W and the total power per day would be the same. Is that LLM consuming 24 kW when working for me? No idea but I hope it's less than that.

Of course I could do all of my coding alone again, but I would be slower. It's like walking to the mall several times per week, several hours per time, instead of once or twice per week with a car, three cumulative hours. I trade a higher energy consumption for more time to do other things and the ability to live far away from shops.

[digitalPhonix]

If you as a human are coding for 24 hours a day as the benchmark for LLM efficiency we have other problems.

[pmontra]

I believe that we consume 100 W on average no matter what we do, except intense physical activities, which consume more.

[digitalPhonix]

Right, but you do stuff other than work 24 hours a day, right? You have fun, relax, etc.

Counting the 100w for 24 hours for a human doesn’t match up with counting the power usage from “AI” for only the 10 minutes it’s doing a task.

Also - units issue: 100 watts for a day is 2400 watt-hours. It’s a moot point anyway because the power draw for the frontier models is an order of magnitude off that the division by 24 is basically meaningless.

[dudisubekti]

Are you against inefficiency or just LLMs? If it's the former, I assure you LLMs are nowhere near the top of the list lol

You should start from beef industry.

[Andys]

Every piece of progress looks like this to begin with.

[measurablefunc]

The numbers must go up, there is no other way.