Powering AI could use as much electricity as a small country | Hacker News Reader

Powering AI could use as much electricity as a small country

techxplore.com

24 pointswglb36 comments

This is such a strange metric. When light bulbs were first invented, they also used more electricity than entire countries. So what? All that means is countries have very different levels of energy consumption.

According to Wikipedia, China uses 118,000 times more energy than, say, Nauru. Okay, and?

Is A.I. doing useful work in exchange? Absolutely, especially when compared to crypto. On a per output basis it's probably also more energy efficient than human writers and artists who wastefully metabolize cow products.

I don't know that we should be measuring productivity on an output per watt-hour basis to begin with, but if we did, we should at least put it in context.

Training an AI costs 40 home-years worth of energy... that's it?! How is that even newsworthy? The value AI gives us, now and in the future, is way way greater than a random selection of 40 households. Energy well spent.

If we truly cared about energy efficiency, human metabolism is incredibly inefficient. Do we really want to go down the path of comparing algorithms versus biological power usage? It's not going to look pretty for us.

Put another way, if ChatGpt uses 500 MWh a day, that's about a quarter tank of single 747's fuel. It's not even a rounding error.

Is it worth it? Probably. Is it significant? I'd say no.

https://en.wikipedia.org/wiki/List_of_countries_by_total_pri...

Strange metric indeed. "As much as a small country" is about as useful as saying "as big as a small rock".

I think its somewhat settled these days that people should be able to eat even if they are not productive, right? Like, wouldn't all the artists and writers still need to inefficiently metabolize things, even if out of a job?

Yes, and that's part of why I don't think we should measure productivity on a output per watt hours basis.

Humans are terrible at S3 standby.

What do you think would be better? How many LSAT question's answered per hour?

If it were up to me, cat memes per BTU?

> This is such a strange metric.

This metric has been used for years by dstractors of cryptocurrency for years without anyone batting an eye.

Because crypto produced nothing of value. AI does.

That's funny. I asked ChatGPT if crypto produced anything of value and it seems to disagree with you.

Hell, I disagree with myself if you give me a week or two to forget.

A small country to power something that you can run on a Raspberry Pi using sqlite is another level ;-). All because there is no one you can trust.

Christmas lights also use as much electricity as entire countries.[1] If the energy usage is what these people actually care about... where's all the moral outrage about Christmas lights?

Obviously something used by millions of people will use a proportionate amount of energy. Not sure why that's so surprising to people.

[1]: https://phys.org/news/2015-12-christmas-energy-entire-countr...

That’s not bad considering they use way more on crypto in which the usefulness is bordering on nil so far

no, that's your knowledge that is only nil. millions use it already. at very basic level it helps sci-hub funded, but u can't even comprehend that. It also help millions access to digital usd.

Figure of speech my man

> in which the usefulness is bordering on nil so far

Not at all.

The usefulness of AI is borderline zero too.

> I see a future where teaching jobs are extinct and everybody learns with some AI/LLM that can tailor a unique learning experience to each student based on their current knowledge level and learning style instead of having to constantly cater to the lowest common denominator in classrooms. Imagine every child anywhere in the world whether its some poor malnourished African kid, or Elon Musks 50th child, they will all have access to the same high-quality education. Wont matter how wealthy you are or where you are located, everybody will have the same opportunities to educate themselves.

Excerpt from conversation I had with someone

I would pay thousands of dollars for this, literally. I'm currently paying many thousands of dollars for fairly mediocre real life human teachers who barely try to understand my question before they repeat what they already said.

ChatGPT isn't there yet, but I'm eager for the day when I can count on some future AI to give me correct answers to questions on advanced topics.

Sure, that said it seems like major AI co's are some of the ones most interested in using clean energy to power their work, so probably not the worst thing in the world.

[1]: https://www.theverge.com/2023/9/26/23889956/microsoft-next-g...

Renewable or clean power does mean unlimited power.

If the demand for power is outpaces the supply, the cost goes up. And not only does this incentivize continued use of worse sources, it makes other uses of power more expensive.

An AI boom may be an even larger and more lasting strain on power markets than crypto was during its biggest spikes. Given that we don’t really know where this tech will get us, that could be bad. It’s worth tracking and thinking about.

It could also motivate more renewable installations. Over-building on the energy production side should make the energy storage side easier (won’t need to dip into the batteries as often if we can power everything even when the panels are only operating at 50%).

Currently I guess they probably run full out all the time. But really AI training seems like it ought to be a pretty shiftable workload.

A workload that can produce good economic returns while eating up as much power as we have available during excess times, but which can be shut off easily when we need to shave off the top of the demand peak, could be really nice for the grid.

Hopefully Altman and Microsoft backend Helion can solve this problem. Excited to see the future of that company as well.

Lots of comments here mention bitcoin. Here's something funny: the author of the paper mentioned in the article owns a website called Digiconomist[0], which seems to have been exclusively focusing on "environmental impact of cryptocurrency". Their latest article is about AI. Looks like a pivot to me.

0: https://digiconomist.net/

And? This sort of article are trying to feed a narrative & misguide the lesser educated / non-critical thinkers. It's got no place on this site.

If we weren't to burn coal, but rely on nuclear / wind / solar, and amp up recycling across the board, how much power we use when and for what, wouldn't be a question. So that is the topic that needs attention, and advocating.

> This extreme scenario is unlikely to happen in the short term because of the high costs associated with additional AI servers and bottlenecks in the AI server supply chain, de Vries says.

If we suddenly start ignoring the costs, it could get very expensive, indeed.

It's a good thing we have this thing called the sun, and storage mediums such as batteries and kinetic energy storage systems now widely in place and growing exponentially.

I think I’ve just solved the problem of what to do with people who lose their jobs because of AI.

Imagine a power plant that is just vast rows and rows of stationary bicycles. We can pay people to ride the bikes and produce energy to power AI. They get paid by the watt. They could wear masks with tubes to sequester their CO2 output. A small country’s worth of people could generate immense clean energy.

The cost of AI services are obviously going to reflect the energy inputs.

Unlike blockchain, AI actually has some utility in the non-criminal economy. I don't think we need to worry too much about the energy expenditure at this point.

It would be great if these economists and other non-practioners would consult any legitimate domain expert before writing these dumb columns. David Patterson has been writing and presenting on this topic a lot in the last 2-3 years because some widely-cited paper overestimates the cost of training and inference by between 50x and 1000000x depending on how you look at it. His deck is pretty interesting: https://chips-compilers-mlsys-22.github.io/assets/slides/10%...

All this energy and waste so I can talk to my computer and get it to write me an essay. Pointless. Expend the energy automating away sweatshop jobs.

Seems like a better use than bitcoin mining, at any rate.

That's not a high bar. =)

Is this even a concern? I occasionally see headlines like this but with all the energy resources available to us (clean and fossil), I don't know why it even matters to discuss how much energy a new tech consumes.

My argument is that the servers that run AI or crypto algos will still exist and operate even if AI or crypto went away.

Reply on news.ycombinator.com