The Obscene Energy Demands of A.I
newyorker.com
newyorker.com
Bitcoin: 145B = 145,000M kWh/year
ChatGPT: 0.5M * 365 = 182M kWh/year
Based on the numbers from the article, ChatGPT is using 3 orders of magnitude less electricity for something that provides high utility for vastly more people.
There are of course many other uses of AI aside from ChatGPT and more cryptocurrencies aside from BTC, but these are very different power consumptions.
>Ethereum will consume 6.56 GWh of electricity annually. To put that into perspective, the annual electricity consumption of the Eiffel Tower is 6.70 GWh
There is no need to hate all crypto because you hate the energy consumption. It’s as easy as people buying and using coin #2 instead of #1 and the problem all goes away. The only reason Bitcoin continues to be mined is that it is profitable. If the coin value goes down, mining goes down.
We have a really nice mechanism of distributed utility measurement called "prices" which suggests otherwise.
Here's an estimation: Video card uses 200W for around 8-10 seconds to generate an image. Let's say 10. That's 6 images/minute, 360 images/hour, 1800 images per kwh.
At 0.20 Euro/kwh in Europe (so not really cheap), an image costs 0.00011 Euro to generate.
Let's say we're further doing a boring brute force approach of choosing the best image out of 100. This puts our costs at 1 cent per final image.
This IMO compares extremely favorably to almost all alternatives. Drawing by hand for hours, using paints or inks, using good paper, etc, will easily cost more than 1 cent per picture.
And AI can be done purely on renewables, while inks, paints and paper have ecological costs.
Anecdotally, 10K sheets of paper per tree, so almost a thousand trees worth in just paper, before accounting for the other materials, the impacts of their extraction and logging and so on.
I think hardware wins there as well, and the more tech advances the more uneven it gets. Making paper and pigments is a purely physical thing, we're already about as good at that as we can be. And probably getting worse with time, as the most convenient mines of materials are exhausted.
If you have 10k H100s, the failure rate is about 3-5/day. In numbers, that is $90-150k/day or ~11%/year (3*365=1095) on the low end. Of course a lot of these cards can go off for RMA and come back to life eventually, but this is being actually realistic about it. You end up having to manage a whole RMA program and it is a constant stream of cards in and out.
One particular example would be generating an image of a cloud. There are billions of such images in existence, but you could also use a model to generate one, which would take several orders of magnitude more energy than just using your search engine of choice to find one of the billions.
In general though, people tend to reach for convenience so I expect the energy usage to increase rather than decrease due to the demand created by the "ease" of AI.
On the other hand, if humanity because productive such that we "solve" energy than the point ends up being moot, since any energy increase will be offset by the invention of nuclear cold fusion, or other such things.
There are contrived examples one could conjure where using AI is slower but ultimately uses less energy than due to the complexity involved, since searching is well, searching preexisting things. If you need something new and you're not good at searching it might require less time and compute in the end to just ask a model than potentially wasting days, weeks or years flailing about.
Let's say that AI is a massive massive success and increases energy demand and ends up consuming 10% of electricity. I don't know if that's realistic, it's probable high, but let's assume it's correct within 10x.
Then, if that ramps up energy demand a lot, what happens? We deploy far more energy generation. Is that a problem? No, actually that will accelerate our energy transition to zero carbon. Already, nearly all of the new energy added in 2024 will be solar, batteries, and wind, because that's what's cheapest. Solar and battery will be 81% of the power capacity added (comparing different techs by power instead of energy has limitations, but assume a 20% capacity factor for solar to get to energy, 50% for gas, and 90% for nuclear):
https://www.eia.gov/todayinenergy/detail.php?id=61424
Further, since solar, batteries, and wind are all on technology learning curves, the more we build of them, the cheaper they become. So if we accelerate building and deployment, the faster they get cheaper.
That means, the faster we build more renewable tech, the quicker it gets to replacing carbon-based energy, on the merits of cost. Some natural gas plants, right now, cost more in fuel to run than it would cost to build a solar farm next to them, and use the solar when the sun is shining, and gas at night. The cheaper it gets, the sooner solar can economically replace more and more existing natural gas generation sites
Similarly for storage, right now it's an extra $90/MWh or so to store a MWh, and the sooner that drops in price, the more existing carbon sources can be replaced. Data centers that need their own backup power, and run on a combo of solar plus batteries, are a perfect application for this tech.
Meanwhile, existing energy infrastructure is loooong lived. Unlike computer tech, which depreciates quickly, you will have generations equipment in place for decades. So that means that the replacement rate for electricity generation is low, unless there is a big increase in demand.
Fortunately electrification of many industries is increasing electricity demand, and accelerating solar/battery/wind deployment and tech curves, but accelerating this with even more demand from AI would likely have positive effects rather than negative effects.
Without models, even me taking my car somewhere could be bad news. But what is the alternative? More polluting or less?
Could you explain the argument?
One made up possibility: maybe you can spend some energy on AI to optimize a supply chain, and that saves 50x as much energy every year after the supply chain is optimized.
The driving example was just that: an example. Driving obviously consumes energy, but what if, for example, you're a home energy efficiency contractor, and you're driving to a job to insulate a house, and that's going to lead to a net decrease in energy much greater than the energy you consumed to get there.
Poster then offers fragments of a parallel scenario (in which he drives his car) and in which models might also be used. Details of this scenario are not clear to me, but presumably there was some alternative? Unclear.
To look at the energy consumption of current model is extremely short sighted. If AI create a new material, a new solar cell, advance fusion reactor is all of humanity that jump forward.
Furthermore new generation of AI accelerators and new algorithms will improve efficiency by order of magnitute, it's still early days.
The material created will be a better poison/virus. The algorithm to keep the fusion tokamak from going boom will be at best 99% correct. The new solar cell? More exotic materials required than the current.
Blind optimism is a vice we cannot afford.
Reasonable is when you communicate the tradeoffs. More efficient, but more complex supply chain. Easier to immediately build, but we have to quadruple the computing power per cubic meter of reactor. More frequently discovered materials, but Uncle Ted can nail down the synthesis of the next VX, and a dispersal system in his garage.
Someone comes to me talking net gains. My next question is, "What's under the sheet you don't want to address out loud?" Under there is most often the real meat of the endeavor.
It's much easier to destroy and to enslave than to create something good. There I was thinking of the possible positive vs electricity cost.
Yes we might doom us all with AI. I hope the good win, wild time.
Putting aside that AI is terrible at novel discovery relative to its abilities otherwise, the direction of new technology is governed by incentives, not hypothetical good for humanity. Take social media as an example, which certainly had potential for good. How was it deployed? By which actors? To what end? And what were the unintended consequences? Not saying it’s all bad, but the utopia of the world becoming more connected and empathetic from the social-topological changes is clearly not the dominant outcome.
How do they even calculate it? I heard that bitcoin mining isn't profitable using fossil fuels. The average reward earned per kwh is something like 5 cents.
You multiply this number by the power consumption of the ASIC and divide it by average emissions for that amount of energy.
That being said, the issue with AI is its convenience: countless times I've seen coworkers not think by themselves for 2-3 seconds, but ask ChatGPT instead. Or worse, do it live in meetings when I had the answer from experience.
It's actually been a little eerie going back and watching TNG lately. Just a few years ago, the computer's conversational abilities were complete science fiction fantasy. Now it seems quaint.
But I worry that lack of education for younger devs (and other folk for that matter) about the dangers of over-delegating to AI is going to be a major cause of decline in analytical capacities.
> From 2010 to 2022, we signed more than 80 agreements totaling approximately 10 GW of clean energy generation capacity—the equivalent of more than 31 million solar panels. Now, as we enter our third decade of climate action, we’ve set a goal to run on 24/7 carbon-free energy on every grid where we operate by 2030, aiming to procure clean energy to meet our electricity needs, every hour of every day, within every grid where we operate. Achieving this will also increase the impact of our clean energy procurement on the decarbonization of the grids that serve us.
How long? How many degrees of global temperature growth long will it be? Will SV value the development of their personal god-head over the environmental destruction such development creates?
You've got a class that can build the AI systems AND build their own escape bunkers and soon their own space vehicles. What incentive exists to care about such issues? Any effeciences created won't drop the over-all wattage used by AI, it'll just increase the amount of AI used. The footprint will never shrink.
This is the entirely wrong question to ask. We will not be able to conserve our way to net zero. Billions of people in the developing world want to improve their life style. They want cars, air conditioning, and other conveniences that the developed world enjoys.
The energy demand will go up. We need to build massive amounts of zero carbon energy production, storage, and distribution facilities.
The question is not if humanity will find more uses for energy. We will. The question is will we be able to meet the demands without increasing CO2 and if we are realistic and plan on increased energy usage, we will.
And the problem, specially when we are talking about the future, it is not so much energy usage but fossil fuel based energy generation. AI may eventually, by the time it gets within an order or two below crypto usage, be fueled by clean energy. Or not, but it still is the elephant in the room vs the little mouse you try to scare about.
This is race-to-the-bottom thinking. All forms of civilization needs energy. More is better. Characterizing it as "obscene" smacks of anti-humanism and is completely wrong-headed.
We need to use more energy - a billion times more. We just need it to be clean.
If we have a future where governments decide what is good energy vs bad energy use instead of clean or dirty energy production, then we will really enter a deep new dark age of human civilization.
It's possible I used the word slightly wrong as the definitions I've now looked up often refer to the "middle class", though as I understand that could historically refer to people who we'd think of now as very upper class, just not actual nobles. Nevertheless, the above reading is what I intended
The major religion in the West for the past two thousand years teaches the concept of original sin and goes on about how we're inherently bad, in need of saving, wealth is bad, etc.
Now we have a lighter version of that with this new wave of secular anti-humanism. At least there is some slow, grinding progress.
Interestingly, Nietzsche predicted all of this in the Genealogy of Morality.
This is a self-correcting "problem".
https://e360.yale.edu/features/artificial-intelligence-clima...
If you’re an engineer like Carmack, I can imagine this is a problem that would be fruitful to tackle.
We are back at saving computation to get real world difference in performance and energy usage.
Having AI consume some energy is, in my opinion, very different than people speculating on bitcoin burning coal faster than most countries.
In the long term, GPUs won't be used to accelerate LLMs.
In light of this, articles with this sort of headline are just scaremongering nonsense. McCarthy didn't tolerate this intentional ignorance [2], so we shouldn't, either.
[1] https://www.osti.gov/biblio/1212086
[2] http://jmc.stanford.edu/commentary/progress/nuclear-faq.html
Whereas Ethereum has clearly shown that you can have a safe chain without this craziness.
But from reddit to Bitcoin talk to Bitcoin core development, there's a handful of people that control everything and want the status quo preserved.
https://www.nytimes.com/2023/03/08/opinion/noam-chomsky-chat...
of course there are others who disagree with him but they're wrong of course
With AI, on the other hand, efficiencies help. Of someone comes up with AI hardware that does inference or training 10x more efficiently, people will incorporate that and the energy usage for the same amount of work will decrease.
[0] https://journals.iucr.org/a/issues/2008/01/00/sc5015/ ("Busting out of crystallography's Sisyphean prison: from pencil and paper to structure solving at the press of a button[...]")
(Well, long order, actually. Short order once the technology ripened on its like fifth iteration)
1. Spend hours and $ hiring a graphic designer to make one
2. Spend hours learning/doing it myself
3. Spend time searching the internet looking for images that aren't copyrighted, or spend time finding an image I like and spend time and $ figuring out how to pay/license it
4. Spend a few seconds prompting AI
I know this comes off snarky but it's been decades now and we've got to be the change we want to see in the world. "business person at the office" banner image is completely useless.
But I've found a pretty interesting middle-ground/sweet spot (mainly with less technical audiences) where a nice graphic can really improve attention and retention, especially on a slide deck. For a recent example, I wrote about the importance of passwords for a "security awareness" training (the audience here are people in Pakistan who have only recently started using the internet). To help illustrate/reinforce the idea that if somebody gets your password they can "hack" you, I generated some colorful images of hackers, computer security, and even one with a person being arrested (in that section I was describing the law that could hold them accountable for breach of others personal information if their negligence allowed an adversary to gain access to the system). Personally to me it's fluff, but I ran a small (and not very scientific) A/B test with the first and second group and the version with the graphics was shockingly better at it's goal of raising awareness.
That being said, visual input and aesthetic are deeply intrinsic human things and expecting us to move beyond that is very unrealistic.
The best we can do is optimise the IG methodology and/or use open, perpetually royalty free images and clipart for illustration.
I’m not sure gen ai even helps with increasing productivity yet - since tons of existing tools and services do all the same things without having much of this ai (ai chatbot better than a web search? gimme a break…)
Just recently, I used generative AI integrated into databricks to build complex sql queries via English, used Google Gemini in Google sheet to formula a tricky pivot table, chatGPT build a cloud formation schema for new infra for a service & to navigate some tricky database primarily key setup for dynamodb; used copilot integrated into my IDE to speed up unit test and simple refractors and code complete every other line written.
I nor the other millions of subscribers to these tools are not subscribed because it barely works it's because it lets me work more efficiently.
It save tons of time across many product for many things that used to be sources of frustration or slow mundane tasks.
"Hype" is all it takes for that to be the case. Then every product had to "add AI" features to keep ahead of competitor's marketing.
- Call Customer Service, go through a crazy flow chart and be redirected to 10 different people.
- Go shopping, sift through thousands of reviews/images/fake products to find one that works.
- Manually create graphics/designs for unique rooms
- Request a junior employee to write some basic code.
- Request a junior employee to perform task X.
The AI also has the advantage that it can scale up to 24/7, on-demand work without an additional engineer.
(Then click a little more haha)
Many people are named McCarthy.
I disagree, because 1. I live close to a former Uranium Mining site
https://climate.mit.edu/ask-mit/what-are-health-and-environm...
2. Chernobyl
I live in the fallout area and mushrooms and boar meat is still contaminated 37 years later. Over 1000 km away from the plant.
Please google Thorium and Minor Actinide Incineration.
>Chernobyl
The manifest incompetencies of a literal Communist government are none of the US' concern.
I'll be happy to compromise with you, though. If you support nuclear in spite of Chernobyl, I'll support the death penalty for any capitalist that oversees a nuclear meltdown not caused by an ICBM despite your lack of knowledge about the history of the USSR, and nature of communism. I might even support burning at the stake or something if that's what it takes to make the obviously correct decision for human progress. McCarthy wrote that link decades after Three Mile and Chernobyl. He wouldn't have done so if he didn't also know what I'm relaying to you now. Received opinion is the least grounded kind.
EDIT: Lol I missed what thread this was in. Again, was Three Mile Island a Thorium reactor? We'd all save so much time if questions were directed to Google instead of me.
I am not anti nuclear btw. I just call out BS, when I see BS.
And I just shared what affected me personaly, if you want to know more about potential and actual disasters:
https://en.m.wikipedia.org/wiki/Lists_of_nuclear_disasters_a...
Multiply that with x, if you increase the number of plants.
And watch out for Murphy:
I understood nuclear to be, from most any lifecycle analysis, the modality with the fewest externalities of any grid-scale tech mature enough to field, and somehow almost cost-competitive even while regulated (as it is today) to safety thresholds many orders of magnitude more conservative than the incumbent technologies.
It sounds like you’re informed differently than me. What’d the better way that you propose in the last sentence?
[0] https://zionlights.substack.com/p/everything-i-believed-abou...
It remains to be seen if any can deliver. Early indications have all been extremely negative.
However, as for the massive logistics and management problem of actually constructing a nuclear reactor, it's unclear if having government employees would be better. It's clear that the current contractors are terrible, but that's no guarantee that the US gov would be better.
Also, basing energy supply considerations solely on a dollar cost notion is questionable, even if the dollar cost values were true. It's turning out that 'high nuclear energy cost' FUD is part and parcel of much of the 'green' movement and based on faulty LCOE assumptions that are spread uncritically by self-styled 'greens' on social media.
Also, this assumes dollar cost is the only relevant factor which is deeply unserious considering that we're dealing with a global long-term issue.
It's a bit strange to rally against cherry-picking and then referring to projects outside of relevant jurisdictions. For some reason we cannot replicate their results, we can have it as a goal perhaps but as a basis for a cost calculation I would opt for jurisdictions that are home or close to home (in a cultural sense).
> Also, this assumes dollar cost is the only relevant factor which is deeply unserious considering that we're dealing with a global long-term issue.
For investors it is the only relevant factor, which is what you initially replied to.
Ignore the greens, and definitely do not adopt their innumeracy, as so man nuclear advocates have done.
The real foes of nuclear right now are the economics and construction costs. The high costs are not "cherry picked," they are the data that are exhaustive and what everybody has available for decision making.
The people that need to be convinced to build nuclear are utility execs, not the public. There are a plentitude of sites that would be happy to add reactors to their existing pair, if somebody would come with the money.
That's the real fight here.
We are bad at construction, and productivity in the entire sector has been stagnant for half a century. In the meantime, we have seen soaring productivity in manufacturing, meaning that labor costs go up too, but without the corresponding productivity increase.
For a nuclear peoject, where so many of the very high up front costs are labor, this is devastating to the cost structure.
And of course, delays are devastating to costs when financing costs are more than 50% of the cost of the entire project.
Fix construction, and maybe nuclear will make financial sense. But not until construction is fixed.
Not as much "labor-saving" as bureucracy expanding by automating "bullshit jobs" jobs that shouldn't exist in the first place.
But does that usefulness really warrant the CO2 emissions?
You need to wonder why someone is trying to pull the heart strings of those concerned about the environment. Who would benefit from getting the greenies mobilized against AI? Who wants to shift the conversation to concerns around the consumption of AI and why?
For a concrete example, because my company helped build this, there’s power plants in the USA wasting less electricity because AI is better able to combine data to predict demand. This is megawatts per day that never needs to be generated because of software running on a couple graphics cards in a server rack using a few hundred watts.
At least AI is more useful than crypto currencies but if we mainly get more spam text and fake images the energy costs aren't worth it.
Here is a “radical” idea for you. Let markets decide what energy is useful for or not!
Obviously there can be rules and regulations around how clean energy is and pollution, but the absolutely last thing the world needs is governments regulating what energy use is “good” or “bad”, unless you are itching to live in serious dystopian future.
The bad news is lots of profitable things have bad consequences but markets don't care.
The entire function of markets is to discover usefulness!
The primary role of government is to provide a rule of law for markets.
Should there be pollution? No
Should there be snake oil or harmful consumer products? No
Should there be fraud and deception? No
Guess what when those happen those are GOVERNMENT failures! Not market failures. Any profits from those activities are not the reason they happened, the governments failure is.
In an economy that has a functioning government with law and order and regulations to protect property rights and public goods, but no intervention, then:
Profits == Usefulness
It’s absolutely outrageous how current western culture has failed to teach basic economic facts. As I have said numerous times, this is equivalent to believing in creationism and teaching it in schools as truth while having overwhelming evidence of evolution. That’s the situation we are in.
Collectivism, socialism, communism, social democracy, social justice, whatever name you assign to this idea a central planner who should adjust the system, intervene. they are all fairy tales that are being taught as facts while we actually have undeniably scientific and hard evidence of how human societies function economically and structurally, amongst honest people, honest intellectuals, there is no question about what works and why. Well regulated markets, NOT with interventions or redistribution, are the single most positive force in human history. Unfortunately culture and politics are holding us back in western civilization with a rival of the failed fairytales of previous times.
The reason why communism fails is the same why the market fails, because humans aren't perfect.
By your logic a crime is never to blame on the perpetrator only on the government.
Do you know who also shifts responsibility like that?
Serial killers! "I did it because you didn't stop me."
If you recognize the product you are selling hurts or kills people and your first response is, as long as it is legal I don't care, then the failure doesn't lie in the government, the failure is you.
BTW there is no market, there are only people and people are responsible for what they are doing
However in markets those transactions are voluntary. In central planning they are involuntary.
When the transaction harms the buyer the reason that happens is the seller has deceived them and is certainly the underlying guilty party, that I agree.
I agree perfect governments are also a fairy tale.
All that said you don’t seem to recognize the heroic role of incentives and profits and markets in finding usefulness. That’s what triggers me. When people voluntarily pay for something it’s because they find it useful.
What you appear to claim is that governments need to be the arbiter of usefulness since the people cannot be trusted. That is a seriously dystopian approach.
Since a lot of power is still dirty, high energy usage should be curbed.
On the flip side, restricting energy usage impacts homes significantly more than businesses and creates poverty.
Look at the UK for energy poverty examples.
https://www.theatlantic.com/technology/archive/2024/03/ai-wa...