Surprised not to see a whole chapter on the environment impact. It's quite a big talking point around here (Europe, France) to discredit AI usage, along with the usual ethics issues about art theft, job destruction, making it easier to generate disinformation and working conditions of AI trainers in low-income countries.
(Disclaimer: I am not an anti-AI guy — I am just listing the common talking points I see in my feeds.)
Yeah, it would be really useful to see a high quality report like this that addresses that issue.
My strong intuition at the moment is that the environmental impact is greatly exaggerated.
The energy cost of executing prompts has dropped enormously over the past two years - something that's reflected in this report when it says "Driven by increasingly capable small models, the inference cost for a system performing at the level of GPT-3.5 dropped over 280-fold between November 2022 and October 2024". I wrote a bit about that here: https://simonwillison.net/2024/Dec/31/llms-in-2024/#the-envi...
We still don't have great numbers on training costs for most of the larger labs, which are likely extremely high.
Llama 3.3 70B cost "39.3M GPU hours of computation on H100-80GB (TDP of 700W) type hardware" which they calculated as 11,390 tons CO2eq. I tried to compare that to fully loaded passenger jet flights between London and New York and got a number of between 28 and 56 flights, but I then completely lost confidence in my ability to credibly run those calculations because I don't understand nearly enough about how CO2eq is calculated in different industries.
The "LLMs are an environmental catastrophe" messaging has become so firmly ingrained in our culture that I think it would benefit the AI labs themselves enormously if they were more transparent about the actual numbers.
> Global AI data center power demand could reach 68 GW by 2027 and 327 GW by 2030, compared with total global data center capacity of just 88 GW in 2022.
"AI's Power Requirements Under Exponential Growth", Jan 28, 2025:
To assess the env impact, I think we need to look a bit further:
While the single query might have become more efficient, we would also have to relate this to the increased volume of overall queries. E.g in the last few years, how many more users, and queries per user were requested.
The training costs are amortized over inference. More lifetime queries means better efficiency.
Individual inferences are extremely low impact. Additionally it will be almost impossible to assess the net effect due to the complexity of the downstream interactions.
At 40M 700W GPU hours 160 million queries gets you 175Wh per query. That's less than the energy required to boil a pot of pasta. This is merely an upper bound - it's near certain that many times more queries will be run over the life of the model.
> ... I then completely lost confidence in my ability to credibly run those calculations because I don't understand nearly enough about how CO2eq is calculated in different industries.
There is a lot of heated debate on the "correct" methodology for calculating CO2e in different industries. I calculate it in my job and I have to update the formulas and variables very often. Don't beat yourself over it. :)
If I were an AI advocate I'd push the environmental angle to distract from IP and other (IMO bigger and immediate concerns) like DOGE using AI to audit government agencies and messages, or AI generated discourse driving every modern social platform.
I think the biggest mistake liberals make (I am one) is that they expect disinformation to come against their beliefs when the most power disinformation comes bundled with their beliefs in the form of misdirection, exaggeration, or other subterfuge.
The biggest mistake liberals have made is thinking leaving the markets to their own devices wouldn't lead to an accumulation of wealth so egregious that the nation collapses into fascism as the wealthy use their power to dismantle the rule of law.
The topic as a whole isn't overlooked but I think the societal impact is understated even by Hollywood. When every security camera is networked and has a mind of its own things get really weird and that's before we consider the likes of Boston Dynamics.
A robotic police officer on every corner isn't at all far fetched at that point.
Every time I have seen it mentioned, it has been rolled into data center usage.
Is there any separate analysis on AI resource usage?
For a few years now it has been frequently reported that building and running renewable energy is cheaper than running fossil fuel electricity generation.
I know some fossil fuel plants run to earn the subsidies that incentivised their construction. Is the main driver for fossil fuel electricity generation now mainly bureaucratic? If not why is it persisting? Were we misinformed as to the capability of renewables?
There's a couple of things at play here (renewable energy is my industry).
1. Renewable energy, especially solar, is cheaper *sometimes*. How much sunlight is there in that area? The difference between New Mexico and Illinois for example is almost a factor of 2. That is a massive factor. Other key factors include cost of labor, and (often underestimated) beautacratic red tape. For example, in India it takes about 6 weeks to go from "I'll spend $70 million on a solar farm" to having a fully functional 10 MW solar farm. In the US, you'll need something like 30% more money, and it'll take 9-18 months. In some parts of Europe, it might take 4-5 years and cost double to triple.
All of those things matter a lot.
2. For the most part, capex is the dominant factor in the cost of energy. In the case of fossil fuels, we've already spent the capex, so while it's more expensive over a period of 20 years to keep using coal, if you are just trying to make the budget crunch for 2025 and 2026 it might make sense to stay on fossil fuels even if renewable energy is technically "cheaper".
3. Energy is just a hard problem to solve. Grid integrations, regulatory permission, regulatory capture, monopolies, base load versus peak power, duck curves, etc etc. If you have something that's working (fossil fuels), it might be difficult to justify switching to something that you don't know how it will work.
Solar is becoming dominant very quickly. Give it a little bit of time, and you'll see more and more people switching to solar over fossil fuels.
I guess for things like training AI, they can go where the power is generated which would favour dropping them right next to a solar farm located for the best output.
Despite their name I imagine the transportation costs of weights would be quite low.
Thank you for your reply by the way, I like being able to ask why something is so rather than adding another uninformed opinion to the thread.
whats the lifetime environmental impact of hiring one decent human being who is capable enough assist with work. Well a lot, you gotta do 25 years with 30 kids to get one useful person.
You get to upgrade them, kill them off, have them on demand
I saw a fun comparison a while back (which I now cannot find) of the amount of CO2 it takes to train a leading LLM compared to the amount of CO2 it takes to fly every attendee of the NeurIPS AI conference (13,000+ people) to and from the event.
Bluesky. They use Expo on top of React Native, use React Native for Web (with a desktop and mobile), and for mobile native apps.
Let's note that because the clients are fully open-source and on GitHub, people from Expo and React Native are helping the little team behind the clients improve performance over time: it's not their final form!
"It is time for Postgres to care about customers!"
Wow. It's one thing to create value on top of one of the most respected open source projects in a field, it's quite another to bash it as the opening sentence of your blog post.
Could you have some serious sources about Google dropping support for Flutter and Dart?
I mean, words have meanings. It's not because Google laid off part of the teams that worked on those technologies that it means that they dropped support of them.
"Controller and processor
Data controllers must clearly disclose any data collection, declare the lawful basis and purpose for data processing, and state how long data is being retained and if it is being shared with any third parties or outside of the EEA."
I don't see the word "personal" in this sentence, only "any data collection". It's clearly not in the users' interest/benefit to activate this data collection, and not required for the normal functioning of the browser/websites. So activating this silently is minimum _very unethical_ and probably illegal, but I'm not a lawyer.
A rather disappointing read. I was expecting an analysis explaining why there is this trend towards very sparse interfaces, or practical ways of designing interfaces that are denser in the face of design trends that are pushing all product teams to do ever more spacing out.
Instead, what I found was a reminder of the ‘laws of design’, which are certainly interesting, but which are only tangentially linked to this drift (in my opinion); and to take the most extreme example of sparse interfaces (the Bloomberg Terminal), without really any concrete elements that could help bring a little density back to our user interfaces.
...not to mention what ends the article, a lunar explanation along the lines of ‘Google's very high stock market valuation compared to Yahoo can be explained by the lack of density of its home page interface’ - really? Come on.
Agreed. Seems like a long winded lead up to what reads to me like a mildly condescending Gestalt 101, followed with the same examples I've seen in countless other blog posts over the last 15 years and very little in terms of actually discussing design trends.
Well, Google lost against Oracle too, so it appears a mere API specification can be closed down arbitrarily; than is the world we live in. Unless the US gets a lot more tech literate and open minded judges and officials, I doubt that will change for the better. And, looking at their presidential candidates… well.
The court decided the opposite--that APIs are copyrightable. However, the Supreme Court ruled that Google's usage was fair use, so I would agree that Google mostly won. The Supreme Court didn't consider whether APIs are copyrightable (the lower court ruled that) because Google would win regardless because it was fair use.
So I'm not sure it matters much whether APIs are copyrightable when what Google did was ruled fair use. I'd prefer if the courts ruled APIs weren't copyrightable, but I think it was still a good result because doing what Google did probably covers about any use case anyway.
I wonder if making a device which uses Threads could be considered fair use in the same way, because implementing threads is required for interoperability with many devices.
The Federal Court took up the appeal from Alsup case, accepted Oracles arguments that copying headers & using then same variable makes made the Java reimplementation a copyright violation (incompetent losers), sent the question of fair use back to a jury trial, the jury decided yeah it was fair use, the Federal Court ignored the jury and decided to ignore everyone hollering at them that they were being idiots & ruled for Oracle anyways.
Then the Supreme Court ignored the copyrightability aspects & ruled for a Google on some fair use grounds.
I've skimmed the write up from the ever excellent always recommendable Mark Lemley, Interfaces and Interoperability After Google v. Oracle, and really hope I can go a bit deeper into the history & trial at some point. Section 2 The Long Saga of Google v. Oracle starts on page 27 of the inner pdf. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3898154
It's been incredibly disappointing watching courts like the Federal Circuit be so unable to handle even basic technical matters with even an iota of comprehension, having them bungle up things so badly in the face of so much easy to rely on precedent. Being sweet talked by Oracle's lawyers into believing a header file is anything greater than interface definiton is either incompetence, or some really vicious pro-business hellworld shit.
Governments have no problem allowing de facto specifications to be closed. 5G is heavily patent protected, for example. MPEG is heavily patent protected.
Typically patents "essential" for a standard are licensed on "fair, reasonable and non-discriminatory" (FRAND) terms. But you do still have to go and pay for the license (sometimes from all the individual companies that have patents, sometimes from a consortium that represents the entire patent pool for a standard).
Here in Paris (France), I always use my credit† card, for everything from a drink in a bar to groceries to burgers to whatever. Most of the time, I use the contactless feature of my plastic card (works for anything under 50€), and sometimes, using my phone using Google Play, which is linked to my card, but doesn't have the 50€ ceiling for regulatory reasons.
I used to carry some cash for my nearest automatic laundery machines, but even that one now features a contactless payment dongle, so it's nearly useless for me.
† I say "credit card", but most plastic card WE carry on France are actually debit card, i.e. there is no specific credit card balance to top or anything, it's directly withdrawned from your account. (There is something hybrid called "differed debit card", that mimic some of the credit card experience, but it's basically buffering your withdrawals, without any "true" credit happening).
(Disclaimer: I am not an anti-AI guy — I am just listing the common talking points I see in my feeds.)