On Mon, Jun 1, 2026 at 7:35 PM Ludovic Courtès <[email protected]> wrote: > > Hi Greg and all, > > Greg Hogan <[email protected]> skribis: > > >> > One > >> > expressed concern for generative AI is over datacenter energy and > >> > water use. The Guix project currently employs multiple large compute > >> > clusters running tens of thousands of builds. > >> > >> Are you comparing ~50 GPU-less nodes worldwide (that I know of, but > >> let’s imagine there are big secret Guix build farms and the figure is > >> 1,000 nodes) to Colossus¹ & co.?… > > [...] > > > The Colossus-scale models have billions of global users. Guix has > > perhaps a few thousand? The analysis would be a speculative comparison > > of the per-capita energy use of the Guix project with and without the > > use of LLMs. > > The carbon footprint of Guix build farms (meaning the greenhouse gas > emissions to produce the hardware of those build farms and to run them) > is proportional not to the number of users but to the number of > packages. > > If Guix had a billion users, with a comparable number of packages, the > build farm size would comparable to what we have today. We would need > to use a content delivery network (CDN), for example, which adds to the > carbon budget of the infrastructure, but relatively little. > > The carbon footprint of facebook.com, with its billion users, is > certainly very high, but its infra is a toy compared to what companies > like OpenAI, xAI, and Anthropic are building. The popularity rise of > Facebook or even YouTube etc. did not lead to a shortage of RAM and GPUs > on the market contrary to what we’re experiencing today. > > You may say: “with AI, I might do in 1 day what I would otherwise do in > 10 days, proof that it’s more efficient and thus reduces my carbon > footprint”. That’s not the case due to Jevon’s paradox, as Simon nicely > explained: > <https://codeberg.org/guix/guix-consensus-documents/pulls/13#issuecomment-15843803>. > > In the 2015 Paris Agreement¹, countries pledged to act to keep global > surface temperature increase below 1.5℃ compared to pre-industrial era. > To achieve that we need to significantly reduce GHG emissions (for > Western Europe where I live, we’re talking about going from ~10 TCO₂e to > 2 TCO₂e per capita—this is huge). “AI” takes us in the opposite > direction, and fast.
As you note, "per capita", which is why the number of users in the denominator is critical. If Guix had a billion users like many American and Chinese tech companies then we could justify the ecological impact of the project's compute clusters and local developer consumption. As to Jevons paradox, this is an observation rather than a general scientific law. And Jevons specific observation summarized by Wikipedia was wonderfully wrong. Coal use has plummeted due to technological progress. It is only because we stopped technological progress for non-technological reasons that any coal, oil, or gas is used today. "In 1865, the English economist William Stanley Jevons observed that technological improvements that increased the efficiency of coal use led to the increased consumption of coal in a wide range of industries. He argued that, contrary to common intuition, technological progress could not be relied upon to reduce fuel consumption." [https://en.wikipedia.org/wiki/Jevons_paradox] > *** Why are we discussing all this in the context of Guix? *** > > The point I am trying to make in GCD 008 on this topic is that free > software doesn’t exist in a vacuum. Yes, LLMs (+ agents) can automate > away tedious packaging tasks. But we can acknowledge that the > environmental cost of these tools *alone* is a problem for society, and > we can refuse *collectively* to not make ourselves dependent on it. If the motivation is to reduce the ecological impact of the Guix project then we should collaborate on that goal. Otherwise, this looks like a pretext. > After all, we’ve been running this project without LLMs for 14 years, so > it shouldn’t be too hard, right? Yes, we have work to do to improve > package maintenance automation; luckily, we already have solid > foundations for that—more solid than most other distros. This argument equally justifies eliminating the Guix project. Just as with Guix, LLMs provide novel capabilities. > I emphasized “collectively” above: the GCD has never been about policing > people—as individuals, we have to make do and sometimes compromise on > what we think is “right”, from running non-free software to using LLMs; > that’s okay and explicitly acknowledged in the document. But this is policing the process. Else we would not be focusing on user tools but rather outcomes including licensing of contributions, package quality, and ecological impact, none of which are concerns specific to LLMs. > I hope I sufficiently clarified and contextualized what I had in mind > regarding this point. > > > *** Is the GCD about “ecology”? *** > > No it’s not. The environmental cost of “AI” is an important point IMO > but there are several more. > > The other points delve into social, ethical, and legal issues. Legal > issues have been widely discussed already, but the rest much less. The last I have heard on legal issues are references to discussions on a private GNU mailing list. > I would invite people to take a look at these other points and the > references in the document. Summarizing it as “AI is bad” would not > only be dismissive, it would miss out on issues that affect free > software projects. > > Free software is a social endeavor with an ethical compass. Just like > some were famously critical of the non-free software running in their > printer or pacemaker and what it means for their agency, questioning how > these technologies affect us and society seems totally legit. I don't perceive this effect on agency. LLMs provide far more agency than the machines we are communicating with. The output of an LLM is still human readable and reviewable. We have agency to use LLMs to benefit the project and community, or to harm. > *** “My motivation is different!” *** > > That’s okay! The pledge in the document derives from my motivation. > Perhaps with a different motivation you reach similar conclusions and we > can come up with a pledge that addresses both our concerns. > > The “Motivation” section gives context, but what matters in the end is > the pledge. I agree with the pledge but disagree with some wording of the commitments. Banning LLMs is not required to support the pledge, and package quality and licensing are general topics not specific to LLM-assisted code.
