Hi all,
Reordering the conversation a bit.
On 4/5/26 17:11, [email protected] wrote:
On 2026-04-05 02:46, Hugo Buddelmeijer via wrote:
>
>> Agreed. I do want to engage with you more at some point to find a
way forward, exactly because of you being passionate.
>
> Are you accusing your opponents of hysteria? Are we who disagree
> "tired and emotional"? Do we "just not understand"?
No, and I'm sorry I gave you that impression; I tried to write carefully
and clearly failed.
I would like to talk to Ian more because it seemed that Ian is someone I
could learn from, because Ian has a strong opinions that are different
from my own. 'Passionate' as in someone who cares.
Which is exactly what befits a humanist project, I think.
> Is this "good faith"?
Yes. And I assume good faith. I wouldn't bother otherwise.
What wording would have conveyed my intent to you better? (This is a
pretty meta question, but losing human connections is a risk I see of
LLMs, so I've decided to be more proactive in countering that.)
With that out of the way, some content.
> Yet you have not named even one benefit.
True. I feel like we are already shutting down the discussion before
any arguments are brought forwards.
I don't feel fully equipped to properly list the benefits of LLMs and
agentic coding, because I'm not up to date with the latest tech.
Nevertheless, I can give it a start, as I said I should.
Let's limit the scope a bit, to my favorite pastime: refreshing Guix
packages. Guix has over 30000 packages. Involving agents in
maintaining them would make that massively more manageable.
This small list is just scratching the surface, but these are things I'd
like to do already now:
- Agents help with learning and gaining understanding, for example an
unknown code base.
- Agents often (not always) produce much better code. Especially for
simple things, so we as humans can focus on the more complex / higher
level things.
- Agents can check human written code. They can catch many minor
problems, so the humans can focus on the big problems (assuming the LLM
doesn't catch those also).
- Agents are so much faster for simple things, like writing patches for
broken packages. This frees up human time for more complex things.
Note that weighing the cons against these benefits is essential, but
outside of the scope of this initial list.
Also note that there are many more benefits beyond this limited
use-case; this is just something I've personal experience with.
this block quote is not responsive to Ian's claim. What proposal is
being blocked here?
There was indeed not an explicit proposal here. So for clarity, and
specificity: I propose to use LLMs and Agents in updating Guix packages.
From earlier interactions, I believe Ian to be opposed to using LLMs
for Guix in all capacity, but I cannot find an explicit claim right now,
so maybe I'm mistaken.
I will name some costs:
Thank you.
LLM companies have repeatedly knocked over many other web services with
bad crawlers that disrespect robots.txt as widely reported on Mastodon
and elsewhere over the last 5 years;
Yes, that scraping is pretty bad and we should not contribute that. I
don't actually know whether the LLMs I use do this agressieve crawling;
I should check, thanks.
If people would use tools from companies that do crawl responsibly,
would that be OK?
several documented cases of suicidal ideation encouraged by LLM leading
to self-harm or murder, reported on regional news outlets and collected
by Caelan Conrad;
Yes, we don't want people to harm themselves or others. I don't think
it is a big risk for free software though.
Other companies also make tools that can be used to (accidentally) kill
people. Like cars; but I don't think we can ask free software people to
stop using cars when performing free software related activities.
I think this falls under personal objections. There are also people
being helped with their mental health through LLMs. I believe the good
will outweigh the bad here. But I don't have a car.
poisoning of discourse by ax-grinding cultists intent on appeasing
Roko's Basilisk and deploying their lie machines to Look Big, visible in
many venues
This is a complicated sentence; but I'll assume good faith.
We indeed don't need anyone "looking big". Keeping it to the Guix
packaging example; it would be easy to fix hundreds of broken Guix
packages with any of the agents and then boast about it. That would be
pointless indeed. E.g. doing that manually would require judgement,
like deciding whether the package should be removed instead.
About Roko's Basilisk, I had to look this up [it is an hypothetical AI
that will punish anyone that did not help in its creation], is that
actually something you fear might happen? If so, then how would not
using LLMs help?
Good news! There aren't sides. There is a project (actually a great many
projects) and there is a cult trying to disrupt the project for the sake
of a product from a company they worship.
Why would you say that?
Who is "trying to the disrupt the project for the sake of a product from
a company they worship"? Where do you find these people? The people I
meet are all trying to make the world a better place.
This statement is difficult to interpret in good faith.
On 4/5/26 04:36, Ian Eure wrote:
You are welcome to research the multitude of well-documented harms
done by LLMs yourself.
We need actual, concrete, objections from Guix contributors so we can
investigate and see whether we can resolve those. Not vague 'someone somewhere
wrote something about it'.
"Look something up" isn't vague. People have already brought up actual
objections, which you are ignoring.
OK. I'll do the work then.
> "emits carbon, both in staggering quantities"
I finally looked this up, because the people who make the claims didn't
give any numbers. Apparently the training of ChatGPT 4 resulted in
15,000 ton of CO2 [1]. My local airport (Schiphol) apparently causes
2,600,000 ton of CO2 per year [2]. An inference costs 0.06 Wh [1].
I'm honestly surprised by how low these numbers are. I was expecting
them to be much higher, given how worried people are. The energy costs
are trivial compared to the benefits. Or are these numbers wrong?
> "consumes fresh water"
About 10 mL per query [3]. Here I'm surprised how high this is!
However, a human drinks about 125ml per waking hour (2000 / 16), so back
to the 'fixing Guix packages' example, it might just be that an LLM is
more water-efficient than a human. (Although I propose to also keep the
human alive, so it adds up. But it is a comparison that helps me judge
how much water this actually is).
> "on infringing intellectual property"
Could be a problem, and I think an important reason to be a bit careful
for now.
But this is a very diverse subject, so it would be good if there are
some specific points to discuss.
E.g., back to the Guix packaging example, I doubt there is many
proprietary Scheme code around. And most Guix package definitions are
probably not copyrightable anyway (as they are facts).
> "including Free Software"
Yes. That's a good thing.
Turn it around: it would be more acceptable to use an LLM for Guix if it
was not trained on Free Software?
> "putting national economies at risk"
I wouldn't want to contribute to putting national economies at risk.
(Or well, given the black-red flag of your email provider, maybe we
should?) I need some pointers though, as I cannot find much concrete
information about it.
There is no need to speculate about whether "it" is well-documented.
Stop typing and go read the documents.
OK, I just did, and it doesn't seem to bad at all actually.
We should also differentiate between personal objections and what we as a
project want. We can only do that by actually writing out a list of objections
first.
Neither of these sentences is true. Even if the former were true, the
latter does not follow.
Yes we should differentiate. I have a whole list of philosophies that I
think everyone should follow. I don't enforce them through free
software projects.
> You want a list of objections only so that you
> can "overcome" them with more slop.
What slop? Again, trying to assume good faith: code is something LLMs
do very well, often better than people.
I can make up a list of potential issues that people might raise, but that
feels kinda the wrong way around. I'm planning to do so anyway at some point,
because I don't like this stale mate.
There is no stalemate. There are not sides. LLMs are anti-free and are
not software. They don't belong in Guix or in any distribution of
software that respects users.
Back to the Guix packages example: why would having hundreds of broken
packages be more respectful than fixing them with an LLM?
We should be able to discuss the use of LLMs in a civil way.
I have not even settled on a specific opinion. (Which is why I only
responded reactively, instead of proactively.)
There are (IMO) much more important and more interesting arguments
against LLMs that we have not come close to discussing. But I'll save
those for another day, as this is long enough.
Thanks,
Hugo
[1] https://www.sciencedirect.com/science/article/pii/S1364032125008329
"A systematic review of electricity demand for large language models"
[2]
https://www.duurzaamwaddinxveen.nl/hvh/8-s/164-sschiphol-is-de-een-na-grootste-vervuiler-in-het-veel-meer-co2-producerende-nederland
[3] https://arxiv.org/abs/2505.09598
"Benchmarking Energy, Water, and Carbon Footprint of LLM Inference"
