Re: LLM usage in Debian -- ballot options comparison

--R0k8QcJj3+H9nByw
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
Content-Transfer-Encoding: quoted-printable

Hi Johannes,

On 30/07/26 at 12:57 +0200, Johannes Schauer Marin Rodrigues wrote:
> I see you saw my "Kitten" example in the other mail. I used this example
> because I think (hope!) that we can agree that killing baby kittens is re=
ally,
> really bad and we should not be doing that. Going back to being more real=
istic:
> if I would now try to weigh costs versus benefits of use of genAI then on=
 the
> one hand we have for the costs a long list containing items like: all the=
 big
> and powerful models being owned by billion dollar corporations, big enoug=
h to
> sway politics, big enough to sway public opinion, that make the rich rich=
er
> (with a surprising number of rich people with clearly anti-democratic fas=
cist
> view points), the poor poorer (no everybody can be an artist!), support t=
he
> climate crisis with data centers requiring more electricity than nation s=
tates,
> use and pollute water, requiring supercomputers to train (not something w=
e do
> as part of a source package build), scraping our web because suddenly
> everything is free game (server admins hate this trick), making me want t=
o buy
> faster hardware because my browser has to solve anubis riddles on every w=
ebsite
> i use, oh no the hardware is now super expensive so you can only afford i=
t if
> you have enough cash around, the whole copyright dilemma, the whole
> hallucination issue, the issue with inherent sexism and racism (why are t=
hese
> points so low on my list?), the problem that if the LLM is wrong I cannot=
 "fix
> it", pushing back on the idea of "free software needs free tools", the de=
ep
> fakes, the problems for my students with learning because the LLM just ha=
nds
> them the answer (yes the smart ones will still be smart but many will jus=
t not
> put the effort anymore) and so on...  Oh and there is also the small thin=
g of
> my hobby being replaced by a machine but that issue really pales in contr=
ast to
> the long list of *real* problems.
>=20
> And on the list of benefits we have, very cynically speaking: it now was =
much
> easier to create a 17 kB HTML table.
>=20
> Now you can go and you can challenge any of my points above. And yes, you=
r HTML
> table will not make the oceans boil or democracy fail. But the things we =
do do
> not happen in a vacuum. The things we do influence others and they normal=
ize
> how all of us (even outside Debian) do things going forward. The free sof=
tware
> movement also started small and see where we are now. I would not like th=
is to
> become another thread where we discuss how local models are more ethical =
than
> claude/chatgpt (they are) or how things will be so much more efficient in=
 the
> future (they probably will be). The point I'm trying to make is this:
>=20
> Most of us are in a very, very privileged position. Many of us can work f=
or
> Debian because we either have secure jobs where we are paid to do this or
> because we live a sheltered and financially secure enough life to do this=
 in
> our free time. Compared to most of us on this planet, we can count us as
> *extremely* well off. And I understand how for probably all of us, creati=
ng
> software is something very important to us because otherwise we wouldn't =
put so
> much of our time into it. But at the end of the day: it's just software. =
 It's
> just an interesting sequence of zeroes and ones. Being able to produce th=
is
> sequence faster is a first-world-problem of such a high degree of privile=
ge
> that it pales against most of the issues I have with LLMs even if all of =
the
> issues in my list above would only be half as bad as I think they are.
>=20
> For myself this means that I rather "limit myself" and write software "sl=
owly"
> or that I don't do some things at all. That way, I'm not part of the prob=
lems I
> see and instead am part of the world that I want to live in.
>=20
> Sorry Lucas that this long (and emotional) mail was now a direct reply to
> yours. I *highly* respect your work and your technical expertise. Just to
> affirm, I do not think that you acted in bad faith and I don't think that=
 your
> messages were written badly. It is useful for me to read your messages.

Thanks again for this detailed write-up.

I think it would be useful to explain how I see AI, and why I'm in the
pro-AI camp in this discussion despite acknowledging all the costs that
you list.

I view the spread of AI we've seen over the past two years as a
revolution similar in impact to the generalization of Internet access to
the wider public since the mid 90s. Several parallels can be drawn:

1/ There were a few large providers trying to sell a locked-in version
of their Internet (AOL, Infonie in France, apparently T-Online in
Germany), where real web access was only secondary. At the same time,
there were initiatives (some very bottom-up) to empower Internet users
to become actors by publishing their own web pages (GeoCities; Mygale or
Altern in France). Later, there were also initiatives to build community
ISPs, to retake ownership of Internet access.  More recently, there have
also been initiatives to encourage self-hosting services and counter the
SaaS & Cloud worlds.

Of course, the landscape of AI corporations is frightening. But there's
hope to build alternatives, with open-source agents and open-weight or
maybe even open-source models.

2/ Both the Internet and AI dramatically change our access to knowledge.
Before the Internet, one had to use super-expensive paper encyclopedias,
and there was a real divide between those who had access, and those who
didn't. Similarly, AI fundamentally changes our ability to acquire
knowledge. Sure, it's not perfect yet. But do you remember the articles
claiming that Wikipedia would never match the paper encyclopedias? Also,
both challenge teaching and evaluation practices (like calculators
before), but that's a pedagogy problem, not a technology problem.

3/ Both seem to involve a financial bubble. My bet would be that the AI
bubble will be similar to the Internet bubble: it might burst at some
point and kill some companies, but all the solid work that is happening
around AI will continue.

4/ Both have a huge impact on society at large, such as jobs getting
transformed (for better or for worse), suppressed, or created. This is
not just a new technology for our field; this is a technology from our
field that has major consequences for society.

On your point about privilege: you are right that, today, benefiting
=66rom AI is a first-world concern. But the same was true of Internet
access in the 90s, and it was democratized over time, and in some ways
helped narrow the divide between the first world and the rest. That
should be the goal for AI too.

Also, it's worth noting that it might be a bit early to weigh costs and
benefits of AI. If you look at the Internet again, the technology's
"costs" include facilitating [online] harassment, causing or increasing
various addictions, causing the control/theft of insane amounts of
personal data, facilitating various illegal networks such as paedophiles
or terrorists, facilitating misinformation/disinformation, mass
surveillance, killing many shops due to online deliveries, concentration
of power in big-tech monopolies, massively increasing the digital divide
between those privileged with access and the knowledge to use and the
rest, environmental cost (datacenters, devices). If we had known all of
this 30 years ago, wouldn't it have sounded like a good idea to ban the
Internet? One difference with AI is that we seem to have a better
understanding of its costs. But that means we are in a better position
to work on them and shape the technology early, not that we should ban
it.

So, despite all the costs you listed, I think that the Free Software
community should explore how to build freer, more independent and better
AI, even if not everything will be super-free and perfect from day one
(after all the Free Software movement wasn't). And in the context of
Debian, this means accepting its use so that Debian contributors who
want to experiment can do so and contribute to the fight for better AI;
this also means understanding where it can help us, what it can bring
us, where it fails to deliver, and what are its downsides in our
specific context.

Lucas

--R0k8QcJj3+H9nByw
Content-Type: application/pgp-signature; name="signature.asc"

-----BEGIN PGP SIGNATURE-----
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=d5BY
-----END PGP SIGNATURE-----

--R0k8QcJj3+H9nByw--