Re: LLM usage in Debian -- ballot options comparison

On 04/08/26 at 15:06 +0200, Didier 'OdyX' Raboud wrote:
> Le lundi, 3 août 2026, 14.49:22 h heure d’été d’Europe centrale Lucas Nussbaum 
> a écrit :
> > Personal anecdote:
> > I was looking for an application to help a child with learning
> > multiplication tables (clearly a boring thing to learn that children
> > go through between ages 6 and 8). There are plenty of such apps on
> > Android and Apple's app stores, and also educational games such as
> > GCompris (free software and packaged in Debian), but none of them really
> > matched my needs. I wanted something simple and focused on
> > multiplication tables, with the ability to select which tables to learn,
> > define whether I wanted questions like 4x5=? or 4x?=20, etc. Also I
> > wanted different game modes (streak mode, timer).
> > 
> > I ended up vibe-coding such an app, together with the target user.

For context, I was replying to a subthread about vibe-coding, where
Aigars shared an anecdote about an non-programmer's use of vibe-coding.
I was not implying that vibe-coding makes sense in the context of
software created for Debian, or that vibe-coded apps should be packaged
in Debian.

> One way to look at this is that you used a 2026 proprietary Software-as-a-
> Service to produce a piece of software for yourself (and relatives). Said SaaS 
> comes with enormous externalities, but happens to be currently cheap-enough 
> for your level of revenue.

So I don't think I tracked which model and which provider I used to
generate that app, but the same result can be achieved using an open
source local agent such as OpenCode, and an Open Weight model that can
be self-hosted. Of course, the required hardware to self-host might not
be cheap, but that's something that can be amortized over different
users.

> One important point for me here is that I grew convinced that the Free 
> Software movement is much more about the _movement_ that about the software. 
> In other words: I care about the Debian project primarily because it's a 
> project of commons (we build something collectively, and it benefits all), and 
> only secondarily because it happens to be good software. (I'm well aware that 
> not everyone in the project sorts these two arguments in the same order, 
> that's likely why we have a vote).
> 
> To get back to your anecdote above: you built useful software, but you didn't 
> build a community. Again, no offense meant: not all software needs a 
> community, pushing to Github doesn't imply collaboration, etc. But I find that 
> it's a too common trait of LLM-produced code: it's only good code.

I shared the same anecdote in another discussion, and some people
pointed out that it would make sense to share the app in communities of
teachers, such as https://ressourcerie.forge.apps.education.fr/ or
https://primaire.forge.apps.education.fr/, explicitely pointing out that
most of the applications there were vibe-coded. So it looks like you can
build a community around the need (the specification) even if the code
itself is generated.

> I want a Debian project to which it's _easier_ to participate to, not harder. 
> And I'm afraid that allowing LLM usage will deepen a divide between "those who 
> can afford the price of LLMs" and "those who are stuck with just their aging 
> laptop at hand" (caricature intended).

As I mentioned already in this thread, I've been working on trying to
get inference providers sponsor access to Debian developers (aligned
with https://lists.debian.org/debian-devel/2025/06/msg00060.html), and I
recently got some success. So it looks like we should be able to provide
access to state-of-the-art Open Weight models to Debian contributors.

Lucas