On Fri, May 15, 2026 at 11:20 AM Ludovic Courtès <[email protected]> wrote:
>
> Hello Guix!
>
> I am submitting GCD 008: “Standing up for human crafting” (draft
> attached).

My comments are based on reading the most recent revision, 6f3990db8e.

Large Language Models are clearly more than stochastic text
generators. One recent application was solving the Erdős Unit Distance
Problem [0], a problem that had challenged mathematicians for 80
years. To achieve consensus a GCD motivation should either be fairly
balanced or narrowly focused on the project. I would prefer the
latter. 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. If the
motivation is to reduce our environmental cost then an accounting of
our current situation is required before consideration of mitigating
solutions.

GenAI may break free software but it need not break apart the free
software community. The purpose of free software was never
exclusivity, which is why permissive licenses and non-copyright works
are compatible with GNU. If we followed our principles this would not
be in question but it is human nature to want to compel the speech and
actions of others. It is “free” software, free as in speech.

For Guix there is a deeper question as to the direction of the project:

On Wed, May 20, 2026 at 3:33 AM Konrad Hinsen
<[email protected]> wrote:
[...]
> My take on this is likely to be unpopular here: I think that Guix is
> already too big and diverse in its current state. I'd prefer Guix to
> concentrate on a minimal core with just the packages necessary for a
> minimal viable system. For such a core, it would be much easier to be
> strict about rejecting vibe-coded software. Everything else could be in
> separate channels, and each of these channels could have its own genAI
> policies.

How minimal could this minimal core with just the packages necessary
for a minimal viable system be? The bootstrap to Guix, Guix System,
and the installer? Guix can shrink or it can continue to grow, why not
both?

The second gate (after licensing) for package inclusion into Guix is
not quality, it’s user demand. So a single-user “vibe-coded” package
likely isn’t appropriate for inclusion today irrespective of the code
quality.

What if we could have this single “core” channel while allowing users
to select their main channel? Have one main channel dependent as today
purely on manual submissions, no GenAI contributions or packages, with
out-of-date and vulnerable libraries and core packages. Have another
main channel using GenAI to update packages. Have another main channel
that includes all of the free software in the world.

> 5. We acknowledge that the project’s sustainability depends on automation for 
> all the mechanical, labor-intensive tasks such as package updates.  We will 
> keep **improving hackable tools and services to automate some of the package 
> collection maintenance work**.

Tools are important and used by agents as well as humans. This
proposed automation eventually leads to development of artificial
intelligence. The project should be recording the knowledge required
to build and update packages, knowledge that must be collected,
interpreted, and kept up-to-date. The project never found consensus in
doing this the human way.

The discussion over licensing is highly reminiscent of Microsoft’s FUD
campaign over the “virality” of the GPL. Copyright is not required for
free software. Many have claimed “uncertainty” (the ‘U’ in FUD) but we
have yet to receive an explanation or description of the uncertainty
regarding inclusion of LLM output in free software.

[0] https://mathworld.wolfram.com/ErdosUnitDistanceProblem.html

Reply via email to