On Wed, Aug 12, 2026 at 05:33:57PM +0900, Simon Richter wrote:
> The other technical failure mode I can see is the exact opposite: it allows
> a non-ideal state to persist longer, by making workarounds more available.
> Basically, if we had had LLMs twenty years ago, there would have been little
> reason to develop debhelper, as we would have had a sufficiently low-effort
> mechanism to generate and modify the debian/rules "boilerplate" code.
I don't agree, and I think that the discussion so far misses this
crucial point. LLMs _can_ do this, but many LLM proponents only want to
use LLM assistance to do what humans already do. They would want to use
LLM assistance to _develop_ something like debhelper in your example,
instead of advocating for direct LLM maintainence of old-style
debhelper-free debian/rules.
This is an approach taken by many LLM users specifically to tackle the
many LLM output quality issues raised here. So banning LLM use
altogether is not necessary to deal with the quality issues. Guiding LLM
use to produce the result we want should be sufficient.
It bothers me that we have proposals (eg. Marc's proposal E) that
specifically consider this ("...satisfy the same standards
of...**maintainability**...") [emphasis mine], but those that disfavour
LLMs don't seem to see this distinction. Essentially proposals like E
are already making the compromise.
Admittedly, there are others who are LLM maximalists who think that
giving LLMs direct control instead of via reviewable code is the way
forward. I'm not convinced and notably we do not even have a GR proposal
that would permit this in Debian either.
This applies to your debhelper example. I believe that there is
consensus that reducing debian/rules to boilerplate by using debhelper
is more maintainable than the alternative. Even in the most
LLM-permissible GR proposal options we have at the moment, creating
debhelper would therefore have been preferred over individual LLM-based
maintainance of debian/rules.
> If we exclude any uses where the output does not pass expert review, not
> much is left, which makes it even more difficult to argue that the negative
> effects are somehow worth it...
You haven't considered the case where an expert uses LLM output as an
intermediate step to speed up their work. There are cases in which they
can correct the intermediate LLM output to produce a final output that
does pass expert review at a lower overall cost.
Robie
