The novel utility of LLMs is that they bring natural language processing from an 80% solution to a 99% solution. It's extremely inefficient, a bit like if Knuth back in the 60s just threw up his hands and said "well any context-free grammar can be parsed with enough backtracking, CPU time is cheap, developer time is expensive", but it more or less works, with the caveat that the output is unfortunately also usually natural language and there is no guarantee when you will or won't hit the 1% where it doesn't work.
As a compression technique, it's way too lossy (and inefficient) to be particularly useful for that purpose, unless you consider the lossiness to be a feature since it matches the retardation of copyright with its own retardation of "hurr I added a green tint to the simpsons so it's not plagiarism" or "downscaling to 240p is fair use".
So for something like summarizing documentation, it's... okay, I guess. You have to bear in mind that it still adds that 1% error, though, and if that's not acceptable you'll have to verify it yourself anyway. For "generating", like with all lossy compression, it really depends on how many bits it spent on the particular pattern you're looking to plagiarize.
In short, it may be another tool in the toolbox for some use cases. If that were all, I'd shrug and say "maybe I'll play around with it later, I've got work to do right now and other tools to learn / build", but unfortunately the general public is involved, and so are megacorporations. The former is stupid and the latter is, shall we say, less than selfless, and not particularly concerned for either me or the general public. If I see an issue filed and it's just "I got this error message, here's what $LLM says" and the remaining 97% of the issue is just a poor summary of what I already know, that shows that the submitter either does not value my time or legitimately believes that he is helping. The former is annoying, the latter is incorrect, neither are helpful. If you wish to ask the computer madness machine what steps you should take to debug, go for it, but what I actually need to see is the data you collected. Do not hide it in a sea of generated sludge.
Unfortunately issues are not the only area where the general public interacts with people trying to actually get work done. This used to be the realm of newbies and amateurs, people who could be mentored into greater competence. But people who show up using LLMs for contributions are not only seldom competent, they seldom even wish to become competent; they have been deluded into believing that competence is not necessary. And the portion of the general public willing to believe corporate advertising and be a nuisance to stroke their ego is far larger than the portion actually interested in learning.
And social interactions have always been the bane of free software. Which means you can't even politely tell them to leave unless the hivemind all agrees.
In Guix specifically, copyright issues have to be a factor, and there's just no sane way of deciding that. On one extreme the software could become nonfree, and on the other extreme the license could become meaningless. It's impossible to say which extreme, or both, will be taken because copyright has always been made-up judicial bullshit with no basis in reality.
tl;dr: I am inclined to say "no LLMs anywhere" solely out of spite. If they end up being useful, then people who actually understand the code will be able to make contributions indistinguishable from those not using them anyway. It filters out the timewasters and provides plausible deniability for copyright issues.
