AI-generated contributions and generated workload on maintainers

Jesse Rehmer jrehmer at blueworldhosting.com
Sun Nov 8 16:42:07 UTC 2026


On 11/8/26 4:53 AM, Julien ÉLIE wrote:
> Hi all,
>
> Wishing to share some thoughts with you.
>
> There is more and more AI-assisted development in today's world. I am 
> not against it; I do see the value it can give, that's not the problem.
>
> As far as INN is concerned, with very few active maintainers compared 
> to its large codebase, I am overwhelmed by the amount of work recent 
> PRs and issues take.  Besides, it does not seem temporary.
> There are several AI agents from OpenAI and Anthropic running on INN 
> codebase, submitting bug fixes and security reviews. Honestly, I 
> cannot cope with that.  Too much mental load, having to read very 
> lengthy discussions, review code, make some research, answer, commit, 
> merge, etc.
> I no longer take any fun in all of that, and my interest begins to 
> decrease.  I do not want to spend my spare time and my evenings 
> reading and discussing with an AI on PRs and the issue tracker. It 
> responds within minutes with questions, decisions to make, proposals, 
> etc.
> It even said yesterday "I'm not a C expert myself.  Your review 
> remains the real check."  Well, it knows C far better than I and does 
> not take any responsibility in its work...
> It announced tons of upcoming patches, concluding with "@Julien-Elie 
> be prepared 👯"

I have barely dipped my toes in the AI water, but my recommendation for 
you is find an agent/model you feel comfortable with and have it perform 
the bulk of reviews. Having multiple models perform reviews seems most 
beneficial in my limited experience.

> Well, this mail is a bit of a call for help.  I cannot work under 
> these conditions.  I naturally do not want to see INN development 
> stopped, notably when most (but not all) of the AI stuff is 
> interesting to merge.
>
> I say "most" because I spotted weird things and inaccuracies in some 
> submitted patches.  And a few minutes after highlighting these, the 
> responses came:
> "You are right, the reproduction steps are wrong: newsgroups: and 
> post: can't be combined."
> "my patch added the field name to its new error message but indeed 
> missed this one"
> "I mis-copied it"
> "I withdraw the sentence. I haven't substantiated it."
>
>
> Sigh...
> But at least it eventually finds the problems and to fix them itself 
> after some iterations.  Humans also do that, including me (I already 
> had to write fixes several times because of an oversight).

Been working with go-while and AI to modernize Diablo. After feeding it 
RFCs, Russ's site, providing a live feed on test machines, and 
instructing to compare behavior/functionality with INN and other NNTP 
tools, we're squashing a ton of bugs and adding major improvements. 
After 360+ commits by AI, the code is working in production (for me), 
and almost as feature complete as INN.

My initial impressions are throwing an agent at an entire codebase with 
a blanket purpose to find bugs in bulk results in a lot of noise. When 
working on a specific bug or improvement (narrower focus) the output 
quality is higher. As we've worked through bugs and feature 
implementation the models get better at understanding the code base in 
its entirety.

> So, what to do next?
>
> A proposal to discuss: couldn't we give push/commit access to the 
> GitHub users running AI agents?  This way, they could automatically 
> work and improve the code as well as open and close PRs and tickets.
>
> I would then just only prepare releases (if not also automated!) as I 
> do not see what else I can do these AI agents cannot already do.  They 
> have ideas and preferred solutions on all the subjects. They could 
> even answer questions on mailing-lists and newsgroups better than I 
> could! 

Their work needs still review, either human or another model directed by 
experienced humans. Mark Johnston from the FreeBSD team gave a talk this 
week on how they're dealing with the onslaught of AI generated bug and 
security reports. (Talk starts at about 1:38:00 - 
https://www.youtube.com/live/cnkmRe4nsOU?si=HeEEAtBkZQx0tEm8&t=5879)

I concur with Richard's sentiment in another thread: "It is absolutely 
possible to do better, with today's AIs, than "let Julian sort out the 
mess"."

Certainly my approach to troubleshooting and reporting issues will be 
drastically different. Prior to AI, posting to news.software.nntp, this 
list, and GitHub Issues are how I sought assistance. Much of the time my 
issues were misconfiguration or misinterpretation of documentation. Now 
I am able to identify issues caused by operator error without 
assistance, produce more intelligent and thorough bug reports, and even 
provide patches that fix the bug. Much better than sending a S.O.S. into 
the ether and hoping a kind and helpful soul would reply (though I still 
appreciate those).

Cheers,

Jesse



More information about the inn-workers mailing list