Hey folks!
There’s growing interest for This Week in Fedora, it’s been discussed in the Fedora Podcast and they mentionned it’s hard to discover. I wonder if it would be useful to put it under a fp.o domain, such as thisweek.fedoraproject.org (to mimic https://thisweek.gnome.org/).
The articles are in the repo, so I can easily make an openshift app that builds the HTML from the repo and serves it, the technical part is not a problem. I would still be generating the contents on my own machine because of the token required and the manual review that I do.
My question is more: do we want to make it official, knowing that the content is generated with a closed-source LLM (currently Gemini)?
6 Likes
As it already replaced Community weekly reports I’m for making it more visible.
2 Likes
You note that summaries are AI generated, and you spend time hand-editing them—that’s good enough for me.
Thanks for these, they’re extremely useful, and the only way I manage to keep up with everything that’s happening in Fedora
(Unfortunately newsboat seems to get stuck at anubis, so I can’t read get to the RSS feed there easily, but that’s a different issue :))
+1 for making it official!
1 Like
It is great that you are aware of the problems here.
I am -1 to continuing down the slippery slope of incorporating closed source software and AI into The Fedora Project.
I certainly appreciate the manual review you do, but I do not think that you review the full text meeting logs and then compare them with the Google output for every meeting.
2 Likes
Understood, thanks! If the model used was an open-weight model (such as Llama2), would it be acceptable for you? For the record I personnaly consider open-weight to be closed source (similar to freeware), my goal was to use Olmo from AI2 because they provide the training data but until we have hardware to run it I haven’t found a way to finance hosted use of the model.
Indeed I do not, I check the text and the links to verify it’s plausable. Sometimes I check the data sources because I found that most LLM errors came from insufficient source data, but not all of them. I don’t really have a strict process besides reading it, looking at the link previews, and clicking on some of them. Besides that it’s more of a “feeling” thing, I investigate when something looks suspicious to me.
I’d probibly be fine with making it have it’s own hostname…
If for some reason we didn’t want to fully associate it with fedoraproject.org (for ai model reasons or whatever), we could also use some other domain we control. But that might not make it that more discoverable I guess.
1 Like
It is great that you are aware of the problems here.
I am -1 to continuing down the slippery slope of incorporating closed source software and AI into The Fedora Project.I certainly appreciate the manual review you do, but I do not think that you review the full text meeting logs and then compare them with the Google output for every meeting.
I also really appreciate the work you’ve put into this! But I have
similar qualms to MatH. I would be more amenable if there was a very
prominent disclaimer that the content is 100% AI-generated and could
contain errors, but then that begs the question if something error-prone
and unpredictable (due to it being generated by an LLM) is something we
want to make more official in the first place. For some things, having a
probably-mostly-accurate AI summary is maybe okay, but I am more worried
about having AI summaries for project governance discussions and
decisions in FESCo and Council and other bodies, since those bodies are
not officially reviewing, verifying, or endorsing the summaries.
From abompard down-thread:
If the model used was an open-weight model (such as Llama2), would it be acceptable for you? For the record I personnaly consider open-weight to be closed source (similar to freeware), my goal was to use Olmo from AI2 because they provide the training data but until we have hardware to run it I haven’t found a way to finance hosted use of the model.
I think that addresses the closed source problem but not the accuracy
issue (if anything it could make it worse if the open source and open
weight models are sometimes of lower quality?). In any case, maybe the
AI/ML SIG would be interested in collaborating to get those resources
set up.
1 Like