Well, ok, let's have this discussion. Last night I submitted two FEPs about federating location for review, after working on them for several days, trying to make both the FEP proposals and our implementation as good as possible for other people to build on. Today I was told that FEP proposals written with the help of LLMs "will likely be rejected in the future". I honestly find it hard to understand how a project that has a stated goal to "improve interoperability and well-being of diverse services, applications and communities that form the Fediverse" could exclude projects like ours. We have tried to align with other FEPs, and were planning to continue this effort in the coming months, because we believe that fedi can only benefit from standardization and interoperability and it's in our main concerns and project goals, but this response stunned me. If FEPs exclude projects working with LLMs, it's not as if we'll stop using them - we'll just stop bothering with FEPs. Is this for the benefit of fedi?
#activitypub #fediverse #fep
@liaizon @panos It's not really about me or other facilitators - we don't fully read every submitted FEP. We check metadata and skim the text, but a thorough review is a task for the community.
And now we're getting feedback that people are tired of AI generated FEPs. Here's the most recent discussion: https://codeberg.org/fediverse/fep/issues/972
@silverpill@mitra.social @liaizon@wake.st @panos@catodon.rocks an "AI" tag could be good. You will still get the submissions, some good, some slop. And people who want nothing to do with it can filter it out.
@panos I'll admit I could see the LLM flavor in your proposals, but I read a lot of LLM text. I'm currently editing documentation for my project that I originally had generated. This is because while I spent time on establishing a tone up front, and review every page generated for accuracy (which it always has been), I want to make sure it's written in a way that doesn't waste people's time.
@panos I do think that for proposals it's very important people can feel you were deeply involved with the ideas, organization, and implementation of your proposal, in order to take it seriously. LLM flavored text can undermine that. To be fair, I think it's reasonable for people to be skeptical. I am still glad to have your FEPs though.
@stephen@planetearth.social I didn't try to hide the LLM flavor. A FEP is a document with information, not a literature piece. I am still in awe at this tbh - that FEPs can be rejected not because they are poorly written, not because they are inaccurate, but only because an LLM was used.
I get it when people argue about poor bug reports or poor PRs, and that's why I did my best to not waste the reviewer's time by just giving a prompt and letting it do whatever. But in this case I haven't received any negative feedback on the FEPs themselves - just one structural issue that I addressed. So I really can't understand why future FEPs "will likely be rejected" not based on their content, not based on whether they accurately describe a platform's behaviour so that others can implement compatible features, but only from what was used to write them.
@panos I can't speak for anyone else, but I do get fatigue reading un-optimized prose. It may not be literature, but well crafted words really do convey ideas with more clarity and less effort for the reader. The gap between good prose and serviceable prose does take a toll on the reader, and I try to be sensitive to that.
@panos I think your being a bit harsh Panos. Silverpill asked if the text was generated by an LLM not if there were LLMs involved. A standardiation process needs the humans who are doing the standardization to understand the material they are working with. silverpil is doing a huge amount of work reviewing in the fep process. If proposals were getting dumped in there at the rate LLMs can generate shit the whole thing wouldn't work.