How much of F-Droid is LLM generated?
10 points by lobo
10 points by lobo
The author says "Any kind of agentic infrastructure automatically lands an app in this tier as I do not believe it is possible to use AI responsibly from within a coding harness."
Which seems like kinda a weird take to me. I can understand a general sentiment of wanting to know how much slop to expect in a project. I can understand a principled objection to any AI use. But that seems like a weird midpoint line in the sand to me. I tend to think:
Dev infrastructure seems like a high-gain low-risk place to see agent use. Testing, reviews, CI/CD, etc... all were frequently places that got neglected and collected human-produced slop even before agentic tools. If the agent is bringing in some quality, rigor, or just action to those things where it wouldn't have existed at all in its absence, I'm not sure I'd object to that.
I particularly don't understand the "from within a coding harness" bit, as if you're accepting genai content at all, it seems like the harness is just likely to steer it towards better quality. A tool can be used or misused, but this feels like blaming the tool itself.
I feel this problem grows out of one of the long term limitations of F-Droid: the only metadata the client gives users trying to find apps is how recently the app version was uploaded, and which Free Software "anti-features" the app might contain.
In the pre-LLM era this already meant many of the apps which appeared first were people's pet projects that were not really ready for anyone else to use them. Back then I would click through and read the repos to try and figure this out (a technique which depends on you being software development literate, with spare time to burn.)
Now in the LLM era the problem is turbocharged, and vibecoded projects can release as often as they tell the agent to do so. (The author's methodology was also based on recent uploads, and I suspect this creates some sampling bias compared to the whole of F-Droid - but that doesn't matter because the F-Droid client applies the same bias!)
The volunteers behind F-Droid seem hardworking and dedicated, and I hope they use this challenge as an opportunity to provide users with additional insight beyond "does this app connect to a Non-Free service?"
(Such as: curated collections, user reviews, links to external reviews, metadata about the repo and contributors, etc. The latter would help flag slop apps without needing to try and automate the LLM-or-no process that the author went through here.)
I guess a different model would to be to try and make a FOSS app store that's more like a Linux distro: app authors don't upload their own apps but instead there's a group of volunteer "app store developers" who do this part.
(Don't know if that's been tried before, free idea for anyone who wants to!)
This can be done in F-Droid even. Just have to create an alternative repository that is maintained this way.
I've been a bit disappointed in the f-droid app itself... it seems to almost actively hide things you'd want to know before installing something. I've always made a habit to research apps on their website, which has been a bit more forthcoming with details, before installing.
F-Droid has always had an app quality problem in general if you ask me. I'm talking broken apps, apps only working on old versions, unmaintained apps, abandon ware, outdated versions up for weeks (that last one is technically due to F-Droid's own build schedule iirc).