The source is heavily modified to clean up non-FLOSS from the Mozilla code. It’s not build reproducible vs Firefox no.
@Licaon_Kter That’s good background information for anyone who didn’t know that, thanks. Can anyone answer the questions I asked though?
So is all code required to perform the auto-translation and auto-summary there in the Fennec repo?
That’s not touched afaik, still straight from Mozilla
So … are you saying these auto-translation and auto-summary functions work without generative models? Which would be exciting news. Or are you saying nobody is checking that this code isn’t downloading proprietary “Open” models from Hugging Face during install, or interacting with proprietary services at runtime?
This example illustrates exactly why F-Droid needs a policy on generative models. To clarify what criteria they need to meet to be libre, and to reassure the community that we’re not downloading software that can’t actually be fully reproduced from scratch.
No, they work as Mozilla programmed them. Afaik, we only fix non-FLOSS code.
I support labelling AI-assisted projects. The main goal should be transparency, even if the label may become less meaningful as AI use becomes commonplace.
This does not answer the question, at all. If the honest answer is that you don’t know whether this app is downloading proprietary binaries at build or install time, or connecting to nonfree services at runtime, why not just say so? Because these evasive semi-answers give the impression that this is exactly what’s happening, and that you know this is what’s happening, but for some reason you don’t want to be transparent about it.
Again, I’m not sure what is being communicated by this statement. But what it implies is that you’ve stopped caring about software freedom as a general set of principles, and confined yourselves to whether or not the published source code for apps is under libre licenses. Again, if this is the policy, maybe it ought to be spelt out somewhere?
I’m directing you to the source of the answers you seek: Mozilla
you can also just read our changes in relan / fennecbuild · GitLab which is also the proper place to ask more about Fennec
So you think the toot will reveal what exactly?
What did Mozilla say when you asked? Which part of fennecbuild was unclear?
I wish people stopped recklessly using that stupid word “AI”. We don’t have actual artificial intelligence.
AFAICS we still don’t have technologically anything more advanced/complicated than doing a ImageMagick pass over a screenshot fragment before feeding it to Tesseract running on a fixed training model. To use modern words, long “prompts” inputted to an unchangeable black box. I’ve been doing that shit for decades. That’s not what humans do when doing the “AI’s” job manually and once you know how to train something as simple as Tesseract and generate the model, you learn to see the limitations of the process everywhere.
So why not call those things neural network, large language model, or whatever the proper name of current gen technology is.
Instead of feeding into this “AI” bubble for VCs and OpenAI on one side, also making it larger-than-life for everyone who fears it on the other one.
I think these are two different issues.
For F-Droid, the main thing should be whether the app and any included AI model are properly licensed and can be built from the available source. If the AI model is included with the app, its license and files should also be clear.
The generated-code part is more complicated. I don’t think code automatically gets all the licenses from the data used to train an AI model. It would be very difficult to prove that a specific piece of generated code came from a specific training example.
For F-Droid, it may be better to focus on things that can actually be checked: source code, licenses, included model files, and reproducible builds.
An anti-feature for apps that include generative AI could also be useful. It would simply tell users that the app contains AI, without saying whether that is good or bad.
There are lot of practical difficulties in verifying this and you can’t expect f-droid to be a pioneer in this regard, so right now the only options are either removal or an anti-feature tag. Or wait and see I suppose, wouldn’t be the first or the last bitter binary blob pill everyone other than the most pious foss monks swallows.
I’m not sure how you counted them, I think I’ve got one more 
- Apps that come with a local LLM which needs a model (trained data). The trained data is a binary blob of 10-1000MB usually, which F-Droid can’t reasonably reproduce (problem = F-Droid’s own ethics only)
That’s like having an apk of Tesseract OCR engine which needs a tessdata file of “rules” for how to OCR your pictures.
I agree with your opinion on that. My own issue is, should the trained data blob be embedded in the APK or downloadable. That’s kind of a big question for me since it’s usually huge (compared to most apks), but if hosted elsewhere it can be compromised or die from linkrot.
- Apps which were coded with the help of the LLM. The apk is reproducible from the code, but you would need the massive multi terabyte/petabyte training model from Copilot, GPT or whatever to reproduce the coding process (general ethics & legal problem).
That’s like writing a book by copying cool parts of other books, then arguing (AFAICS) that some of the books are so famous and old that nobody cares, while others are rare enough that no one will notice.
- I’ve also noticed apps that make some kind of local wrapper around a remote LLM of ChatGPT type. I don’t think there was any kind of local LLM in the app I tried, I think it wraps a prompt around your prompt to get the features it has, I haven’t looked into code. I used one of these to use a remote LLM anonymously until the app (or I?) ran out of free tokens.
Can’t see a licensing problem here, but it’s not great that I’m encouraged to spend money on API keys for a big corpo with tracking interests. This is the biggest one ethically if you worry about contributing/enabling to the “AI” hype
(It also makes my personal use case pointless, because I wanted an anonymous LLM that doesn’t guess who I am / what I like and answers independently of that. Can’t have this with a personal API key paid by a CC in my name.)
@hackermen4567 I’m pretty sure you were talking about my issue 1 but some other posts here are about issue 2 and/or 3, so it gets confusing. I think maybe mods should separate this into different threads
it’s all piled together in the OP itself, can’t split that, gotta start a new thread if you want it clean
One thread is enough imo.
Quick concrete question, since this keeps getting discussed in the abstract.
I maintain N-Zik, an Android music app on F-Droid ( GitHub - N-Zik-Group/N-Zik: A multilingual YouTube Music frontend for Android, focused on performance, UI/UX refinements, feature expansion, and long-term support. · GitHub ).
My workflow is AI-assisted (AGENTS.md etc.), but everything is human-reviewed before merge. I’m fully open about it.
Where would something like that actually land once this settles disclosure tag, antifeature, or fine as-is?
Not trying to derail, just curious how a case like mine reads to people here.
Prolly AI as a tag, with Anti Feature on.
Kind reminder, in 2.0 AFs are just a filter option