2026-07-06 09:03:53
2026-07-04 19:40:23
2026-07-04 19:34:03
18868477
RE: toot.fedilab.app/@apps/1168632…
Some people are wary of AI, and I understand that. But LibreTranslate, one of the most ethical translation services and already used by #Fedilab on its own server, was still built through AI, with pretrained neural models. The real question is not the technology itself, it is where your data goes. And here nothing leaks, since everything stays on my own server: a local model, no cloud and no third party.
The feature is only here to help disabled people, and only you can validate the content.
An idea for #Fedilab: a feature to generate a description for your photos to help accessibility. You post an image, and a suggested alt text comes back for you to review and edit before publishing. It would run on a self-hosted open source vision model (#MoonDream: github.com/vikhyat/moondream) on my own server, in a closed loop with no cloud, no third-party service and no image stored. Would this be useful to you?GitHub - m87-labs/moondream: tiny vision language model
tiny vision language model. Contribute to m87-labs/moondream development by creating an account on GitHub.GitHub
Etch
in reply to Fedilab Apps • • •Fedilab Apps
in reply to Etch • • •No, but that would be a simple addition, pointing to a different server for such services. Thanks for the suggestion.
Ulrich_the_Elder, 🇨🇦,🇺🇦,🇵🇸
in reply to Fedilab Apps • • •Fedilab Apps
in reply to Ulrich_the_Elder, 🇨🇦,🇺🇦,🇵🇸 • • •This feature is mostly there to help people who do not describe their images at all. Fedilab already warns by default when an image has no description, and going further, it could suggest one to educate. Descriptions matter because blind readers rely on them to know what an image shows. The human always validates before posting, so nobody is replaced.
Paco Velobs
in reply to Fedilab Apps • • •I'm not sure you can choose what "The real question" is here.
It sure is the one that matters to you. But, for others it might be the origin of the data it's been trained on.
Being the most ethical is not equivalent to being ethical, far from it.
I recon the accessibility quest is noble.
But I think this feature should be opt-in, not in or opt-out.
Fedilab Apps
in reply to Paco Velobs • • •Of course it will be disabled by default. The idea is only to help people who want help writing a description, nothing will be forced.
Paco Velobs
in reply to Fedilab Apps • • •Like I told my boss the other day, I can't forbid others to use it but I'll stay as long as others can't force me to use it either.
Sam
in reply to Fedilab Apps • • •You'll have to put on your boxing gloves when you come around here supporting AI.
Defending a niche output of AI (translation services) when the greater Machine is physically digging up rural communities, sucking down resources (WATER) and making billionaires richer, is showing that LibreTranslate is not a good steward of resources and lacks a moral compass in relation to those who ultimately benefit.
"since everything stays on my server..." is a fallacy. Just wait.
Fedilab Apps
in reply to Sam • • •The environmental cost of AI is real, but it comes from the big training, not from running a small model that already exists. I train nothing and I collect nothing. It is only local, on a server I already have. This is very far from the big datacentres you talk about. I will also try to share where the training data came from.
Sam
in reply to Fedilab Apps • • •Trying to share 'training' data going forward is a nice non-gesture. What LLM are you are using, and what company is it traced back to? Who really owns you at the end of the day?
Rémi Letot
in reply to Fedilab Apps • • •No, the environmental cost is also on inference, not only in training. Specifically, you'll have to monopolize ressources to run the model and do what you intend to do.
Also, using a model "that already exists" validates its creation. So you share part of the cost of training it and its successor.
And last, good luck finding who they stole the training dataset from. I'm interrested, if you ever find out.
@Nead
Fritz Adalis
in reply to Fedilab Apps • • •Willow (she/her)
in reply to Fedilab Apps • • •not knowing a ton about the specific language used to differentiate neural model, etc... But as someone who has a hard time sometimes with writing alt text... I am curious.
Before the ai hype I found machine learning so interesting and the applications fascinating. I'd have concerns about the source of the training - was the model developed ethically, and what is the climate impact. Those are my two fundamental issues with ai I think.
wortezimmer
in reply to Fedilab Apps • • •Lauma Pret ♡ ᪲᪲᪲🌈
in reply to Fedilab Apps • • •