The models are open source meaning you can download them and run them. But the training data and code to train the model is not. So, they stills control the model, as there is no way to replicate it.

[–] rockerface@lemm.ee 3 points 3 weeks ago (5 children)

So if you can't replicate it, it by definition isn't open source, is it?

[–] Cochise 4 points 3 weeks ago (1 children)

The model is, in the sense you can modify it. Further train it, integrate in your app, etc. But the recipe to make the model is not.

And yes, it's less open source than we can think at first sight.

[–] rockerface@lemm.ee 3 points 3 weeks ago (2 children)

Isn't every software binary open source then? Since you can open it in a hex editor and modify it

[–] Cochise 3 points 3 weeks ago

But tou don't have permission to do. And hacking a binary is much more difficult than specializing a model, for instance.

[–] hendrik@palaver.p3x.de 2 points 3 weeks ago* (last edited 3 weeks ago) (1 children)

Yeah, that's kind of AI companies' definition of open source... Other companies just have "open" in their name for historical reasons. The FSF doesn't really agree ( https://www.fsf.org/news/fsf-is-working-on-freedom-in-machine-learning-applications ) and neither do I. It's "open weight". Or I'd need to see the datasets and training scripts as well.

[–] rockerface@lemm.ee 4 points 3 weeks ago

Yeah, "open weight" seems a more appropriate label. It still seems better than a fully proprietary system, but calling it open source without clarification is misleading.

load more comments (3 replies)