@vt52 I would disagree on the transparency there. One of the points of the ethical assessment is:
> Is the training data available and free to use?
Consider how StackOverflow is basically arguing that the stuff people wrote on the site is "free to use" (as it is on a CC By-SA license), but the community outcry seems to suggest that they are not exactly on board with that interpretation.
LocalAI gets a Green rating, for example. But I cannot find info on the training data… 👀
@vt52 also, there are multiple lawsuits currently that focus on the copyright question. If they go a particular way, it might potentially mean that:
AI models are derivative works of the training data they are trained on.
Anything done with the input from these AI models are in turn derivative works of these models.
IOW, if works licensed under CC By-SA (Wikipedia, Wikimedia Commons) are in the training corpus, suddenly anything created with these models might be CC By-SA.
@m0bi @Michał "rysiek" Woźniak · 🇺🇦 I see no mentioning of origins and ownership of their training data playing any role in their ethicality rating.
They do only offer
"Is the software (both for inferencing and training) open source?
Is the trained model freely available for self-hosting?
Is the training data available and free to use?"
> I see no mentioning of the origina of their trainigh data
There is no "their" (Nextcloud) LLM so no "their" data used to train. There are integration options only. You decide what you use with your Netcloud instance, if you want LLM in Nextcloud. Its optional.
Secondly, it mentions it can use Llama, but also mentions that "the training data is freely available". I doubt both of these can be true. Unless I missed something major, Llama training data have not been released by Meta.
There they mention models like "text2image_stablediffusion2". So this applies: https://en.wikipedia.org/wiki/Stable_Diffusion#Training_data
> An investigation by Bayerischer Rundfunk showed that LAION's datasets, hosted on Hugging Face, contain large amounts of private and sensitive data.
@rysiek@nextcloud They integrated it in all the places(tm) and you just have to plug in an API. LocalAI, o4 whatever other $thing you may fancy. And then all the funny functions work.
@rysiek@tfiebig@nextcloud Tobias is correct. Nextcloud isn't an AI company, we don't have our own LLMs. First, all AI features can be disabled because some folks don't like AI. Second, you can bring your own model now with Nextcloud Hub 8, or even integrate with OpenAI API. It's really up to you what flavor of AI you prefer to use. Your data, your choice.
The thing I am confused about, I guess, is the messaging around the "AI" assistant being the "centerpiece" of Nc Hub 8. What I understand from what you said and what I found in the docs, the Assistant requires some model to be available. Is there a model distributed with Nextcloud for that purpose?
Add comment