#aihorde — Public Fediverse posts

Nivenly, the cooperative that runs hachyderm.io, is having a community discussion and vote on what to do about Haidra and AIHorde, a distributed generative AI system.

@quintessence's Our First Community Discussion and Vote: Haidra has a summary and information about next steps. @Db0's The Birth of Haidra and the Join With Nivenly has background.

The discussion's open to all but voting is restricted to members They're also tracking discussion on social media in the Discord chatrooms but ask that official questions be submitted via Github discussions or email to [email protected]. They're gathering questions between now 8 Aug 2023, and will be. posting followup blog pots, followed by a general member discussion from August 14 - 25 August 2023 and an "election" after -- specific mechanism TBD,

Eligible voters including project members, trade members, and general members who pay $7/month). Something that's always interesting to watch in situations like this who's at the table. For example Nivenly is tech- and open-source focused. How well represented will artists, photographers, writers, actors, and musicians -- the people who tools like Haidra exploit -- be in the discussions and voting? What happens if somebody used #AIHorde ("a crowdsourced distributed cluster of Image generation workers and text generation workers") to generate fake avatars and profiles, sign up 1000 new accounts and voted to ban these exploitative tools?

Here's an excellent thread from @hrefna with some initial thoughts and questions https://hachyderm.io/@hrefna/110833561126744694). The "ACM" references in the thread are to the ACM Code of Ethics and Professional Conduct at https://ethics.acm.org/

Nivenly's discussion potentially affects cross-intance communication as well, and is a good chance to think about AI-related issues in general. In the Github discussion, for example, somebody brought up a question about whether posts and images that federate to hachyderm (or another nivenly-run instance) would be used to train future AIs. With Haidra, based on the discussion so far, the answer appears to be no; but this isn't the only situation where there's a risk of using federated data without consent to train AIs. What kind of ground rules do other instances want to have, and do we need to have conventions such as extending the nodeinfo structure (which gives basic data about every fediveres instance) to include information about whether or not an instance is using data for AI training?

So it's certainly a discussion worth tracking -- and participating in, if you're interested.

#haidra #nivenly

EDIT: A response in the github discussion describes why data federating to Hachyderm isn't being used to train Haidra's models, so I updated the text above to make that clear. It's still a good example of how this conversation highlights more general issues. If other instances using images and text they receive to train AI models, should they have to disclose it?