this post was submitted on 09 Aug 2023
379 points (100.0% liked)
Technology
37702 readers
291 users here now
A nice place to discuss rumors, happenings, innovations, and challenges in the technology sphere. We also welcome discussions on the intersections of technology and society. If it’s technological news or discussion of technology, it probably belongs here.
Remember the overriding ethos on Beehaw: Be(e) Nice. Each user you encounter here is a person, and should be treated with kindness (even if they’re wrong, or use a Linux distro you don’t like). Personal attacks will not be tolerated.
Subcommunities on Beehaw:
This community's icon was made by Aaron Schneider, under the CC-BY-NC-SA 4.0 license.
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Practically you would have to separate model architecture from weights. Weights are licensed as research use only, while the architecture is the actual scientific contribution. Maybe some instructions on best train the model.
Only problem is that you can't really prove if someone just retrained research weights or trained from scratch using randomized weights. Also certain alterations to the architecture are possible, so only the "headless" models are used.
I think there's some research into detecting retraining, but I can imagine it's not fool proof.
I kind of think that as proof-of-concepts, the AI models are kind of interesting. I don't like the content they produce much, because it is just so utterly same-y, so I haven't yet seen anything that made me go "wow, that's amazing". But the actual architecture behind them is pretty cool.
But at this point, they've gone beyond researching an interesting idea into full on commercial enterprises. If we don't have an effective means of retraining the existing models to remove the data that isn't licenced for commercial use (which is most of it), then it seems the only ethical way to move forward would be to start again with more selective training data, including only what is commercially licenced. Now the research has been done in how to create these models, it should be quicker to build new ones with more ethically sourced training data.