240
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
this post was submitted on 09 Oct 2023
240 points (87.5% liked)
Technology
59438 readers
4317 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each another!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
Approved Bots
founded 1 year ago
MODERATORS
Where? Closest I can find is a reference to an "AI-generated nude of someone who looked like Margot Robbie, and another image of Lopez."
"Looking like" a person is not at all the same as "reproducing an exact replica of an input."
I think the idea is that if the model doesn't know what Jennifer Lopez is, it couldn't make imitations of her naked.
Realistically that ship has sailed and AI is capable enough now that even if the data wasn't there it could be pretty easily added.
It will need to become a simple fact of life. If we can imagine something now, we can have pictures of it. There is no putting this back in the bottle.
"Knowing what Jennifer Lopez looks like" is a very distinct thing from "reproducing an exact replica" of training data. OP appears to be arguing that the former is not true because he thinks the latter is true, but it's actually the opposite. That's the crux of what I'm arguing here, OP is simply factually wrong about his position.
Edit: OP has pointed out that he doesn't actually think there are exact replicas being produced, which just makes this even more confusing.
Your misread their first comment, I think.
They were saying that DESPITE the common arguments that AI only learns and doesn't copy exactly it might still be good to require consent for people's content to be in training data.
Oh, it's a reference to that paper.
Firstly, that paper was written in January and examined a Stable Diffusion model that was already obsolete due to its poor training even back then. Secondly, even with that poor model they had to move heaven and earth to find a handful of examples out of hundreds of millions of training examples where they could get a blurry replica out.
Here's a Reddit thread from back in the day discussing how, really, this sort of thing just proves how difficult it is to do this.
Secondly, as mentioned, that model is long obsolete due to issues exactly like this. Modern models work better in part because they have better curated training sets that eliminate this sort of "overfitting." There's no indication in this article that the website in question is using one of those old models, it's just presented as a hypothetical concern.