Hacker News Reader

Comment by JacobiX | Hacker News Reader

It’s the same problem with those ML models, the other day someone generated a children’s book using GPT3, turned out that there is a real children's book with the same name and a very similar content: The Very Lonely Firefly by Eric Carle.

View on HN

Other thing I'm worried about: how to retract facts from ML model? I guess it's impossible, you need to retrain from scratch with part X removed from training set. Or... people could invent layered ML models similar to docker - each layer would be marked what data it was trained with. Then at least you'd have some cache of trained model to re-use in next training session. Nasty stuff.

Or instead of inventing complicated layered ML models Github could just use each repo's license information to decide what's okay to use. Detecting licenses is already a feature on that site.

Many licenses requite attribution, which would be hard to track.

Interesting, it's a big question I've had for a while, how 'original' stuff coming from these AI systems is, and also the distribution of uniqueness over many answers. I haven't dived into it yet, but I find it surprising how little this comes up when these systems are discussed (ie here on HN).

Does anyone even know? Can we even check? What if 1 in a thousand, or one in a million outputs is (very close to) something existing? I find this especially relevant when generating faces.

Reply on news.ycombinator.com