3 ms·
This is disingenuous. Large models are observed to reproduce works from their training data, sometimes verbatim, often close enough to constitute a derived wor
by klyrs 4y ago
This is disingenuous. Large models are observed to reproduce works from their training data, sometimes verbatim, often close enough to constitute a derived work. You scraping metadata would never be able to do that.
- marssaxman 4y agoThe dataset in question is not a model, and contains no images; it is a long list of URLs, with associated tags. Someone who wished to train a model could use this information to download a lot of pictures from various places on the internet, then create training and validation sets; but LAION themselves do not provide images or model weights. The photographer went after the wrong target here; if he wants his images taken off the internet he should contact whoever is hosting them at some publicly accessible URL.
- archontes 4y agoA model being able to produce an infringing work doesn't mean the model is an infringing work. I can produce work myself that is infringing or not infringing as a result of learning.
- Canada 4y agoI don't think your questioning of my sincerity is helpful. I agree that generative models have been observed to reproduce training input verbatim, but that's not what I understand to be at issue in the article. The issue here is whether you can presume the consumer of a list will do something illegal with an item on the list without any evidence, and therefore threaten the compiler of the list.