I’m skimming the paper and Nightshade is not doing what a reasonable layperson might assume from the various headlines I’ve seen. It attacks models that use open source sub-models. The models that are the biggest threats to artists (as far as I can see) don’t use open source sub-models.
And it’s only a bit better than just mislabeling an image. And I had a good idea of that before I started reading, because I know just enough about these things to apply some rules of thumb. “Mammals with thicker fur are usually from colder climates” or “most prime numbers are odd” level stuff.
We’re going to keep getting excited (positively and negatively) about the wrong stuff in this field until we sit down and figure out a general sense of how it works, and the service journalism in that direction – that I’ve seen – has mostly been very weak so far.
This feels like one of those classic things where every link in the chain between the lab, the press office, the journalism, the headline, and social media did a reasonable amount of rephrasing, but the idea left in the reader’s head is not supported by the source material.
This being the public internet, here are a few standard disclaimers: Breaking the sense of inevitability in “AI” PR is good. Research in this direction is vital. Artists should make good livings. Our current intellectual property system is not ideal. Overall, whatever you think, I mostly agree.
Eh, it's a lab built on hype that churns out mediocre products. A lot of self aggrandizement. Also, their previous products have reused code in ways contrary to the GPL that it was written under.
