No comment.
The furniture issue was one of the funniest bugs I've ever worked on. For a couple weeks I felt like I was being gaslit because the test model on DreamStudio was turning everything into furniture. Want a picture of the jungle? How about a painting of a jungle above a couch.
Copyright in training is tricky. I'm extremely sensitive to laws/regulation/precedent which would make only the biggest (well-funded) players able to compete on models. We have to figure out how to reward/protect creators without creating a new set of content monopolies.
As for U.S. law, I do believe training is "transformative" under fair use. The worst-case scenario is that we literally blind our robots to the world in the name of copyright. Models and machines should be free to observe what's observable.
People who use AI to blatantly break copyright should fall under the same rules as people who use Photoshop or literal paint to do so. I don't think going after the tool makers is the right move.
Learning from the observable world should not be gated, who can we trust with that power? The authority to even attempt to enforce what's allowed to be witnessed terrifies me.
A robot's view of Times Square after DRM for reality takes hold.
