I find this take interesting. So would you also argue that saving an image into computer memory is the same as memorizing an image for a human? Those processes are viewed very different by the law, but if we anthropomorphize computers should we not view them the same?
Also I wonder where you get the view that future ML systems will not require large amounts of learning? I don't see any development in current systems that would allow that, or do you mean you have a network trained on large amounts of data which can then adjust to a style from a single image? If that's the case we are still at the same question, how was the original model trained.
> Those processes are viewed very different by the law, but if we anthropomorphize computers should we not view them the same?
Not only do I think the two processes are essentially the same, but I can't think of any laws in my jurisdiction (the UK) which actually distinguish between them.
E.g. we are allowed to make copies of digital media for personal use.
I'm actually not sure about that. It's certainly against the terms of the theatre, but assuming you are only using the recording for personal use, I'm not sure if you would be breaking any laws.
For the second question, yeah exactly, as long as you've trained the rest of the system to a certain degree you can certainly do one-shot training on top of that now already for object recognition for example and you would be able to do it for style acquisition for diffusion models as well soon I think (you can already pretty quickly do overfit training on them, at home, in a couple of minutes with 10-20 images).
Essentially this is what the brain does when you do oneshot learning of traffic signs or characters when learning a new alphabet etc. (yeah sometimes it's not that easy but still it's "theoretically" possible :). The rest of the recognition pipeline is so general that styles and objects etc are just a small icing on the cake to learn on top, you don't need to retrain all the areas of the brain when adding a roadsign to your driving skill set.
But my point was that you could train the rest of the network on more general public data and not greg rutkowski. Hooray. Then someone shows it a single greg image and you're back to square 0.
Also I wonder where you get the view that future ML systems will not require large amounts of learning? I don't see any development in current systems that would allow that, or do you mean you have a network trained on large amounts of data which can then adjust to a style from a single image? If that's the case we are still at the same question, how was the original model trained.