Hacker News Reader

Comment by superkuh | Hacker News Reader

It's extremely obvious that the trained models for image generation are operating under fair use. There's absolutely no way a 3GB .bin file actually contains any substantial fraction of any one person's intellectual property. This is just magical thinking by some loud ignorant subset of the artist population. Or maybe not even that but just some lawyers claiming to represent them with a couple big name clients. I would not presume to think these loud ignorants represent even a substantial fraction of artists who probably understand the technical aspects of this and appreciate the new tools available.

Viewing a website and the multi-media files displayed on it is not inherently evil or bad and giving it the name "scraping" does not make it so. Let's not adopt this lawyer-ese framing. These people seem to want to have their cake and eat it too.

View on HN

If artists collectively notice (and can prove) a decrease of value of their work then I'd wager fair use is definitely broken, particularly: 4. the effect of the use upon the potential market for or value of the copyrighted work.

Right.

Just like there is "no way" that a 3 Lb hink of gray cells can memorize epic poems or figure out how to design ships to fly to other celestial bodies.

IOW: It doesn't work that way.

The GPTs do not store representations, they store what are effectively generation instructions. A formula ir program can generate nearly unlimited terabytes of data, but we don't need to store the output to recreate it; we need only store the formula or code.

So the "only 3GB" argument does not mean that sufficient instructions have not been stored via ingesting an artist's works to effectively infringe on that artist's works without literally storing copies.

>sufficient instructions have not been stored via ingesting an artist's works to effectively infringe on that artist's works without literally storing copies.

That's exactly what it means. It means no complete representations of the artists work are stored. The definition of fair use. Just like I'm not stealing anything when I go to an art museum and vague memories of some of the exhibits later influence a drawing I make.

> That's exactly what it means. It means no complete representations of the artists work are stored. The definition of fair use.

That's... not the definition of fair use.

> Just like I'm not stealing anything when I go to an art museum and vague memories of some of the exhibits later influence a drawing I make.

In the AI example, who is the ordinary human who viewed each of these works and is recalling them from memory?

The 3GB bin file.

Um, fair use is quoting a SMALL excerpt, or commenting on the piece.

Fair use is not reproducing the same painting in the same style, regenerating almost verbatim large blocks of code or text, etc.

But, while we can argue the boundaries of fair use all day, that is not the key point.

The key point is: the argument that "3GB is too small to hold anything that could qualify as fair use" does not hold water. Just because you or I don't know how it would work, does not not mean that it cannot work. And it has indeed been shown to work in recreating art, text, and code.

That OpenAI are now banning use of directives resembling "in the style of" is good evidence that they consider it a legitimate problem, although that is likely to be ultimately a speed-bump measure.

> but we don't need to store the output to recreate it; we need only store the formula or code.

This doesn't exempt it from information theory. If a 3GiB model was trained on 5 billion images, it fundamentally cannot have more than ~5 bits of information per training image on average, regardless of whether those are instructions for its recreation or a compressed copy of it.

Over-representation of some images (at the cost of the under-representation of others) is an issue, but it certainly cannot "produce a copy of any of its Training Images—which could number in the billions" as is claimed in the lawsuit against Stable Diffusion.

> So the "only 3GB" argument does not mean that sufficient instructions have not been stored via ingesting an artist's works to effectively infringe on that artist's works

Copyright applies to fixed works in a tangible medium of expression - and explicitly not formulae, procedures, or instructions like recipes. I don't think there's a bright line here, but this wouldn't be the direction I'd argue in if I wanted to classify machine learning as infringement.

Machine learning is incredibly good at lossy compression, 3GB is a lot of entropy to work with.

True that ML is good for lossy compression, but it's not "produce a copy of any image with less than a byte average of information" good, like is claimed in the Joseph Saveri Law Firm lawsuit against Stable Diffusion:

> Because a trained diffusion model can produce a copy of any of its Training Images—which could number in the billions—the diffusion model can be considered an alternative way of storing a copy of those images. In essence, it’s similar to having a directory on your computer of billions of JPEG image files.

I'm assuming you are not an artist?

Reply on news.ycombinator.com