Comment by csallen | Hacker News Reader

Comment by csallen | Hacker News Reader

Parent Full thread

Most people know 20,000-40,000 words. Let's call it 30,000. You've learned 99.999% of those 30,000 people from other people. And don't get me started on phrases, cliches, sentence structures, etc.

How many of those words do you remember learning? How many can you confidently say you remember the person or the book that taught you the word? 5? 10? Maybe 100?

That's how brains work. We ingest vast amounts of information that other people put out into the world. We consume and it incorporate it and start using it on our own work. And we forget where we even got it. My brain works this way. Your brain works this way. Artists' brains work this way. GPT-4 works this way.

The idea that a visual artist can somehow recall where they first saw many of the billions of images stored in their brain -- the photos, movies, architecture, paintings, and real-life scenes that play out every second of every day -- is laughable. Almost all of that goes uncredited, and always will.

This is what it is to learn.

I tend to fall more on the "training should be fair use" side than most, but your comment seems to be missing the point. Nobody is arguing that models are violating copyright or social norms around credit simply because they consume this information. Nobody ever argued/argues that the traditional text generation in markov models on your phone's keyboard runs afoul of these issues. The argument being made is that these particular models are now producing content that very clearly does run into these norms in a qualitatively different way. You cannot convincingly make the argument that the countless generated "X, but in the style of Y" images, text, and video going around the internet are exclusively the product of some unknowable mishmash of influences -- there is clearly some internalized structure of "this work has this name" and "these works are associated with this creator".

To take it to an extreme, you obviously can't just use one of the available neural net lossless compression algorithms to circumvent copyright law or citation rules (e.g., distributing a local LLM that helpfully displays the entirety of some particular book when you ask it to), you can't just tweak it to make it a little lossy by changing one letter, or a little more lossy than that, etc., while on the other hand, any LLM that performs exactly the same as a markov model would presumably be fine, so there is a line somewhere.

A company hires an artist. That artist has observed a ton of other artists' work over the years. The company instructs that artist to draw, "X but in the style of Y", where Y is some copyrighted artwork. The company then prints the result and puts it on their packaging.

A company builds an AI tool. That AI tool is trained on a ton of artists' work over the years. The company opens up the AI tool and asks it to draw, "X but in the style of Y," where Y is come copyrighted artwork. The company then prints the result and puts it on their packaging.

What's the difference?

I'd argue there isn't one. The copyright infringement isn't the ability of the artist or the AI tool to make a copy. It's the act of actually using it to make a copy, and then putting that out into the world.

The artist has a claim for production of a derivative work and for passing off against the other artist.

> What's the difference?

Ultimately, only high courts in each jurisdiction can decide. I can imagine a case where some highly advanced nations decide different interpretations that cause conflict. Then, we need an amendment to the widely accepted international copyright rules, the Berne Convention. Ref: https://en.wikipedia.org/wiki/Berne_Convention

Okay, but then that's an an argument subject to the critiques made upthread that you were initially trying to dismiss? You can't claim that AI doesn't need to worry about citing influences because it's just doing a thing humans wouldn't cite influences for, then proceed to cite an example where you would very much be expected to cite your influences, and AI wouldn't, as evidence.

I never argued that AI doesn't need to worry about citing influences. If I am a person using a tool to create a work, and the final product clearly resembles some copyrighted work that I need to reference and give credit to, what does it matter if my tool is a pencil, a graphics editing program, a GPT, or my own mind? I can cite the work.

Like I said, this is exactly what the comment you first replied to was explaining. It is very clearly not the same as a pencil or a graphics editing program, because those things do not have a notion of Cat & Girl by Willem de Kooning embedded in them that they can utilize without credit. It is clearly not the same as your mind, because your mind can and, assuming you want to stay in good standing, will provide credit for influence.

Again, take it back to basics: do you believe it is permissible to share a model itself (not the model output, the model), either directly or via API, that can trivially reproduce entire copyrighted works?

I'd say that a tool itself can't be guilty of copyright infringement, only the person using the tool can. So it doesn't matter if the GPT has some sort of "notion" of a copyrighted work in it or not. GPTs aren't sentient beings. They don't go around creating things on their own. Humans have to sit down and command them, and that point, whoever issued the command is responsible for the output. Copyright violation happens at the point of creation or distribution, not at the much earlier point of inspiration or learning.

So yeah, of course imo it should be permissible to share a model that can reproduce copyrighted works. Being "capable of being used" to violate a law is not the same thing as violating a law.

A ton of software on my computer can copy-paste others' work, both images and words. It can trivially break copyright. Hell, there are even programs out there than can auto-generate code for me, code that various companies have patent claims for. Do I think distributing any of this software should be illegal? No. But I think using that software to infringe on someone's copyright should be.

(Note: This is different than if the program distributed came with a folder that included bunch of copyrighted works. To me, sharing something like that would be a copyright violation.)

I'm not sure how to explain this any clearer. I am talking about neural net compression algorithms. As in, it is literally just a neural net encoding some copyrighted work, and nothing else. It is ultimately no more intelligent than a zip file, other than the file and program are the same. You can't seriously believe that these programs allow you to avoid copyright claims, can you? Movie studios, music producers, and book publishers should just pack it in, pirates just need to switch to compressing by training a NN, and seeding those instead, and there's no legal precedence to stop them? If you do think that, do you at least understand why nobody is going to take your position seriously?

A neural net designed to do nothing other than compress and decompress a copyrighted work is completely different than GPT-4, unless I'm uninformed. To me that sounds like comparing a VCR to a brain. GPT-4's technology is clearly something that "learns" in order to be able to produce novel thoughts and ideas, rather than merely compressing. A judge or jury would easily understand that it wasn't designed just to reproduce copyrighted works.

> It is clearly not the same as your mind, because your mind can and, assuming you want to stay in good standing, will provide credit for influence

I forgot to respond to this, but it's not true. Your mind is incapable of providing credit for 99.9% of its influence and inspiration, even when you want it to. You simply don't remember where you've learned most of the things you've learned. And when you have a seemingly novel idea, you can't always be aware of every single influential example of another person's work/art that combined to generate that new idea.

> A neural net designed to do nothing other than compress and decompress a copyrighted work is completely different than GPT-4, unless I'm uninformed.

Compression and the output from LLMs are cousins. The model tries to predict what continuations are likely, given context. Indeed, it takes a lot of effort to make LLMs less willing to just output training data verbatim. And conversely, you can get compression algorithms to do things similar to what LLMs do (poorly).

Whether this also describes most of human cognitive process, is subject to debate.

Individual words aren't comparable to the things people are worried about getting copied. People are much more able to tell you where they learned about more sophisticated concepts and styles.

The same principle applies, though. They can tell you maybe a dozen, maybe a few dozen, concepts they've learned and use in their work. But what about the thousands of concepts they use in their work they can't tell you about? The patterns they've noticed, the concepts that don't even have names, but that came from seeing things in the world world that were all created by other people?

For example, how many artists drawing street scenes credit the designer at Ford Motors for teaching them what a generic car looks like? How many even know which designers created their mental model of a car?

That is again a single word.

There is a strong correlation between how copyrightable a concept is and how well you can point to where you learned it.

Reply on news.ycombinator.com