2,745 karma · joined November 27, 2019
There are uncountably many countable collections of symbols.
Edit:
The above is true if your symbols are countable. If they are uncountable, then obviously there are uncountable finite sets of your symbols (As the singleton sets are a subset, and those are uncountable).
The main one that my previous post was getting at was (A) is not an axiom in your ethical framework; it's derived. And I think you'll find if you try and build up to (A) from base principles, sentience won't be a condition of the ethics of an action.
A secondary one is that (C) is true regardless of what you replace LLMs with, because you've outright state that you don't have a means to judge what sentience is. You could take your argument and use it for:
- Humans
- Cats
- Ants
- Car engines
- Rocks
- Air
And it will hold, since you aren't relying on any particular properties of LLMs to make your case.
You can assume that other people are conscious like you, and so you put ethical value on that. But then you're really putting ethical value on "being like me", irrespective of whether other people are or are not conscious.
I don't think a definition of consciousness is possible in any objective sense, as consciousness is by definition, a subjective experience.
A more useful question is why ask if LLMs are conscious?
ex:
1. Are you trying to decide whether shutting them off is ethical?
2. Are you trying to figure out how to manipulate them to ask your questions correctly?
3. Are you trying to answer whether they can replicate anything a human can do?
When the contextual question that prompts the word "consciousness" is defined, we can make progress in determining whether something is "conscious".
For example, if it's a question of ethics, then you follow up by defining your ethical framework.
Can you explain how this is any different than what Google does?
It's unbiased by definition of "does the output reflect the input"?
It's not unbiased by definition of "does the output reflect reality"?
Can you point me to a ruling where this form of expression is successfully defended via fair use?
Seems unlikely that a judge would agree that a simple transformation between "digital" vs "non-digital" medium avoids copyright infringement (ex: if I produce a digital image, I definitely maintain copyright over digital and non-digital forms of that image; you can't just start printing and selling posters of my image).
You could try to argue that the change of medium is transformative, but that's not a generic argument for "digital" vs. "non-digital"; that's particular to the circumstances of the case.
I think you misunderstand me. Despite my feelings about Stable Diffusion and laws around copyright, I believe the law will find in favour of the artists, assuming they have the funds/will to see it through. The law unfortunately doesn't really care how I feel about it.
Granted, they're up against some fairly large corporations that have both time and money to stall for a while, so I don't know how far their lawsuit will go.
This doesn't really have much to do with copyright.
> Is literally anyone using Stable Diffusion with the intent to reproduce one of the plaintiff's copyright protected works?
Intent isn't really a factor in terms of damages for copyright infringement. But I assume we'll figure out the extend of the damages if/when damages are awarded, since that's when they measure it out. Granted, I expect the liable parties to settle far before it gets to that; the payout seems cheap (~100k - 1m) relative to the risk (penalties for copyright infringement are per unit, and it's easy to hit millions of units on the internet).
Yes, and so is Winzip, and the code that trains Stable Diffusion. But that doesn't have much to do with the resulting data that is used and generated by those files.
> There were and still are other torrent sites that have non-infringing utility which are of course allowed to exist.
Can you point me to a torrent site that withstood a legal challenge for copyright infringement?
Or derivative works.
> not the instructions that could lead to the image,
And yet torrent sites have been found to violate copyright, despite not hosting any of the copyrighted content.
> not the 1s and 0s that make up the image.
Now this is plainly false. I cannot start selling a Mickey Mouse JPEG stripped straight from Disney's site.
> You'll notice that CafePress still has t-shirts with the DeCSS code for sale: https://www.cafepress.com/+,954530397
1) This is covered by a trade secrets law, and not copyright.
2) It still being for sale doesn't mean it doesn't violate copyright; only that the copyright holder hasn't litigated to stop it. Put differently, getting away with something doesn't make it legal.
Look, I think Stable Diffusion is pretty neat. But all of this "it's just math"/"it's just data"/"it also creates non-copyrighted images" stuff doesn't cut it as far as the law is concerned. If your point is that being in possession of the model isn't violating copyright, sure.
I'm talking about distributing or otherwise providing a service that uses the model to generate images. If they figure out how to get it to stop generating training data, they'll be fine. But with existing approaches, that'll be nigh impossible.
Also, that usage of Mickey Mouse might be copyright infringing, but fall under fair use (probably parody), which is a specific defense of copyright infringement (similar to "self defense").
What I am saying is that:
1) If your model returns images which look near-identical to your training data, then any copyright infringement that applies to the training image will also apply to your image.
2) If your model can consistently return copyrighted imagery, there's little difference between explicitly sharing those images (with a password) and implicitly sharing them (via a model + prompt).
If you're convinced that a photo of Mickey Mouse with slightly larger ears, or slightly reddish pants isn't copyright infringement, then sure, neither is any of this stuff. It would also mean that republishing copyrighted images with lossy compression algorithm (IE, JPEGs) would also not violate copyright.
I would suggest looking at the actual laws around copyright instead of relying on what you feel copyright should be.
Here's a whole paper, complete with citations:
https://arxiv.org/pdf/2212.03860.pdf
> These models do not contain copies. One way to describe the data is they contain a statistical breakdown of the artwork, which is substantially different from a JPEG -> PNG conversion you mention.
I don't understand the distinction you're making. What legally separates a "statistical breakdown" representation from a zip file representation, JPEG representation, PNG representation?
For (1), I think it's handy to think of an encrypted zip file that contains a bunch of images. If you give it a valid password, it'll give you a corresponding image. If it produces an image that violates copyright, it clearly violates copyright, even if most of the images contained in it don't.
It's possible that if the model required sufficient instruction to reproduce Mickey Mouse ("create a line from (0,20) to (20,30) of thickness 2, then a line from ..."), and could not do so without that, then the argument wouldn't apply. And there's definitely a grey area once we're outside of the territory of reproducing images. But we definitely aren't anywhere close to that with these models, and looking at the "AI Safety" work, it doesn't seem possible to get these models to "behave".
> This would be a good thing but wouldn't prevent AI image generation from broadly doing the things that the plaintiffs want to stop it doing.
Yeah, I agree that the plaintiffs aren't suing solely because the model can easily regenerate their images. But I think we'll find its insanely hard to get these models to avoid producing copyrighted images. They may be forced to license the training images.
If someone was distributing a password protected zip-file where if I type in "Mickey Mouse", I get a photo of Mickey Mouse, the zip file violates copyright. It doesn't matter that the underlying representation of Mickey Mouse is distorted in storage and recovered later. It doesn't matter that in order to retrieve the image, I have to go through a series of operations.
Of course, if I opened the zip file and put the image on a t-shirt, I would also be violating copyright. But the zip file itself is copyright infringing.
If I take a JPEG of Mickey Mouse and then turn it into a PNG, it's not a "copy", as the bits are different. But it still contains copyrighted material.
You can try to argue that the bits of PNG itself aren't an image of Mickey Mouse, but rather the algorithm that reads the PNG produces an image of Mickey Mouse. But that isn't really a relevant distinction in so far as copyright is concerned.
In addition, this statement is false:
> The important distinction is that the model does not contain a copy of a copyrighted material.
It has been shown repeatedly that the model produces copies of training data. The copies are of course not stored as JPEGs/PNGs in the model, but they are retrievable from the model, given the correct password (prompt).
And to preempt a likely counterargument, intent is not a required factor for establishing copyright violation.
We can break this into (1) and (2):
(1)
The model is a copyright violation as it contains untransformed copyrighted material. You can show this by simply asking for a picture of Mickey Mouse.
To preempt a few (obvious) arguments about what it means to be "transformed", keep in mind an encrypted zip file containing images of Mickey Mouse would not be considered transformative, even if it happens to contain images that aren't copyrighted.
(2)
> then it does so to exactly the same degree as it would if it wasn't AI generated.
This is the conclusion of an argument, but I'd need to see how you show that "it does so to exactly the same degree". Right now, it looks like you're taking that as a base assumption.
To be clear, my argument doesn't show that statistical generative approaches to art are violating copyright by using copyrighted data. It only shows that the argument "humans learn on copyrighted stuff therefore it's legal for AI to do it" isn't a sound argument.
Another argument that relies on the definition of derivative work might be able to show that you can train a model on copyrighted data without producing works that violate copyright.
If I take a picture of Mickey Mouse, add 1 to each RGB value, it would still be covered under copyright.
I think the term 'learn' is doing the heavy lifting in your argument here. Specifically, your argument seems to be:
1. We use the term 'learn' as shorthand for 'fitting a machine learning algorithm to data'.
2. We also use the term 'learn' for a human looking at copyrighted material to practice their craft.
3. Humans using copyrighted material to learn doesn't make their works derivative.
4. Therefore, machine learning algorithms that are trained on copyrighted material aren't producing derivative works.
This argument holds an implicit assumption: if we use the same term for a process, it should covered by the same rules/laws. Which I think isn't very sound.
I interned at a dating site a while ago and there's no need to taint anything. People are far too complicated to predict beyond who will send a message to who. Also, the problem is adversarial (IE, people misrepresent themselves as the person they want to be as opposed to the person they are), so trying to find the signal is a continuous game of cat and mouse.
Hate to break it to you, but it passed into the mainstream years ago.
How would this be different in experience than Google's existing "summary cards" that pop up?
What kind of questions are you having trouble answering via Google?
- The Acorn: Not in Gastown. In South Main.
- The Naam: Not in Gastown. In Kitsilano.
- Chau Veggie Express: Not in Gastown. In Granville Island.
- The Essential: Doesn't exist.
- The Wilder Snail: Not a restaurant, not vegetarian, not in Gastown. In Japantown, which is at least walkable from Gastown.
It can replace queries like "what's the tallest mountain in France", but that's not really a profitable query for Google.
Whereas it's unclear to me how it replaces "best vegetarian restaurants in Gastown". It can string together a coherent statement, but then I have to go search to see whether it actually recommended vegetarian restaurants and if they are in fact any good. At which point, I'm back to using a search engine.
And yes, you could wire it to show related articles. At which point you've just recreated a local search index (with GPTChat providing the Google summary card), which you now have to:
1. Figure out how to solve SEO hacking for your model.
2. Update the model regularly with new information.
Putting you in the exact same position as Google, minus twenty years of experience doing it.
> whenever i harsh Buddhism's vibe, some fucker brings me a "well actually Buddhists believe the opposite of that", which is way too easy given what i have now decided is the Third Law of Buddhism:
> for every Buddhism there is an equal and opposite Buddhism somewhere in Asia
Unity has a much larger community and will have much more in the way of intro tutorials, so I started there.
But regardless of the engine you choose, you'll probably spend the better part of a year building tiny games (like tic-tac-toe -> pong -> breakout -> Tetris -> etc) before you can execute on anything more complicated (like a 3d RTS).