Live data from Hacker News

Judge said Meta illegally used books to build its AI

wired.com

261–270 of 352 posts

Re: Judge said Meta illegally used books to build its AI

#261
post #189

Earlier quoted context omitted.

“Copy” is ambiguous here. Of course data is copied during training. That said, OP is referring to whether the resulting model is able to produce verbatim copies of the data.

A good way of thinking about this is: consider the case where the data in question is illegal. Could you get into trouble for not only having access to it but also making copies of it? There’s plenty of case law there…

I would argue that as an individual, real person, obtaining content without a license and personally consuming that content is significantly different than a corporation doing the same. My rational is that distribution of that content is (or should be) the primary offense. If I work for a company and they direct me to collect a bunch of content without a license and then I pass that to other members in my team to train a model, I've now distributed that content at the direction of my employer. That should be the offense the company is tried for.

Using content to train an LLM is not copying the content. I'm ignoring the silly "but actually" arguments about the content being in RAM so it's "copying". It's using the content to generate a statistical model of token (word-ish) relationships and probabilities. If you write content that is so original in it's wording and I train an LLM against it, then there is certainly the possibility that the LLM could be provoked the recall the exact words you used. You'd have to set the parameters just right to make it happen and I think that proper training would drastically lower if not remove that possible scenario. But even if it doesn't, the LLM doesn't have a copy of that original content. All it has is weights representing those relationship probabilities. Yes, the minutia is more complex, but that is the essence. If my LLM were to generate enough of this essentially verbatim unique content and I tried to publish or copyright it, then I as the user should be on the hook. But then you get into a discussion about how many words in a unique sequence does it take to be infringement?

Obviously, I am not a lawyer.

My summation in all of this is that new laws need to be put into place to handle this stuff because the existing ones are sufficiently non-definitive and/or ill-suited such that every party is forming strong opinions about how old laws apply to new situations and causing massive friction.

Re: Judge said Meta illegally used books to build its AI

#263
post #187
post #161

Earlier quoted context omitted.

Don't they mean that LLMs cannot perfectly reproduce the source material?

They're only stochastically lossy compression -- so sometimes it can.

Given a unique enough arrangement of words and a low enough entropy in token selection.

Re: Judge said Meta illegally used books to build its AI

#264
post #55

Let me make a clarifying statement since people confuse (purposely or just out of ignorance) what violating copyright for AI training can refer to: 1. Training AI on freely available copyright - Ambiguous legality, not really tested in court. AI doesn't actually directly copy the material it trains on, so it's not easy to make this ruling. 2. Circumventing payment to obtain copyright material for training - Unambiguo…

I'm not sure if Meta did anything illegal in 2. either. I thought the copyright infringement was by the people who provided the copyrighted material when they did not have the rights to do so. I may be wrong on this, but it would seem a reasonable protection for consumers in general. Meta is hardly an average consumer, but I doubt that matters in the case of the law. Having grounds to suspect that the provider did no…

Even if your belief that only the person *providing* the content is liable, do you honestly think a single person found all the content, downloaded it, directly trained the model themselves, and then deleted the content? If at any step the content was given or shared to anyone else for any reason, have they not converted into a provider themselves?

Re: Judge said Meta illegally used books to build its AI

#265

Earlier quoted context omitted.

>They can be modeled like that in the same way you can model humans as lossy compression algorithms Humans are totally capable of data compression. This will just devolved into a semantics game of what a data compressor is. LLMs were not developed to be, do not function as, and are not use as data compression utilities. Please, come knocking when a service provider exists that will use LLM's to compactly store your c…

> LLMs were not developed to be, do not function as, and are not use as data compression utilities. Again, from a information theoretic view point, this is exactly what they are doing, how they where developed and how they function. I don't know any serious researcher in ML that would find this claim even remotely controversial. It's really not just "a semantics game", its a part of a foundational understanding of th…

VectorDBs are not LLMs or SQL replacements and RAG is not data compression. Again this is just going to dwindle into semantics and where one draws boundaries. I can randomly remove bits from my HDD and call it compression. If you think humans are data compressors then I have no argument.

Can you get transformers to regurgitate information verbatim? Yes.

Would anyone in their right mind rely on a transformer to do so? No.

Would anyone in their right mind rely on a vectorDB to do so? No.

Would anyone in their right mind use a vectorDB/RAG/SQL/transformer combo to do so? Yes.

Is youtube going to drop VP9 for GeminiEncode to save google billions in bandwidth? No.

Re: Judge said Meta illegally used books to build its AI

#266

Earlier quoted context omitted.

The point is that it isn't compression. Its molding a plain structure iteratively into a ultra complex one. The model starts and ends at 20GB. It might have features that are reminiscent of compression or act like it, but under the hood there is nothing like zip, rar, H.265, or JPEG going on. And yes LLMs can recall exact material, but it is excerpts and fragments. There is statistical significance to it's ordering.…

OK, so we name it something different, you transform inputs into smaller outputs. If I make a script without AI that transforms someones poems without permission so sometimes it outputs the exact poems but sometimes it does it wrong, when is my script fair to use and when what I did is illegal. say my script contains words and matrixes of numbers the original poems are not directly inside, the script transformed them…

YouTube actively filters small channel copied content too.

We should be building robust copyright filters and everyone should be able to contribute their work to it.

but that is a different issue than whether or not an LLM is legally allowed to view a work that is publicly available.

Again, pretty much every artist is capable of off-hand copyright violation on the spot. This has been true forever. We don't bar them from seeing art to prevent this.

Re: Judge said Meta illegally used books to build its AI

#267

Earlier quoted context omitted.

> The RIAA lawyers never had to demonstrate that copying a DVD cratered the sales of their clients. They just got high penalties for infringers almost by default. The argument for 'fair use' in DVD copying/sharing is much weaker since the thing being shared in that case is a verbatim, digital copy of the work. 'Format shifting' is a tenuous argument, and it's pretty easily limited to making (and not distributing) per…

> there's clearly no copyright infringement in a human learning from someone's work and creating their own output, even if it "copies" an artist's style or draws inspiration from someone's plot-line. What do you mean here by "clearly?" This is not at all clear, and court cases have been decided in the opposite direction. This case: https://www.reuters.com/article/lifestyle/marvin-gaye-family... is as far from what yo…

But that case involves publishing of content. If you want to compare them, it seems like you'd argue that the content of Pharrell Williams' head is somehow copyright infringement. I've yet to see anyone credible say that an AI model outputting exact content and then someone publishing that should be allowed. If you manage to make an AI model output copyrighted content, you can't then claim it's not copyrighted. If you sit on one side of a table and read a book out loud and I write it down (with minor transcription errors), that content is almost certainly still copyrighted content and I could not distribute it legally.

Re: Judge said Meta illegally used books to build its AI

#268

Earlier quoted context omitted.

> They are in no way compression algorithms. I'm sorry, but this a fundamentally incorrect view of machine learning (including, but not limited to transformers). From an information theoretic perspective the two are essentially identical with the exception that standard compression algorithms do not have a proper "loss" function other than just trying to minimize reconstruction loss with the resulting compression siz…

>They can be modeled like that in the same way you can model humans as lossy compression algorithms Humans are totally capable of data compression. This will just devolved into a semantics game of what a data compressor is. LLMs were not developed to be, do not function as, and are not use as data compression utilities. Please, come knocking when a service provider exists that will use LLM's to compactly store your c…

> It has long been established that predictive models can be transformed into lossless compressors and vice versa. Incidentally, in recent years, the machine learning community has focused on training increasingly large and powerful self-supervised (language) models. Since these large language models exhibit impressive predictive capabilities, they are well-positioned to be strong compressors. In this work, we advocate for viewing the prediction problem through the lens of compression and evaluate the compression capabilities of large (foundation) models. We show that large language models are powerful general-purpose predictors and that the compression viewpoint provides novel insights into scaling laws, tokenization, and in-context learning. For example, Chinchilla 70B, while trained primarily on text, compresses ImageNet patches to 43.4% and LibriSpeech samples to 16.4% of their raw size, beating domain-specific compressors like PNG (58.5%) or FLAC (30.3%), respectively.

https://arxiv.org/pdf/2309.10668

Transformers are also used in the top algorithm right now on the Large Text Compression Benchmark. https://bellard.org/nncp/nncp.pdf

Re: Judge said Meta illegally used books to build its AI

#269

Earlier quoted context omitted.

Interesting figure that guy. Here's this: >Boies also was on the Theranos board of directors,[2][74] raising questions about conflicts of interest.[75] Boies agreed to be paid for his firm's work in Theranos stock, which he expected to grow dramatically in value.[75][3] https://en.wikipedia.org/wiki/David_Boies That was one of the decisions of all time.

He was also the primary villain of John Carreyrou's account of Theranos' rise and fall -- Bad Blood -- as his firm attempted to bully and hound whistleblowers, and intimidate their families with baseless legal threats. Not a very nice or ethical guy.

He was a secondary villain in Ronan Farrow's Catch and Kill book, as Harvey Weinstein's attorney. He employed similar tactics to protect Weinstein as he did for Theranos.

He was fired by the New York Times after all this came out, because he hired private investigators to go after the NYT reporters investigating Weinstein (Jodi Kantor & Megan Twohey at the NYT broke the story just before Farrow's reporting was published), while also representing the Times.

Re: Judge said Meta illegally used books to build its AI

#270
post #214

Earlier quoted context omitted.

> Except that humans don't make full copies to RAM, or disk or paper. Images on your retina form exact copies. They are scanned and translated into impulses that are then sent to a first set of "neural columns" - that's an exact copy. This is then connected to the visual cortex by the two most high bandwidth links in the human body ("the optical nerve", there's 2 of them of course, always wondered why everybody insis…

> that's an exact copy. Not for the purposes of copyright law. > is that humans don't have separate RAM [or disk] And that turns out to be incredibly important. Humans can't create a lasting, shareable copy of a copyrighted work by consuming it.

Sure they can. You can learn a copyrighted work by hard, even indirectly, then quickly duplicate it by hand. Mozart was originally famous for making a business out of that.
Post reply on HN