The language model isn’t teaching anything it is changing the wording of something and spitting it back out. And in some cases, not changing the wording at all, just spitting the information back out, without paying the copyright source.

You could honestly say the same about most "teaching" that a student without a real comprehension of the subject does for another student. But ultimately, that's beside the point. Because changing the wording, structure, and presentation is all that is necessary to avoid copyright violation. You cannot copyright the information. Only a specific expression of it.

There's no special exception for AI here. That's how copyright works for you, me, the student, and the AI. And if you're hoping that copyright is going to save you from the outcomes you're worried about, it won't.

Good luck breaking down people's doors for scanning their own physical books for their personal use when analog media has no DRM and can't phone home, and paper books are an analog medium.

That would be like kicking down people's doors for needle-dropping their LPs to FLAC for their own use and to preserve the physical records as vinyl wears down every time it's played back.

That's a reasonable critique.

The point is that it's trivial to come up with new words. Put that same prompt into a bunch of different LLMs and you'll get a bunch of different words. Some of them may exist somewhere that don't exist. There are simple rules for combining words that are so simple that children play them as games.

The LLM doesn't actually even recognize "words" it recognizes tokens which are typically parts of words. It usually avoids random combinations of those but you can easily get it to do so, if you want.

A court will decide such cases. Most AI models aren't trained for this purpose of whitewashing content even if some people would imply that's all they do, but if you decided to actually train a model for this explicit purpose you would most likely not get away with it if someone dragged you in front of a court for it.

It's a similar defense that some file hosting websites had against hosting and distributing copyrighted content (Eg. MEGA), but in such cases it was very clear to what their real goals were (especially in court), and at the same time it did not kill all file sharing websites, because not all of them were built with the intention to distribute illegal material with under the guise of legitimate operation.

This so very much. I've been saying it since 2020. People who think the big corporations (even the ones that use AI), aren't playing both sides of this issue from the very beginning just aren't paying attention.

It's in their interest to have those positive to AI defend them by association by energizing those negative to AI to take on an "us vs them" mentality, and the other way around as well. It's the classic divide and conquer.

Because if people refuse to talk to each other about it in good faith, and refuse to treat each other with respect, learn where they're coming from or why they hold such opinions, you can keep them fighting amongst themselves, instead of banding together and demanding realistic, and fair policies in regards to AI. This is why bad faith arguments and positions must be shot down on both the side you agree with and the one you disagree with.

I've hand calculated forward propagation (neural networks). AI does not learn, its statically optimized. AI "learning" is curve fitting. Human learning requires understanding, which AI is not capable of.

Google search results aren't deterministic but I wouldn't say it "learns" like a person. Algorithms with pattern detection isn't the same as human learning.

You may be correct but we don't really know how humans learn.

There's a ton of research on it and a lot of theories but no clear answers.
There's general agreement that the brain is a bunch of neurons; there are no convincing ideas on how consciousness arises from that mass of neurons.
The brain also has a bunch of chemicals that affect neural processing; there are no convincing ideas on how that gets you consciousness either.

We modeled perceptrons after neurons and we've been working to make them more like neurons. They don't have any obvious capabilities that perceptrons don't have.

That's the big problem with any claim that "AI doesn't do X like a person"; since we don't know how people do it we can neither verify nor refute that claim.

There's more to AI than just being non-deterministic. Anything that's too deterministic definitely isn't an intelligence though; natural or artificial. Video compression algorithms are definitely very far removed from AI.

Hehe jailbreak an Android OS. You mean “rooting”.

Judge,I'm pirating them to train ai not to consume for my own personal use.

The ruling explicitly says that scanning books and keeping/using those digital copies is legal.

The piracy found to be illegal was downloading unauthorized copies of books from the internet for free.

It sounds like transferring an owned print book to digital and using it to train AI was deemed permissable. But downloading a book from the Internet and using it was training data is not allowed, even if you later purchase the pirated book. So, no one will be knocking down your door for scanning your books.

This does raise an interesting case where libraries could end up training and distributing public domain AI models.

Oh, so it would be the news?

I wonder if the archive.org cases had any bearing on the decision.

You can train an LLM to generate 1:1 copies

when not in use i have it load images from my local webserver that are generated by some scripts and feature local news or the weather. kindle screensaver sucks.

One point I would refute here is determinism. AI models are, by default, deterministic. They are made from deterministic parts and "any combination of deterministic components will result in a deterministic system". Randomness has to be externally injected into e.g. current LLMs to produce 'non-deterministic' output.

There is the notable exception of newer models like ChatGPT4 which seemingly produces non-deterministic outputs (i.e. give it the same sentence and it produces different outputs even with its temperature set to 0) - but my understanding is this is due to floating point number inaccuracies which lead to different token selection and thus a function of our current processor architectures and not inherent in the model itself.

Billions of dollars, and they can't afford to buy ebooks?

It took me a few days to get the time to read the actual court ruling but here's the basics of what it ruled (and what it didn't rule on):

• It's legal to scan physical books you already own and keep a digital library of those scanned books, even if the copyright holder didn't give permission. And even if you bought the books used, for very cheap, in bulk.
• It's legal to keep all the book data in an internal database for use within the company, as a central library of works accessible only within the company.
• It's legal to prepare those digital copies for potential use as training material for LLMs, including recognizing the text, performing cleanup on scanning/recognition errors, categorizing and cataloguing them to make editorial decisions on which works to include in which training sets, tokenizing them for the actual LLM technology, etc. This remains legal even for the copies that are excluded from training for whatever reason, as the entire bulk process may involve text that ends up not being used, but the process itself is fair use.
• It's legal to use that book text to create large language models that power services that are commercially sold to the public, as long as there are safeguards that prevent the LLMs from publishing large portions of a single copyrighted work without the copyright holder's permission.
• It's illegal to download unauthorized copies of copyrighted books from the internet, without the copyright holder's permission.

Here's what it didn't rule on:

• Is it legal to distribute large chunks of copyrighted text through one of these LLMs, such as when a user asks a chatbot to recite an entire copyrighted work that is in its training set? (The opinion suggests that it probably isn't legal, and relies heavily on the dividing line of how Google Books does it, by scanning and analyzing an entire copyrighted work but blocking users from retrieving more than a few snippets from those works).
• Is it legal to give anyone outside the company access to the digitized central library assembled by the company from printed copies?
• Is it legal to crawl publicly available digital data to build a library from text already digitized by someone else? (The answer may matter depending on whether there is an authorized method for obtaining that data, or whether the copyright holder refuses to license that copying).

So it's a pretty important ruling, in my opinion. It's a clear green light to the idea of digitizing and archiving copyrighted works without the copyright holder's permission, as long as you first own a legal copy in the first place. And it's a green light to using copyrighted works for training AI models, as long as you compiled that database of copyrighted works in a legal way.

The court's ruling explicitly depended on the fact that Anthropic does not allow users to retrieve significant chunks of copyrighted text. It used the entire copyrighted work to train the weights of the LLMs, but is configured not to actually copy those works out to the public user. The ruling says that if the copyright holders later develop evidence that it is possible to retrieve entire copyrighted works, or significant portions of a work, then they will have the right sue over those facts.

But the facts before the court were that Anthropic's LLMs have safeguards against distributing copies of identifiable copyrighted works to its users.