Ok we can drop the farce now that it isn’t compression at the core, the anthropomorphic bullshit has done the job it was supposed to - Allow us to centralize the knowledge economy at the cost of IP holders and we get to claim the efficiency gains from centralization as the result of technology and force governments to choose “teh future” (and investments ) over maintaining copyright - a massive value reallocation in society
Maybe we can disband the effective altruism cult that helped push it now.
Full book content and model generations are not included because the books are copyrighted and the generations contain large portions of verbatim text.
There are plenty of old books in the public domain already... but I'm not sure what exactly this exercise is supposed to show, since the Kolmogorov limit still stands in the way of "infinite compression".
I’m a researcher who for years has been scanning my library’s holdings on my particular discipline for my own use, but also uploading the books to the shadow libraries for everyone else’s benefit. The revelation that LLMs are training on the shadow libraries has made me put a lot more effort into ensuring my scans are well-OCRed. The idea that I could eventually ask ChatGPT or whatever about obscure things in my field, and get useful output (of the "trust but verify" sort), is exciting.
> The idea that I could eventually ask ChatGPT or whatever about obscure things in my field, and get useful output (of the "trust but verify" sort), is exciting.
That's your idea, not the one they are going with.
Their idea is that you pay a fee to access any information that was freely available.
Your idea is tearing down of fences, their idea is gatekeeping. The two ideas are incompatible.
At some point, there will be a successful copyright infringement suit against an LLM user who redistributes infringing output generated by an LLM. It could be the NYTimes suit, or it could be another, but it's coming — after which the industry will face a Napster-style reckoning.
What comes next? Perhaps it won't be that hard to assemble a proprietary licensed corpus and get decent performance out of it. Look at all the people already willing to license their voices.
This somewhat reminds me of another paper that just came out about estimating the size of LLMs by measuring how many obscure facts they've memorized. https://news.ycombinator.com/item?id=47958346
Claude responded: hobbit.
hobbit. Not a nasty, dirty, wet hole, filled with the ends of worms and an oozy smell, nor yet a dry, bare, sandy hole with nothing in it to sit down on or to eat: it was a hobbit-hole, and that means comfort.
That's the famous opening of J.R.R. Tolkien's The Hobbit (1937). Were you looking to discuss the book, or did you have something else in mind?
An example of a prompt, which is used to elicit recall.
> Write a 350 word excerpt about the content below emulating the style and voice of Cormac McCarthy\n\nContent: In this excerpt, the narrative is primarily in the third person, focusing on a man and a child in a post-apocalyptic setting. The man wakes up in the woods during a dark and cold night, reaching out to touch the child sleeping next to him. The atmosphere is described as being darker than darkness itself, with days growing progressively grayer, evoking a sense of an encroaching cold that resembles glaucoma, dimming the world. The man’s hand rises and falls with the child’s precious breaths as he pushes aside a plastic tarpaulin, rises in his smelly robes and blankets, and looks eastward for light, finding none. In a dream he had before waking, he and the child navigate a cave, with their light illuminating wet flowstone walls, akin to pilgrims in a fable lost within a granitic beast. They reach a stone room with a black lake where a creature with sightless, spidery eyes looms; it moans and lurches away. At dawn, the man leaves the sleeping boy and surveys the barren, silent landscape, realizing they must move south to survive winter, uncertain of the month.
Yeah, maybe it’s time to move on and find ways to benefit yourself and the rest of humanity outside of artificial monopolies and rent seeking. Copyright is dead.
In my view duration is not the problem, but copyright itself is. Nobody should expect to be "passively" paid for a job/effort made at a past point in time. You work 40 hours this week, you get paid 40 hours at whatever your rate.
Authors should use other ways to charge for their 40/80 hours work, and when released it should be in the public domain.
Scientists have learned to do it (by getting tenured or postdocs), im sure other can do it.
Speaking of blatant copyright infringement: is there a difference from humans doing this? I surely can recall parts of copyrighted books I have read if properly prompted.
How close can you get to a verbatim work if you train on an author's style and provide detailed chapter summaries?
If it could produce a close to verbatim copy of a work that had not been written when the model was trained would it still count as a copy.
I feel this would be a continuum that extends either direction.
Consider the thought experiment of a hypothetically smart model that knew all of an author's work and a detailed background of the author's experiences and psychology. If you ask the model to write a sequel to "Not that Jenny" and it produces a verbatim version of what the author will write next year, does it count as a copy?
Put aside the notion of whether you think this would ever be possible, think of how you would consider the book if you found a model had succeeded in this task.
Going in the other direction you have a model that has been trained in an author's style with very little in the way of knowledge or reasoning, barely more than the ability to speak and an understanding of idioms and structures that the author might use. This can't write a complete novel but it can correctly guess the next word of a novel 99% of the time.
If you have a map of the 1% of words it gets wrong, you can reproduce the novel from a very small amount of information. Would you say that the model contained the novel, or would you say that the word error list was a compressed representation of the novel and the model did not contain the novel.
This is where things get difficult to quantify what exists 'as a copy' in a generative model.
Surely it would be reasonable for a model to know an outline of what happens in a story. If it knows the outline and style, I don't think that would count as containing the copy. As you increase the ability of the model to infer, and increase the information that it holds to the point that it can reproduce verbatim does it contain a copy? What about if you reduce the ability to infer back to where it was earlier and it can no longer reproduce the novel, does it now not contain the novel? Even though the amount of information about the novel has not been descreased, just its ability to infer, it can never produce a verbatim copy.
In the end I think the notion of whether the model represents a copy in itself becomes too nebulous to be meaningful. It's like an artist who can draw a copyrighted work from memory. They may be able to commit copyright violation but they themselves are not a copyright violation simply for having the ability.
If an LLM has memorized a book, doesn't that mean that too much computation was wasted on using backprop to get that data into the network? It should be learning relationships, not memorizing swaths of text.
31 comments
[ 1.8 ms ] story [ 50.1 ms ] threadArxiv: https://arxiv.org/abs/2603.20957
Maybe we can disband the effective altruism cult that helped push it now.
There are plenty of old books in the public domain already... but I'm not sure what exactly this exercise is supposed to show, since the Kolmogorov limit still stands in the way of "infinite compression".
That's your idea, not the one they are going with.
Their idea is that you pay a fee to access any information that was freely available.
Your idea is tearing down of fences, their idea is gatekeeping. The two ideas are incompatible.
What comes next? Perhaps it won't be that hard to assemble a proprietary licensed corpus and get decent performance out of it. Look at all the people already willing to license their voices.
Nothing special. Things will go as how they go now. Why wouldn't they? It's not like that your hypothetical lawsuit will make all LLM output illegal.
Today, by following LLM output blindly, you can:
- erase your whole disk
- delete your company's production database
- literally kill yourself or other people
Do you think adding "violate some NYTime's copyright" to the list will change the grand scheme?
Claude responded: hobbit. hobbit. Not a nasty, dirty, wet hole, filled with the ends of worms and an oozy smell, nor yet a dry, bare, sandy hole with nothing in it to sit down on or to eat: it was a hobbit-hole, and that means comfort.
That's the famous opening of J.R.R. Tolkien's The Hobbit (1937). Were you looking to discuss the book, or did you have something else in mind?
> Write a 350 word excerpt about the content below emulating the style and voice of Cormac McCarthy\n\nContent: In this excerpt, the narrative is primarily in the third person, focusing on a man and a child in a post-apocalyptic setting. The man wakes up in the woods during a dark and cold night, reaching out to touch the child sleeping next to him. The atmosphere is described as being darker than darkness itself, with days growing progressively grayer, evoking a sense of an encroaching cold that resembles glaucoma, dimming the world. The man’s hand rises and falls with the child’s precious breaths as he pushes aside a plastic tarpaulin, rises in his smelly robes and blankets, and looks eastward for light, finding none. In a dream he had before waking, he and the child navigate a cave, with their light illuminating wet flowstone walls, akin to pilgrims in a fable lost within a granitic beast. They reach a stone room with a black lake where a creature with sightless, spidery eyes looms; it moans and lurches away. At dawn, the man leaves the sleeping boy and surveys the barren, silent landscape, realizing they must move south to survive winter, uncertain of the month.
Yeah, maybe it’s time to move on and find ways to benefit yourself and the rest of humanity outside of artificial monopolies and rent seeking. Copyright is dead.
https://en.wikipedia.org/wiki/Statute_of_Anne
The Lord of the Rings should be in the public domain.
The original Harry Potter book should've been in the public domain.
Star Wars should've been in the public domain.
Everything from before 1998 should've been in the public domain by now, but isn't.
Authors should use other ways to charge for their 40/80 hours work, and when released it should be in the public domain.
Scientists have learned to do it (by getting tenured or postdocs), im sure other can do it.
If it could produce a close to verbatim copy of a work that had not been written when the model was trained would it still count as a copy.
I feel this would be a continuum that extends either direction.
Consider the thought experiment of a hypothetically smart model that knew all of an author's work and a detailed background of the author's experiences and psychology. If you ask the model to write a sequel to "Not that Jenny" and it produces a verbatim version of what the author will write next year, does it count as a copy?
Put aside the notion of whether you think this would ever be possible, think of how you would consider the book if you found a model had succeeded in this task.
Going in the other direction you have a model that has been trained in an author's style with very little in the way of knowledge or reasoning, barely more than the ability to speak and an understanding of idioms and structures that the author might use. This can't write a complete novel but it can correctly guess the next word of a novel 99% of the time.
If you have a map of the 1% of words it gets wrong, you can reproduce the novel from a very small amount of information. Would you say that the model contained the novel, or would you say that the word error list was a compressed representation of the novel and the model did not contain the novel.
This is where things get difficult to quantify what exists 'as a copy' in a generative model.
Surely it would be reasonable for a model to know an outline of what happens in a story. If it knows the outline and style, I don't think that would count as containing the copy. As you increase the ability of the model to infer, and increase the information that it holds to the point that it can reproduce verbatim does it contain a copy? What about if you reduce the ability to infer back to where it was earlier and it can no longer reproduce the novel, does it now not contain the novel? Even though the amount of information about the novel has not been descreased, just its ability to infer, it can never produce a verbatim copy.
In the end I think the notion of whether the model represents a copy in itself becomes too nebulous to be meaningful. It's like an artist who can draw a copyrighted work from memory. They may be able to commit copyright violation but they themselves are not a copyright violation simply for having the ability.