Live data from Hacker News

Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

understandingai.org

151–160 of 326 posts

Re: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

#151

Earlier quoted context omitted.

I think the argument is less about piracy and more that the model(s output) is a derivative work of Harry Potter, and the rights holder should be paid accordingly when it’s reproduced.

The main issue on an economical point of view is that copyright is not the framework we need for social justice and everyone florishing by enjoying pre-existing treasures of human heritage and fairly contributing back. There is no morale and justice ground to leverage on when the system is designed to create wealth bottleneck toward a few recipients. Harry Potter is a great piece of artistic work, and it's nice that…

Capitalism is allergic to second-order cybernetics.

First-order systems drive outcomes. "Did it make money?" "Did it increase engagement?" "Did it scale?" These are tight, local feedback loops. They work because they close quickly and map directly to incentives. But they also hide a deeper danger: they optimize without questioning what optimization does to the world that contains it.

Second-order cybernetics reason about systems. It doesn’t ask, "Did I succeed?" It asks, "What does it mean to define success this way?" "Is the goal worthy?"

That’s where capital breaks.

Capitalism is not simply incapable of reflection. In fact, it's structured to ignore it. It has no native interest in what emerges from its aggregated behaviors unless those emergent properties threaten the throughput of capital itself. It isn't designed to ask, "What kind of society results from a thousand locally rational decisions?" It asks, "Is this change going to make more or less money?"

It's like driving by watching only the fuel gauge. Not speed, not trajectory, or whether the destination is the right one. Just how efficiently you’re burning gas. The system is blind to everything but its goal. What looks like success in the short term can be, and often is, a long-term act of self-destruction.

Take copyright. Every individual rule, term length, exclusivity, royalty, can be justified. Each sounds fair on its own. But collectively, they produce extreme wealth concentration, barriers to creative participation, and a cultural hellscape. Not because anyone intended that, but because the emergent structure rewards enclosure over openness, hoarding over sharing, monopoly over multiplicity.

That’s not a bug. That's what systems do when you optimize only at the first-order level. And because capital evaluates systems solely by their extractive capacity, it treats this emergent behavior not as misalignment but as a feature. It canonizes the consequences.

A second-order system would account for the result by asking, "Is this the kind of world we want to live in?" It would recognize that wealth generated without regard to distribution warps everything it touches: art, technology, ecology, and relationships.

Capitalism, as it currently exists, is not wise. It does not grow in understanding. It does not self-correct toward justice. It self-replicates. Cleverly, efficiently, with brutal resilience. It's emergently misaligned and no one is powerful enough to stop it.

Re: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

#152
post #111
post #57

Earlier quoted context omitted.

Fair use is not a thing in every jurisdiction. In Germany for example there are cases where three words („wir sind Papst“) fall under copyright.

Germany does not have something called "fair use," but it does have provisions for uses that are fair . For example your use of the three words to talk about their copyrighted status is perfectly legal in Germany. That somebody wasn't allowed to use them in a specific way in the past doesn't mean that nobody is allowed to use them in any way.

Of course, but „it’s a short quote so you can use it“ is not true (at least in Germany).

Re: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

#153
post #75

Earlier quoted context omitted.

That may be relevant in the NYT vs OpenAI case, since NYT was supposedly able to reproduce entire articles in ChatGPT. Here Llama is predicting one sentence at a time when fed the previous one, with 50% accuracy, for 42% of the book. That can easily be written off as fair use.

That can easily be written off as fair use. No, it really couldn't. In fact, it's very persuasive evidence that Llama is straight up violating copyright. It would be one thing to be able to "predict" a paragraph or two. It's another thing entirely to be able to predict 42% of a book that is several hundred pages long.

Is it Llama violating the "copyright" or is it the researcher pushing it to do so?

Re: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

#154

Earlier quoted context omitted.

The main issue on an economical point of view is that copyright is not the framework we need for social justice and everyone florishing by enjoying pre-existing treasures of human heritage and fairly contributing back. There is no morale and justice ground to leverage on when the system is designed to create wealth bottleneck toward a few recipients. Harry Potter is a great piece of artistic work, and it's nice that…

Capitalism is allergic to second-order cybernetics. First-order systems drive outcomes. "Did it make money?" "Did it increase engagement?" "Did it scale?" These are tight, local feedback loops. They work because they close quickly and map directly to incentives. But they also hide a deeper danger: they optimize without questioning what optimization does to the world that contains it. Second-order cybernetics reason a…

[flagged]

Re: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

#155
post #37

On the other hand, it’s surprising that Llama memorized so much of Harry Potter and the Sorcerer's Stone. It's sold 120 million copies over 30 years. I've gotta think literally every passage is quoted online somewhere else a bunch of times. You could probably stitch together the full book quote-by-quote.

But also we know for a fact that Meta trained their models on pirated books. So there's no need to invent a hare brained scheme of stitching together bits and pieces like that.

No, assuming that just because it was in the training data it must be memorized is hare brained.

LLMs have limited capacity to memorize, under ~4 bits per parameter[1][2], and are trained on terabytes of data. It's physically impossible for them to memorize everything they're trained on. The model memorized chunks of Harry Potter not just because it was directly trained on the whole book, which the article also alludes to:

> For example, the researchers found that Llama 3.1 70B only memorized 0.13 percent of Sandman Slim, a 2009 novel by author Richard Kadrey. That’s a tiny fraction of the 42 percent figure for Harry Potter.

In case it isn't obvious, both Harry Potter and Sandman Slim are parts of books3 dataset.

[1] -- https://arxiv.org/abs/2505.24832 [2] -- https://arxiv.org/abs/2404.05405

Re: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

#156
post #64

As an experiment I searched Google for "harry potter and the sorcerer's stone text": - the first result is a pdf of the full book - the second result is a txt of the full book - the third result is a pdf of the complete harry potter collection - the fourth result is a txt of the full book (hosted on github funny enough) Further down there are similar copies from the internet archive and dozens of other sites. All in…

> let's not pretend that an LLM that autocompletes a couple lines from harry potter with 50% accuracy is some massive new avenue to piracy

No one is claiming this.

The corporations developing LLMs are doing so by sampling media without their owners' permission and arguing this is protected by US fair use laws, which is incorrect - as the late AI researcher Suchir Balaji explained in this other article:

https://suchir.net/fair_use.html

Re: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

#157
If LLMs are good at summarizing/compressing, what does this say about the underlying text? Why are some passages more easily recalled? Sure, some sections have probably been quoted more times than others, so there's bias in training data, which might explain why the Llama 1 and 3.1 images have similar peaks. Would this happen to LLMs even with no training bias?

Edit: seems the first part is about a memory about being bullied by Duddley. The second is where he's been elected to the quidditch team. Possibly they are just boring passages, compared to the surrounding ones. So probably just training bias.

Re: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

#158
post #156
post #64

As an experiment I searched Google for "harry potter and the sorcerer's stone text": - the first result is a pdf of the full book - the second result is a txt of the full book - the third result is a pdf of the complete harry potter collection - the fourth result is a txt of the full book (hosted on github funny enough) Further down there are similar copies from the internet archive and dozens of other sites. All in…

> let's not pretend that an LLM that autocompletes a couple lines from harry potter with 50% accuracy is some massive new avenue to piracy No one is claiming this. The corporations developing LLMs are doing so by sampling media without their owners' permission and arguing this is protected by US fair use laws, which is incorrect - as the late AI researcher Suchir Balaji explained in this other article: https://suchir…

Yeah, that's literally the title of the article,and the premise of the first paragraph.

Re: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

#159
post #109
post #104

Earlier quoted context omitted.

Of course it is. It's just a form of compression. If I train an autoencoder on an image, and distribute the weights, that would obviously be the same as distributing the content. Just because the content is commingled with lots of other content doesn't make it disappear. Besides, where did the sections of text from the input works that show up in the output text come from? Divine inspiration? God whispering to the ma…

[flagged]

> For one thing, they are probabilistic, so you wouldn't get the same content back every time like you would with a compression algorithm.

There is nothing inherently probabilistic in a neural network. The neural net always outputs the exact same value for the same input. We typically use that value in a larger program as a probability of a certain token, but that is not required to get data out. You could just as easily determinsitically take the output with the highest value, and add some extra rule for when multiple outputs have the exact same (e.g. pick the one from the output neuron with the lowest index).

Re: Meta's Llama 3.1 can recall 42 percent of the first Harry Potter book

#160
post #146
post #141

Earlier quoted context omitted.

> Freedom of speech doesn't give me the right to make copies of copyrighted works. No, but it gives you the right to quote a line from a movie or TV show without being charged with copyright infringement. You argued that an LLM deserves that same right, even if you didn't realize it. > than so do the analogous weights (reinforced neural pathways) in your brain Did your brain consume millions of copyrighted books in o…

Millions? No, but my brain certainly consumed thousands of books, movies, TV shows, pieces of music, artworks, and other copyrighted material. Where is the cutoff? Can I only consume 999,999 copyrighted works before I'm not longer allowed to remember something without infringing copyright? My brain definitely would not exist in its current form without consuming that material. It would exist in some form, but it woul…

> just because it turns out their work was so trite and predictable it could be reproduced by simply guessing the next most likely token.

Well, if you have no idea how LLMs work, you could've just said so.

Post reply on HN