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

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by paxys on Monday June 16, 2025 @12:01AM (#44286511)

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.

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by gpm on Monday June 16, 2025 @12:13AM (#44286558)

I'm pretty sure books.google.com does the exact same with much better reliability... and the US courts found that to be fair use. (Agreeing with parent comment)

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by pclmulqdq on Monday June 16, 2025 @12:20AM (#44286596)

If there is a circuit split between it and NYT vs OAI, the Google Books ruling (in the famously tech-friendly ninth circuit) may also find itself under review.

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by echelon on Monday June 16, 2025 @12:26AM (#44286612)

> 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.

Is that fair use, or is that compression of the verbatim source?

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by TeMPOraL on Monday June 16, 2025 @12:07PM (#44290886)

It doesn't let you recover the text without knowing it in advance, so no.

You can't in particular iterate it sentence by sentence; you're unlikely to go past sentence 2 this way before it starts giving you back it's own ideas.

The whole thing is a sleigh of hand, basically. There's 42% of the book there, in tiny pieces, which you can only identify if you know what you're looking for. The model itself does not.

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by gamblor956 on Monday June 16, 2025 @01:50AM (#44286928)

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.

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by reedciccio on Monday June 16, 2025 @02:01AM (#44286984)

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

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by lern_too_spel on Monday June 16, 2025 @03:28AM (#44287349)

If you distribute a zip file of the book, are you violating copyright, or is it the person who unzips it?

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by TeMPOraL on Monday June 16, 2025 @07:46AM (#44288679)

If you walk through the N-gram database with a copy of Harry Potter in hand and observe that for N=7, you can find any piece of it in the database with above-average frequency, does that mean N-gram database is violating copyright?

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by gamblor956 on Monday June 16, 2025 @12:02PM (#44290845)

If the database is sharing those pieces, it might be yes.

Copyright takes into account the use for such the copying is done. Commercial use will almost always be treated as not fair use, with limited exceptions.

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by TeMPOraL on Monday June 16, 2025 @12:32PM (#44291081)

I'd say no, because you can't reasonably access and order those pieces without already having the work at your side to use as a reference.

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by lern_too_spel on Wednesday June 18, 2025 @07:36PM (#44314168)

Not unless you can reproduce large portions of Harry Potter verbatim from the database. If the 7-grams are taken only from Harry Potter, that is very likely.

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by gamblor956 on Monday June 16, 2025 @12:00PM (#44290838)

You are.

The creation of the unzipped file is not treated as a separate copy so the recipient would not be violating copyright just by unzipping the file you provided.

[ Parent | on HN ]

Re:Meta's Llama 3.1 can recall 42 percent of the first Harry... by int_19h on Tuesday June 24, 2025 @07:14PM (#44372087)

If it can predict the next sentence reliably, that sentence then becomes part of the context, so if you just continue inference, it would eventually produce the entire text verbatim, no?

[ Parent | on HN ]