2 ms·

An AI LLM, on the other hand, is (unless some court or law changes this) a derivative work of its training data. because an AI LLM is in fact copyrightable and

An AI LLM, on the other hand, is (unless some court or law changes this) a derivative work of its training data. because an AI LLM is in fact copyrightable and self-evidently a derivative work. I mean, that's a fine opinion to hold, and you might be right. But so far all you've done is repeat yourself and appeal to "self-evident" which isn't a terribly strong argument. I'll wait for some actual precedent / case-law to solidify my own opinion. As it stands, I can see both sides of the argument, but I don't think the conclusion is as obvious as some folks in this discussion seem to find it. shrug an AI LLM can easily recite its training data, providing further evidence that the AI LLM is a derivative work of that data. OK, I can buy that to a point, so far as arguing that the LLM itself is a derivative work. But I'm not convinced that, in turn, the output of the LLM is also a derivative work in those cases where what it returns is not an exact copy (or even nearly exact copy) of anything in the training corpus.

JoshTriplett 3y ago
Clarifying: the part I'm arguing is "self-evident" is that an LLM is a derivative work of its training data, in the same sense that if you copy the text of a million books into a data file and compress that file reversibly in a way that lets you get most or all of them back out again, the result is clearly a derivative work of those books. That part I made a case for, and it seems like from your last paragraph you agree with that part of the argument. (By contrast, I wouldn't say it's self-evident that a database of blake3 hashes would be a derivative work (leaving aside that it'd probably be fair use), nor is it self-evident that compiling a million books into a Bloom filter that can recognize any random sentence but not output any random sentence would make the Bloom filter a derivative work. I think the unfiltered LLM being able to output near-verbatim copies of parts of the training set makes that case evident.) I agree that the second step, of the output of the LLM being a derivative work of the LLM, is less obvious. And I agree that it's going to take case law before people are sure of the answer to that part. I hope the answer is "yes", and I think it'd do substantial harm to Open Source if the answer is a definitive "no".
- mindcrime 3y ago
  Fair enough. I think the distinction between "the model weights" and "the output of the model" was a little blurred when this first started. Sounds like we're closer to "in agreement" than not for the most part.

JoshTriplett 3y ago

Clarifying: the part I'm arguing is "self-evident" is that an LLM is a derivative work of its training data, in the same sense that if you copy the text of a million books into a data file and compress that file reversibly in a way that lets you get most or all of them back out again, the result is clearly a derivative work of those books. That part I made a case for, and it seems like from your last paragraph you agree with that part of the argument. (By contrast, I wouldn't say it's self-evident that a database of blake3 hashes would be a derivative work (leaving aside that it'd probably be fair use), nor is it self-evident that compiling a million books into a Bloom filter that can recognize any random sentence but not output any random sentence would make the Bloom filter a derivative work. I think the unfiltered LLM being able to output near-verbatim copies of parts of the training set makes that case evident.) I agree that the second step, of the output of the LLM being a derivative work of the LLM, is less obvious. And I agree that it's going to take case law before people are sure of the answer to that part. I hope the answer is "yes", and I think it'd do substantial harm to Open Source if the answer is a definitive "no".

mindcrime 3y ago
Fair enough. I think the distinction between "the model weights" and "the output of the model" was a little blurred when this first started. Sounds like we're closer to "in agreement" than not for the most part.