5 ms·

There are two important differences. Google Books retains the bibliographical information so you can properly cite the authors or contact them for permission t

by nequo 4y ago

There are two important differences. Google Books retains the bibliographical information so you can properly cite the authors or contact them for permission to use their material. And Google Books does not automatically write new books for you that you can then send off to Penguin Books or self-publish on Amazon.

naasking 4y ago
Copilot also isn't retaining the actual content of the source code repositories and then deriving works from that. If I wrote a giant table of token frequencies and associative keywords by analyzing a bunch of source, and sold that to people as a "github code analysis" book, I'm pretty sure that's perfectly fine because it's not a derivative work. I'm not sure that the fact that a program can then take that associative data and generate new code doesn't suddenly make it not ok.
- nequo 4y ago
  Yeah, I guess the court will be the real test of what is allowed here.
- TeMPOraL 4y ago
  > If I wrote a giant table of token frequencies and associative keywords by analyzing a bunch of source, and sold that to people as a "github code analysis" book, I'm pretty sure that's perfectly fine because it's not a derivative work. That sounds to me somewhat close to "if I take an FFT of each of those copyrighted images, glue them together, and sell this as a picture, is that a derivative work?" - I'd say yes, or perhaps even a different encoding of the original work, since you can reverse the frequency domain representation and get the original spatial representation - the original images - back.
  - naasking 4y ago
    Sure, ROT13 encoding is a derivative work because the entire original work is still there, encoded. Ditto for FFT. Large language models are not that. Sometimes parts of the original works are still encoded, which we've seen when some code is reproduced verbatim, and I'm sure that happens to people as well, ie. they see some algorithm and down the road have to write something similar and end up reproducing the exact same thing. Once they iron out those wrinkles, it's not clear to me that a large language model is a directly reversible function of the original works. At least, not any more than a human learning from reading a bunch of code and then going on to have a career selling his skills at writing code. Edit: by which I mean, LLMs are lossy encodings, not lossless encodings.
    - TeMPOraL 4y ago
      > Ditto for FFT. Large language models are not that. They're not, but the "giant table of token frequencies and associative keywords" reminded me of doing FFT on images, and I wanted to communicate the idea that transformations like this can actually retain the original information, and reproduce it back through inverse transform. > by which I mean, LLMs are lossy encodings, not lossless encodings Exactly. And while I doubt most training data is recoverable, "lossy encoding" is still a spectrum. As you move away from lossless, it's not obvious when, or if at all, the result is clear from copyright of original inputs' author. Compare e.g. with JPEG, which employs a less sophisticated lossy encoding - no matter how hard you compress a source image, the result would still likely retain the copyright of the source image author, as provenance matters. (IANAL, though.)
      - naasking 4y ago
        > And while I doubt most training data is recoverable, "lossy encoding" is still a spectrum. [...] Compare e.g. with JPEG I'll just finally note that LLMs are not lossy encodings in the same sense as JPEG. LLMs are closer to human-like learning, where learning from data enables us to create entirely new expressions of the same concepts contained in that data, rather than acting as pure functions of the source data. That's why this will be interesting to see play out in the courts.
        TeMPOraL 4y ago
        My belief is there is no fundamental difference here. That is, learning is a form of compression. Learning concepts is just a more complex form of achieving that much greater (if lossy) compression levels. If the courts will see it the same way too, things will get truly interesting.
        naasking 4y ago
        Yes learning concepts is a form of compression, but I'm not sure that implies there's no "fundamental" difference. I see it as akin to a programming language having only first-order functions vs. having higher-order functions. Higher-order functions give you more expressive power but not any more computational power. You could say a higher order program can "just" be transformed into a first-order program via defunctionalization, but I think the expressive difference is in and of itself meaningful. I hope the courts can tease that out in the end, and we'll see if LLMs cross that line, or if we need something even more general to qualify.
  - robertlagrant 4y ago
    > since you can reverse the frequency domain representation and get the original spatial representation - the original images - back I'd have thought that's exactly what you can't do with CoPilot.
  - Dylan16807 4y ago
    What do you mean by "glue them together"? If you overlap a hundred different FFTs, then the result is likely fine copyright-wise. These networks are not [supposed to] contain much of the original data. Like the trivia point that Stable Diffusion has less than two bytes per source image, on average.
    - TeMPOraL 4y ago
      > What do you mean by "glue them together"? Stitch them side by side. Yes, this is not how those DNNs work, but the example was more about highlighting that "a giant table of token frequencies" by itself is probably reversible back to original data, or at least something resembling it. > Stable Diffusion has less than two bytes per source image, on average. I'm not convinced by this trivia point, though. Stable Diffusion is, effectively, a lossy compression of the training data. Nothing says lossy compression algorithms can't exploit some higher-level conceptual structures in the inputs[0], and applying lossy compression to some work doesn't automatically erase the copyrights of the original input's author. -- [0] - SD isn't compressing arbitrary byte sequences, it's compressing images - which is a small subset of all possible byte sequences as large as the largest image used in training. "Less than two bytes per source image, on average" doesn't sound to me like something implausible for a lossy compressor that is focused on such small subset of possible inputs, and gets to exploit high-level patterns in such data.
      - Dylan16807 4y ago
        > the example was more about highlighting that "a giant table of token frequencies" by itself is probably reversible back to original data, or at least something resembling it That depends entirely on how many frequencies you're keeping. > high-level patterns in such data High level patterns across thousands of images are generally not copyrightable. I might even describe the purpose of stable diffusion as extracting just the patterns and zero specifics.
      - cbigart 4y ago
        >"Less than two bytes per source image, on average" doesn't sound to me like something implausible for a lossy compressor that is focused on such small subset of possible inputs, and gets to exploit high-level patterns in such data. Two bytes would only let you uniquely identify ~65k images though, which to me doesn't sound plausible for a lossy compressor.

by nequo 4y ago

naasking 4y ago

Copilot also isn't retaining the actual content of the source code repositories and then deriving works from that. If I wrote a giant table of token frequencies and associative keywords by analyzing a bunch of source, and sold that to people as a "github code analysis" book, I'm pretty sure that's perfectly fine because it's not a derivative work. I'm not sure that the fact that a program can then take that associative data and generate new code doesn't suddenly make it not ok.

nequo 4y ago
Yeah, I guess the court will be the real test of what is allowed here.
TeMPOraL 4y ago
> If I wrote a giant table of token frequencies and associative keywords by analyzing a bunch of source, and sold that to people as a "github code analysis" book, I'm pretty sure that's perfectly fine because it's not a derivative work. That sounds to me somewhat close to "if I take an FFT of each of those copyrighted images, glue them together, and sell this as a picture, is that a derivative work?" - I'd say yes, or perhaps even a different encoding of the original work, since you can reverse the frequency domain representation and get the original spatial representation - the original images - back.
- naasking 4y ago
  Sure, ROT13 encoding is a derivative work because the entire original work is still there, encoded. Ditto for FFT. Large language models are not that. Sometimes parts of the original works are still encoded, which we've seen when some code is reproduced verbatim, and I'm sure that happens to people as well, ie. they see some algorithm and down the road have to write something similar and end up reproducing the exact same thing. Once they iron out those wrinkles, it's not clear to me that a large language model is a directly reversible function of the original works. At least, not any more than a human learning from reading a bunch of code and then going on to have a career selling his skills at writing code. Edit: by which I mean, LLMs are lossy encodings, not lossless encodings.
  - TeMPOraL 4y ago
    > Ditto for FFT. Large language models are not that. They're not, but the "giant table of token frequencies and associative keywords" reminded me of doing FFT on images, and I wanted to communicate the idea that transformations like this can actually retain the original information, and reproduce it back through inverse transform. > by which I mean, LLMs are lossy encodings, not lossless encodings Exactly. And while I doubt most training data is recoverable, "lossy encoding" is still a spectrum. As you move away from lossless, it's not obvious when, or if at all, the result is clear from copyright of original inputs' author. Compare e.g. with JPEG, which employs a less sophisticated lossy encoding - no matter how hard you compress a source image, the result would still likely retain the copyright of the source image author, as provenance matters. (IANAL, though.)
    - naasking 4y ago
      > And while I doubt most training data is recoverable, "lossy encoding" is still a spectrum. [...] Compare e.g. with JPEG I'll just finally note that LLMs are not lossy encodings in the same sense as JPEG. LLMs are closer to human-like learning, where learning from data enables us to create entirely new expressions of the same concepts contained in that data, rather than acting as pure functions of the source data. That's why this will be interesting to see play out in the courts.
      - TeMPOraL 4y ago
        My belief is there is no fundamental difference here. That is, learning is a form of compression. Learning concepts is just a more complex form of achieving that much greater (if lossy) compression levels. If the courts will see it the same way too, things will get truly interesting.
        naasking 4y ago
        Yes learning concepts is a form of compression, but I'm not sure that implies there's no "fundamental" difference. I see it as akin to a programming language having only first-order functions vs. having higher-order functions. Higher-order functions give you more expressive power but not any more computational power. You could say a higher order program can "just" be transformed into a first-order program via defunctionalization, but I think the expressive difference is in and of itself meaningful. I hope the courts can tease that out in the end, and we'll see if LLMs cross that line, or if we need something even more general to qualify.
- robertlagrant 4y ago
  > since you can reverse the frequency domain representation and get the original spatial representation - the original images - back I'd have thought that's exactly what you can't do with CoPilot.
- Dylan16807 4y ago
  What do you mean by "glue them together"? If you overlap a hundred different FFTs, then the result is likely fine copyright-wise. These networks are not [supposed to] contain much of the original data. Like the trivia point that Stable Diffusion has less than two bytes per source image, on average.
  - TeMPOraL 4y ago
    > What do you mean by "glue them together"? Stitch them side by side. Yes, this is not how those DNNs work, but the example was more about highlighting that "a giant table of token frequencies" by itself is probably reversible back to original data, or at least something resembling it. > Stable Diffusion has less than two bytes per source image, on average. I'm not convinced by this trivia point, though. Stable Diffusion is, effectively, a lossy compression of the training data. Nothing says lossy compression algorithms can't exploit some higher-level conceptual structures in the inputs[0], and applying lossy compression to some work doesn't automatically erase the copyrights of the original input's author. -- [0] - SD isn't compressing arbitrary byte sequences, it's compressing images - which is a small subset of all possible byte sequences as large as the largest image used in training. "Less than two bytes per source image, on average" doesn't sound to me like something implausible for a lossy compressor that is focused on such small subset of possible inputs, and gets to exploit high-level patterns in such data.
    - Dylan16807 4y ago
      > the example was more about highlighting that "a giant table of token frequencies" by itself is probably reversible back to original data, or at least something resembling it That depends entirely on how many frequencies you're keeping. > high-level patterns in such data High level patterns across thousands of images are generally not copyrightable. I might even describe the purpose of stable diffusion as extracting just the patterns and zero specifics.
    - cbigart 4y ago
      >"Less than two bytes per source image, on average" doesn't sound to me like something implausible for a lossy compressor that is focused on such small subset of possible inputs, and gets to exploit high-level patterns in such data. Two bytes would only let you uniquely identify ~65k images though, which to me doesn't sound plausible for a lossy compressor.