4 ms·

The question of whether training is a copyright violation is still an open one. I think Japan is in the right here. IMO, training an AI is analogous to reading

The question of whether training is a copyright violation is still an open one. I think Japan is in the right here. IMO, training an AI is analogous to reading a book. As long as the AI isn’t regurgitating the contents 1:1, it’s not infringement.

dorkwood 2y ago
Yes, I've been saying this. It's like if I invented a robot that can run, and then suddenly people started saying it shouldn't be allowed to run down the sidewalk. What's the difference between a human running down the sidewalk and a robot running down the sidewalk? I guess someone could argue that the robot runs a million times faster than any human can, and that sidewalks weren't built with this in mind, but I personally don't see a difference since my robot is just "running" and humans can run too.
- alanfranz 2y ago
  For running, you’d probably be right. No copyright on running. But for copyrighted content, it’s a thorny issue. You already cannot do whatever you want when you buy a book (or a cd, or a dvd); it’s for personal use unless you pay more. Surely you can quote some paragraphs or take a couple of screenshots, and you can learn the overall content, but you cannot reuse it in its entirety. Also, the usage of those contents is kind of “rate limited” by the fact that an human must spit out those answers. Now with AI you get a kind of an opaque system which “ingests” data in some way that can be considered “compression” (rather than “summarization” or “understanding”) and can spit this out at very large rates. Is this a fair thing? If you got a small model, probably your model has learned the concept, so I could agree that copyright is not a concern; but with LLMs? The NYY lawsuit is very relevant. Also, I suspect that humans will adapt and create content which is less useful or harder to digest for AIs - maybe by making it worse for humans as well. Or they will use AIs to create bad content on their topics of expertise to make AI training data worse. There’s a serious risk of horrible content flooding the internet (and the publishing world). LLMs are useful but risky. We should take care and consider 2nd order consequences.
  - dorkwood 2y ago
    I personally wouldn't want a robot running a million times faster than a human down the side walk. The argument that it's running just like a human runs isn't convincing to me.

dorkwood 2y ago

Yes, I've been saying this. It's like if I invented a robot that can run, and then suddenly people started saying it shouldn't be allowed to run down the sidewalk. What's the difference between a human running down the sidewalk and a robot running down the sidewalk? I guess someone could argue that the robot runs a million times faster than any human can, and that sidewalks weren't built with this in mind, but I personally don't see a difference since my robot is just "running" and humans can run too.

alanfranz 2y ago
For running, you’d probably be right. No copyright on running. But for copyrighted content, it’s a thorny issue. You already cannot do whatever you want when you buy a book (or a cd, or a dvd); it’s for personal use unless you pay more. Surely you can quote some paragraphs or take a couple of screenshots, and you can learn the overall content, but you cannot reuse it in its entirety. Also, the usage of those contents is kind of “rate limited” by the fact that an human must spit out those answers. Now with AI you get a kind of an opaque system which “ingests” data in some way that can be considered “compression” (rather than “summarization” or “understanding”) and can spit this out at very large rates. Is this a fair thing? If you got a small model, probably your model has learned the concept, so I could agree that copyright is not a concern; but with LLMs? The NYY lawsuit is very relevant. Also, I suspect that humans will adapt and create content which is less useful or harder to digest for AIs - maybe by making it worse for humans as well. Or they will use AIs to create bad content on their topics of expertise to make AI training data worse. There’s a serious risk of horrible content flooding the internet (and the publishing world). LLMs are useful but risky. We should take care and consider 2nd order consequences.
- dorkwood 2y ago
  I personally wouldn't want a robot running a million times faster than a human down the side walk. The argument that it's running just like a human runs isn't convincing to me.