3 ms·

I can't think of a single argument that trivially transfers between the two, could you elaborate?

mensetmanusman 3y ago
I think of LLM’s as a vast digital information landscape, vaguely analogous to walking through World of Warcraft, except at a higher dimensionality of existence. A photographer doesn’t create nature, the subjects, the architecture, the universe, the billions of R&D behind the silicon sensors, etc., but they create the situation and push the button. Creating the situation, I.e. drawing from the near infinite of possibilities, is analogous to prompting.
- some_random 3y ago
  People have enforced copyright over photos, the Eiffel tower's lights are copyrighted, some tattoo artists try to enforce copyright over their work, and I'm pretty sure you can't just take a photo of someone's painting and sell it. Any photo of a pine tree is the exclusive copyright of the photographer, nature can't enforce copyright. A photo of a pine tree generated from hundreds of copyrighted photos of pine trees is a completely different question, I think your argument is too reductive.
  - mensetmanusman 3y ago
    You can not take a photo of someone’s art or likeness and sell it in many cases (or if abuse etc.), this is true and should be true of LLMs. In this case the photography analogy holds. For the case of observing copyrighted photos (not directly storing indefinitely) and generalizing a pine tree from that data, I think that is fine: -assume we believe and rule it is not fine, then we will just allow huge companies to generate their own data sets and they will own this space moving forward (would require google maps levels of investment). Also, we wouldn’t be able to tell the difference between illegal SD output versus clean FAANG output without extremely draconian measures.
112233 3y ago
in both cases the building/scene/model training data is made by others. you just pick few parameters (location/light/prompt)
- some_random 3y ago
  Depending on what you take a photo of, your photo may be copyrighted by someone else, I don't think this is a strong argument at all
  - 112233 3y ago
    Why? Photography is an established practice, with existing rules, so it seems like a good aid at reasoning about ml image generators. E.g. by considering in which situations "your photo may be copyrighted by someone else" it might be easier to reason about cases when "your generated image may be copyrighted by someone else"
Cerium 3y ago
I think it is like photography. If an image is created by Dall-e then the copyright of that image should belong to the person that created the prompt and selected the output. Those are creative activities, much like pointing a camera and deciding which negatives to use going forward. Also, like photography, the equipment does not claim photography, be that cameras or background paper, etc.
- some_random 3y ago
  If you take a photo of something copyrighted, that copyright may be enforced on the photo. If Dall-e is trained on loads of copyrighted material (which it is) then the argument is that it's output is owned by the people who's art it was trained on rather than the photographer who just took a picture.

I think of LLM’s as a vast digital information landscape, vaguely analogous to walking through World of Warcraft, except at a higher dimensionality of existence. A photographer doesn’t create nature, the subjects, the architecture, the universe, the billions of R&D behind the silicon sensors, etc., but they create the situation and push the button. Creating the situation, I.e. drawing from the near infinite of possibilities, is analogous to prompting.

some_random 3y ago
People have enforced copyright over photos, the Eiffel tower's lights are copyrighted, some tattoo artists try to enforce copyright over their work, and I'm pretty sure you can't just take a photo of someone's painting and sell it. Any photo of a pine tree is the exclusive copyright of the photographer, nature can't enforce copyright. A photo of a pine tree generated from hundreds of copyrighted photos of pine trees is a completely different question, I think your argument is too reductive.
- mensetmanusman 3y ago
  You can not take a photo of someone’s art or likeness and sell it in many cases (or if abuse etc.), this is true and should be true of LLMs. In this case the photography analogy holds. For the case of observing copyrighted photos (not directly storing indefinitely) and generalizing a pine tree from that data, I think that is fine: -assume we believe and rule it is not fine, then we will just allow huge companies to generate their own data sets and they will own this space moving forward (would require google maps levels of investment). Also, we wouldn’t be able to tell the difference between illegal SD output versus clean FAANG output without extremely draconian measures.

112233 3y ago

in both cases the building/scene/model training data is made by others. you just pick few parameters (location/light/prompt)

some_random 3y ago
Depending on what you take a photo of, your photo may be copyrighted by someone else, I don't think this is a strong argument at all
- 112233 3y ago
  Why? Photography is an established practice, with existing rules, so it seems like a good aid at reasoning about ml image generators. E.g. by considering in which situations "your photo may be copyrighted by someone else" it might be easier to reason about cases when "your generated image may be copyrighted by someone else"

Cerium 3y ago

I think it is like photography. If an image is created by Dall-e then the copyright of that image should belong to the person that created the prompt and selected the output. Those are creative activities, much like pointing a camera and deciding which negatives to use going forward. Also, like photography, the equipment does not claim photography, be that cameras or background paper, etc.

some_random 3y ago
If you take a photo of something copyrighted, that copyright may be enforced on the photo. If Dall-e is trained on loads of copyrighted material (which it is) then the argument is that it's output is owned by the people who's art it was trained on rather than the photographer who just took a picture.