5 ms·

Is it just me, or do the comments in this thread seem to be the exact opposite of the sentiment in the comments on similar Github Copilot threads? I just find

by _gabe_ 4y ago

Is it just me, or do the comments in this thread seem to be the exact opposite of the sentiment in the comments on similar Github Copilot threads? I just find it a bit ironic that programmers are irate about Github Copilot using their copyrighted material to train. However, if it's an ML model training off of copyrighted artists material, clearly its a transformative work. I just find the opposing sentiments for these scenarios a bit funny.

mountainriver 4y ago
Yeah I hadn’t thought of it that way. Curious how designers feel about image generation. I imagine everyone will love AI that doesn’t threaten their job
scarmig 4y ago
I think both are fine. One source of the difference in reception (aside from the selfish aspect) is that the space of work generated by Copilot has fewer degrees of freedom than visual art. So the things it produces are much more likely to be near-identical to existing code than Stable Diffusion is to images. I.e. if SD were just providing near identical copies of artists' work, people would be more sympathetic.
pclmulqdq 4y ago
If you use my copyrighted code to train your brain and reproduce exact copies of functions I wrote, I have a problem with that. Same for art. Exact copying is the copyright problem. If you use my copyrighted (public) code to train your brain, and then produce new functions that do new things, there is no problem. Replace "brain" with "AI" and you get my position on this (and a lot of others' positions). It seems to be a lot easier to transform art than code.
- lelag 4y ago
  Indeed. A difference I see between the GPT-like language models and those image generation models is that it feels like the language models actually hold full copies of a lot of sample of the training dataset (hence their ability to recite existing content), whereas the image generation models clearly do not: Stable Diffusion is 4Gb something, yet it can draw anything. That's the amazing part: the training dataset contains 5B images. Yet it distilled all of them in this mere 4Gb of data and can produce an infinite amount of content with those. It really feel like it learned how to draw, in the same sense a human learn and does not simply reproduce exact copy of what it already saw.
  - kixiQu 4y ago
    Haven't we seen a lot of examples of these things vomiting up Starry Night on demand? Isn't that a clear case of memorization?
    - scarmig 4y ago
      Midjourney did this one: https://i0.wp.com/www.technollama.co.uk/wp-content/uploads/2022/08/starry-ai.png https://i0.wp.com/www.technollama.co.uk/wp-content/uploads/2... Not really a copy. I'm sure you could cajole it into something very similar to the actual piece, but at that point it's more the model + the extensive prompt instead of the model itself.
  - majou 4y ago
    32Gb (4GB)
  - earth_walker 4y ago
    I see what you mean about GPT returning copies. But is there really a difference if your model just happens to 'calculate' an exact copy instead? For example, if you enter in something simple and high profile, what it returns looks pretty close to an existing work... Try, for example, typing in "Banksy", "Brad Pitt" or "Starbucks". So if you type in "Photo of a Coffee", "Impressionist painting of a fruit bowl" or "Blue canvas with one red line in the middle", how do you guarantee that the image you get back isn't actually a copy of someone's work?
    - lurquer 4y ago
      Copyright - at least for images - is going to become moot shortly. Copyright exists to allow an artist time to reap the benefits of his labor. Without this time, they say, no rational person would invest his labor into making art. Dubious… but, whatever. The point is, if you remove the ‘labor’ from the mix, there’s no need for copyright. If I can produce spectacular images with zero individual labor, there will be little reason for me to copy someone else’s work.
      - 4y ago
- thunderbird120 4y ago
  Generative models attempt to model their training data. Essentially, they try to be a model of the underlying data distribution from which all samples in the training data were drawn from. A model of that distribution which cannot reproduce all samples from the original training data given the right prompting/query/seed is an incomplete model by definition. If it can't reproduce all samples drawn from the original distribution then it clearly does not model the same distribution those samples were drawn from. That said, this is very very different from just copying at a conceptual level. This is going to end up being a an interesting legal question going forward. I'm curious to see how it turns out.
grandmczeb 4y ago
Maybe you’re referring to a different thread, but I see plenty of people arguing both sides here and in the copilot one.
dagmx 4y ago
That’s entirely how I feel too. I think it’s either A) almost like a dunning Kruger effect where people are less familiar with a domain and therefore think this has superseded it and cheer it on, whereas they’re more familiar with their own craft and can see where it falls short. B) They see it as technology conquering a new area , but get nervous when it starts infringing on their own. I think it’s a bit of both in all likelihood
Karunamon 4y ago
The difference appears to be that Github Copilot just reproduces, verbatim, snippets of code without actually doing anything transformative to it. You cannot describe a sort algorithm and have it spit out something other than a direct reproduction of someone else's implementation. With these artwork models, they can emulate general styles, occasionally known characters or bits of text will show up in the output, but it is (as far as I've seen) never a 1 to 1 faithful reproduction of the training material with no changes made. This is an important difference from a legal perspective, not just a moral one. Whether or not a use of copyrighted work is transformative is a big part of whether that use is fair or not.
- ShamelessC 4y ago
  > verbatim It actually, by default I believe, now checks for exact collisions with any existing GitHub code (the training data) and removes them. These are somewhat rare in any case, is my understanding. Not to detract from your other points. Just a common misconception I see a lot.
- itronitron 4y ago
  Well, if you crop the image you are more likely to end up with an exact copy of prior work. Similarly, Copilot can be seen to reproduce 'snippets of code' but would probably struggle with a full application. The images generated from these AI models are not transformative. In fact they are not even generating art, they are only generating pictures.
postalrat 4y ago
And wasn't it settled back then? Now I'm ok with it.
kixiQu 4y ago
Sometimes people think it ought to be obvious, bright-line-tests only, to determine what is truly transformative or novel. The divergence you're pondering, if taken real charitably, maybe signals that that is wrong and that it requires some domain expertise to be able to ascertain whether imitation is really problematic.
sinenomine 4y ago
I don't think AI trained on my creative output (not just code, mind you) is a problem per se. What is a problem, in my opinion, is the tendency of large corporations and small circles on top of these to monopolize access to these models, and if some of the functionality gets available to the public, it's going through a very paternalistic, corporate, puritan censorship pipeline. If you train artificial intelligence on our hard-won data, the resulting artifact should be available to us. StableDiffusion executed very well here. Problem is, there may be no stability.ai for general-purpose multimodal AIs that are coming this decade, and this technology is a dystopia fuel when it's owned by a select few. AI should be democratized.
- ahmedbaracat 4y ago
  This reminds me of Lex Fridman talking with Jaron Lanier about Data Dignity https://youtu.be/Fx0G6DHMfXM?t=2609 https://youtu.be/Fx0G6DHMfXM?t=2609
winddude 4y ago
Copilot wasn't released opensource.
- peoplefromibiza 4y ago
  the dataset SD was trained on is not opensource either, it can't be, it includes copyrighted material for which they have no right or license.
zaptheimpaler 4y ago
Copilot is paid and closed source. It takes freely from the commons but doesn't give back as freely. There is a big difference.
- jannes 4y ago
  Your brain is also closed source and you can train it on copyrighted material as well.
napier 4y ago
This is open source. Github Copilot is not.
g42gregory 4y ago
One among many differences between copyright issues for artists and software engineers: software engineers’ copyright designed to protects developers, while art copyright designed to enrich (mostly Hollywood) lawyers.

by _gabe_ 4y ago

mountainriver 4y ago

Yeah I hadn’t thought of it that way. Curious how designers feel about image generation. I imagine everyone will love AI that doesn’t threaten their job

scarmig 4y ago

I think both are fine. One source of the difference in reception (aside from the selfish aspect) is that the space of work generated by Copilot has fewer degrees of freedom than visual art. So the things it produces are much more likely to be near-identical to existing code than Stable Diffusion is to images. I.e. if SD were just providing near identical copies of artists' work, people would be more sympathetic.

pclmulqdq 4y ago

If you use my copyrighted code to train your brain and reproduce exact copies of functions I wrote, I have a problem with that. Same for art. Exact copying is the copyright problem. If you use my copyrighted (public) code to train your brain, and then produce new functions that do new things, there is no problem. Replace "brain" with "AI" and you get my position on this (and a lot of others' positions). It seems to be a lot easier to transform art than code.

lelag 4y ago
Indeed. A difference I see between the GPT-like language models and those image generation models is that it feels like the language models actually hold full copies of a lot of sample of the training dataset (hence their ability to recite existing content), whereas the image generation models clearly do not: Stable Diffusion is 4Gb something, yet it can draw anything. That's the amazing part: the training dataset contains 5B images. Yet it distilled all of them in this mere 4Gb of data and can produce an infinite amount of content with those. It really feel like it learned how to draw, in the same sense a human learn and does not simply reproduce exact copy of what it already saw.
- kixiQu 4y ago
  Haven't we seen a lot of examples of these things vomiting up Starry Night on demand? Isn't that a clear case of memorization?
  - scarmig 4y ago
    Midjourney did this one: https://i0.wp.com/www.technollama.co.uk/wp-content/uploads/2022/08/starry-ai.png https://i0.wp.com/www.technollama.co.uk/wp-content/uploads/2... Not really a copy. I'm sure you could cajole it into something very similar to the actual piece, but at that point it's more the model + the extensive prompt instead of the model itself.
- majou 4y ago
  32Gb (4GB)
- earth_walker 4y ago
  I see what you mean about GPT returning copies. But is there really a difference if your model just happens to 'calculate' an exact copy instead? For example, if you enter in something simple and high profile, what it returns looks pretty close to an existing work... Try, for example, typing in "Banksy", "Brad Pitt" or "Starbucks". So if you type in "Photo of a Coffee", "Impressionist painting of a fruit bowl" or "Blue canvas with one red line in the middle", how do you guarantee that the image you get back isn't actually a copy of someone's work?
  - lurquer 4y ago
    Copyright - at least for images - is going to become moot shortly. Copyright exists to allow an artist time to reap the benefits of his labor. Without this time, they say, no rational person would invest his labor into making art. Dubious… but, whatever. The point is, if you remove the ‘labor’ from the mix, there’s no need for copyright. If I can produce spectacular images with zero individual labor, there will be little reason for me to copy someone else’s work.
    - 4y ago
thunderbird120 4y ago
Generative models attempt to model their training data. Essentially, they try to be a model of the underlying data distribution from which all samples in the training data were drawn from. A model of that distribution which cannot reproduce all samples from the original training data given the right prompting/query/seed is an incomplete model by definition. If it can't reproduce all samples drawn from the original distribution then it clearly does not model the same distribution those samples were drawn from. That said, this is very very different from just copying at a conceptual level. This is going to end up being a an interesting legal question going forward. I'm curious to see how it turns out.

grandmczeb 4y ago

Maybe you’re referring to a different thread, but I see plenty of people arguing both sides here and in the copilot one.

dagmx 4y ago

That’s entirely how I feel too. I think it’s either A) almost like a dunning Kruger effect where people are less familiar with a domain and therefore think this has superseded it and cheer it on, whereas they’re more familiar with their own craft and can see where it falls short. B) They see it as technology conquering a new area , but get nervous when it starts infringing on their own. I think it’s a bit of both in all likelihood

Karunamon 4y ago

The difference appears to be that Github Copilot just reproduces, verbatim, snippets of code without actually doing anything transformative to it. You cannot describe a sort algorithm and have it spit out something other than a direct reproduction of someone else's implementation. With these artwork models, they can emulate general styles, occasionally known characters or bits of text will show up in the output, but it is (as far as I've seen) never a 1 to 1 faithful reproduction of the training material with no changes made. This is an important difference from a legal perspective, not just a moral one. Whether or not a use of copyrighted work is transformative is a big part of whether that use is fair or not.

ShamelessC 4y ago
> verbatim It actually, by default I believe, now checks for exact collisions with any existing GitHub code (the training data) and removes them. These are somewhat rare in any case, is my understanding. Not to detract from your other points. Just a common misconception I see a lot.
itronitron 4y ago
Well, if you crop the image you are more likely to end up with an exact copy of prior work. Similarly, Copilot can be seen to reproduce 'snippets of code' but would probably struggle with a full application. The images generated from these AI models are not transformative. In fact they are not even generating art, they are only generating pictures.

postalrat 4y ago

And wasn't it settled back then? Now I'm ok with it.

kixiQu 4y ago

Sometimes people think it ought to be obvious, bright-line-tests only, to determine what is truly transformative or novel. The divergence you're pondering, if taken real charitably, maybe signals that that is wrong and that it requires some domain expertise to be able to ascertain whether imitation is really problematic.

sinenomine 4y ago

I don't think AI trained on my creative output (not just code, mind you) is a problem per se. What is a problem, in my opinion, is the tendency of large corporations and small circles on top of these to monopolize access to these models, and if some of the functionality gets available to the public, it's going through a very paternalistic, corporate, puritan censorship pipeline. If you train artificial intelligence on our hard-won data, the resulting artifact should be available to us. StableDiffusion executed very well here. Problem is, there may be no stability.ai for general-purpose multimodal AIs that are coming this decade, and this technology is a dystopia fuel when it's owned by a select few. AI should be democratized.

ahmedbaracat 4y ago
This reminds me of Lex Fridman talking with Jaron Lanier about Data Dignity https://youtu.be/Fx0G6DHMfXM?t=2609 https://youtu.be/Fx0G6DHMfXM?t=2609

winddude 4y ago

Copilot wasn't released opensource.

peoplefromibiza 4y ago
the dataset SD was trained on is not opensource either, it can't be, it includes copyrighted material for which they have no right or license.

zaptheimpaler 4y ago

Copilot is paid and closed source. It takes freely from the commons but doesn't give back as freely. There is a big difference.

jannes 4y ago
Your brain is also closed source and you can train it on copyrighted material as well.

napier 4y ago

This is open source. Github Copilot is not.

g42gregory 4y ago

One among many differences between copyright issues for artists and software engineers: software engineers’ copyright designed to protects developers, while art copyright designed to enrich (mostly Hollywood) lawyers.