5 ms·

>compress an artist's painting into a model That's not how image models work.

by hexage1814 3y ago

greentext 3y ago
Painting features => back propagation => weights. Yes it is.
- gaganyaan 3y ago
  No, it's not. It's quite easy to learn this stuff, about as easy as making shit up on a message forum, so there's no reason not to go learn what you're talking about.
  - greentext 3y ago
    Why don't you counter knowledge with knowledge instead of profanity then? We're all waiting.
Loocid 3y ago
It has been shown that image models can produce originals, or at least extremely close to the originals. If the outcome is the same, what is the difference between compression/decompression vs training/generation regarding copyright?
- sdiupIGPWEfh 3y ago
  > It has been shown that image models can produce originals Not in the general case, no. For the study done against Stable Diffusion [1], researchers were only able to reproduce about 0.03 percent of the images tested. Those were also believed to be cases of overfitting on images which were over-represented in the training data and they're not something you'd hit upon by accident. Generative text models seen to be more problematic, depending on the subject. Code seems especially prone to overfitting, probably due to insufficient amounts of it compared to other text sources as well as lots of copying going on between the repos the models were trained on. [1](https://arstechnica.com/information-technology/2023/02/researchers-extract-training-images-from-stable-diffusion-but-its-difficult/ https://arstechnica.com/information-technology/2023/02/resea...)
  - icehawk 3y ago
    They got 94 direct matches, which is 94 instances where copyright infringement could be argued.
    - sdiupIGPWEfh 3y ago
      Could be argued, sure. If you have to already have access to the copyrighted images to find them in the model, the argument seems weak. A sufficiently advanced model could, in theory, generate any image. You could then, again in theory, find an embedding for any image. Does said model then infringe on all copyrighted images? A program that creates Fourier epicycle drawings could be given input that causes trademarked output. An evolutionary algorithm iterating on noise could, given metrics for an image and the right fitness function, generate infringing images. Hypothetically, and admittedly absurdly, you could extract any image in the binary expansion of Pi and share it by "just" providing an index and length. If you have to know exactly what you're looking for and have to perform a substantial amount of computation to get it, it could be argued that the act of infringement is in the effort made by the person seeking infringing content (and distributing the results) rather than whatever it is they're attempting to extract the content from. But hey, courts don't always make sensible rulings, so who knows.
      - jonathanstrange 3y ago
        > If you have to already have access to the copyrighted images to find them in the model, the argument seems weak. That makes no sense. The copyright holder has access to their own inventions, of course. That's the standard in any copyright claim. > A sufficiently advanced model could, in theory, generate any image. You could then, again in theory, find an embedding for any image. Does said model then infringe on all copyrighted images? Without the slightest doubt. You're already violating copyright if you sing a faulty and badly played version of a pop song in a street cafe without paying a license fee. > you could extract any image in the binary expansion of Pi and share it by "just" providing an index and length The method of storing the information is pretty much irrelevant to copyright. Your link argument has been tried by pirates and it's not working too well, although it depends on the country and legislation.
a13o 3y ago
LLM training is a compression algorithm, LLM weights are a compressed dataset of the training materials, and executing LLMs is accessing the compressed source material. That the compression technology relies on parameterizing the copyrighted material, and as a result can produce hallucinations remixing the copyrighted material, is super cool but doesn't change that this is compression at rest. All the hullabaloo about artificial intelligence is science fiction laundering a (cool new) compression algorithm. Copyright law shouldn't apply any differently to a LLM as it does to gzip.
- epups 3y ago
  Can I talk to gzip in natural language and have it produce novel output not contained within any of its source files? If not, I think your comparison is deeply flawed. LLM's are not simply compression algorithms.
  - a13o 3y ago
    Maybe if someone were to build it. You can't talk to LLMs in natural language either. They have a very precise query language. The natural language component is an additional feature bolted onto the front. Also the output of LLMs is not novel, it's a derivative work of the training dataset. LLMs can't produce anything not present in the training set. They are operating on parameterized classifications of artwork instead of the artwork itself, which is the new technology; but they can't do anything except combine those existing Legos in new ways. We also talk to databases in what was considered 'natural language' for that era - SQL. The LLMs are no different, which is why we have prompt engineering same as we have database engineering. Just because the database is processing the dataset and storing it in a proprietary way, approachable via query language, does not change the fact that the original data is still in there, and everything that comes out is derivative.

>compress an artist's painting into a model That's not how image models work.

greentext 3y ago

Painting features => back propagation => weights. Yes it is.

gaganyaan 3y ago
No, it's not. It's quite easy to learn this stuff, about as easy as making shit up on a message forum, so there's no reason not to go learn what you're talking about.
- greentext 3y ago
  Why don't you counter knowledge with knowledge instead of profanity then? We're all waiting.

Loocid 3y ago

It has been shown that image models can produce originals, or at least extremely close to the originals. If the outcome is the same, what is the difference between compression/decompression vs training/generation regarding copyright?

sdiupIGPWEfh 3y ago
> It has been shown that image models can produce originals Not in the general case, no. For the study done against Stable Diffusion [1], researchers were only able to reproduce about 0.03 percent of the images tested. Those were also believed to be cases of overfitting on images which were over-represented in the training data and they're not something you'd hit upon by accident. Generative text models seen to be more problematic, depending on the subject. Code seems especially prone to overfitting, probably due to insufficient amounts of it compared to other text sources as well as lots of copying going on between the repos the models were trained on. [1](https://arstechnica.com/information-technology/2023/02/researchers-extract-training-images-from-stable-diffusion-but-its-difficult/ https://arstechnica.com/information-technology/2023/02/resea...)
- icehawk 3y ago
  They got 94 direct matches, which is 94 instances where copyright infringement could be argued.
  - sdiupIGPWEfh 3y ago
    Could be argued, sure. If you have to already have access to the copyrighted images to find them in the model, the argument seems weak. A sufficiently advanced model could, in theory, generate any image. You could then, again in theory, find an embedding for any image. Does said model then infringe on all copyrighted images? A program that creates Fourier epicycle drawings could be given input that causes trademarked output. An evolutionary algorithm iterating on noise could, given metrics for an image and the right fitness function, generate infringing images. Hypothetically, and admittedly absurdly, you could extract any image in the binary expansion of Pi and share it by "just" providing an index and length. If you have to know exactly what you're looking for and have to perform a substantial amount of computation to get it, it could be argued that the act of infringement is in the effort made by the person seeking infringing content (and distributing the results) rather than whatever it is they're attempting to extract the content from. But hey, courts don't always make sensible rulings, so who knows.
    - jonathanstrange 3y ago
      > If you have to already have access to the copyrighted images to find them in the model, the argument seems weak. That makes no sense. The copyright holder has access to their own inventions, of course. That's the standard in any copyright claim. > A sufficiently advanced model could, in theory, generate any image. You could then, again in theory, find an embedding for any image. Does said model then infringe on all copyrighted images? Without the slightest doubt. You're already violating copyright if you sing a faulty and badly played version of a pop song in a street cafe without paying a license fee. > you could extract any image in the binary expansion of Pi and share it by "just" providing an index and length The method of storing the information is pretty much irrelevant to copyright. Your link argument has been tried by pirates and it's not working too well, although it depends on the country and legislation.

a13o 3y ago

LLM training is a compression algorithm, LLM weights are a compressed dataset of the training materials, and executing LLMs is accessing the compressed source material. That the compression technology relies on parameterizing the copyrighted material, and as a result can produce hallucinations remixing the copyrighted material, is super cool but doesn't change that this is compression at rest. All the hullabaloo about artificial intelligence is science fiction laundering a (cool new) compression algorithm. Copyright law shouldn't apply any differently to a LLM as it does to gzip.

epups 3y ago
Can I talk to gzip in natural language and have it produce novel output not contained within any of its source files? If not, I think your comparison is deeply flawed. LLM's are not simply compression algorithms.
- a13o 3y ago
  Maybe if someone were to build it. You can't talk to LLMs in natural language either. They have a very precise query language. The natural language component is an additional feature bolted onto the front. Also the output of LLMs is not novel, it's a derivative work of the training dataset. LLMs can't produce anything not present in the training set. They are operating on parameterized classifications of artwork instead of the artwork itself, which is the new technology; but they can't do anything except combine those existing Legos in new ways. We also talk to databases in what was considered 'natural language' for that era - SQL. The LLMs are no different, which is why we have prompt engineering same as we have database engineering. Just because the database is processing the dataset and storing it in a proprietary way, approachable via query language, does not change the fact that the original data is still in there, and everything that comes out is derivative.