4 ms·

OpenAI plainly admitted that it is impossible not to do so in a House of Lords inquiry. So, presumably there is no way around it to train models. There is just

by ungut 1mo ago

OpenAI plainly admitted that it is impossible not to do so in a House of Lords inquiry. So, presumably there is no way around it to train models. There is just not enough non-copyrighted data out there.

yorwba 1mo ago
You mean this one https://committees.parliament.uk/writtenevidence/126981/pdf/ https://committees.parliament.uk/writtenevidence/126981/pdf/ where they write "it would be impossible to train today’s leading AI models without using copyrighted materials"? That doesn't mean they have to download those materials illegally. For a billion dollars, you can easily buy one legal copy of each book in Anna's Archive and still have some cash left over to run a whole-of-internet scraping operation.
- spwa4 1mo ago
  I'm pretty sure we would know if they did that. And we don't. Plus this is not legal in the EU (and Canada, and ... let's just say the entire rest of the world, and accept that I'll be wrong for one or two smaller countries). Doesn't that matter? Or is only Mistral disallowed from training on copyrighted materials? Je veux ma chaton fat, goddamit!
  - yorwba 1mo ago
    AI companies legally acquiring books have indeed been in the news: https://news.ycombinator.com/item?id=49330742 https://news.ycombinator.com/item?id=49330742 And where are you getting the idea that Mistral doesn't train on copyrighted data? There's not a lot of code written by people who've been dead for more than 70 years, but somehow Mistral has been able to release coding models anyway.
    - spwa4 1mo ago
      But they have been training on copyrighted data since GPT-2 at least. 2019, and that's when it came out, so before that of course.
      - yorwba 1mo ago
        GPT-2 was trained using data scraped from the web (https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf https://cdn.openai.com/better-language-models/language_model... section 2.1), i.e. copyrighted data provided free of charge to anyone with an internet connection.
        spwa4 1mo ago
        You mean very likely the Anna's archive torrent dump because it's MUCH better quality than the general internet and beyond a certain amount of input data (which is a lot, but much less than the internet) the only thing that matters in training is the quality of the data, to the point that now many labs have thousands of people just making and improving essentially school exercises full time? Hell, I know that for one "lab" (kindof AI lab) since 2020 or so has determined wikipedia quality is dropping fast. It was already dropping slowly before that, but now it's getting bad.
        yorwba 1mo ago
        No, I mean the WebText corpus whose construction from 45 million Reddit post with at least 3 karma is described in section 2.1 of the PDF I linked. They did remove all Wikipedia documents. Anna's Archive didn't exist in 2019.
- tygon 1mo ago
  I wonder if anyone has run the numbers on what the actual cost, both in cash and logistical headache, contacting so many copyright holders would be. That seems like quite the feat to calculate.
  - yorwba 1mo ago
    There's an established network of intermediaries that can supply a large variety of books for a few dollars apiece, so no need to contact copyright holders directly.
    - tygon 1mo ago
      This is very true. As someone with quite the experience with materials published under Penguin, Scholastic, etc. you effectively have a "dictionary attack" on the matter, rather than true "brute force," but that still leaves quite a list to compile to send to each and is easier for larger titles than smaller ones. I wonder how that leads to a bias in what materials get used for training. You are not getting many local self-published books this way. It is almost like we need a "for use for training" agreement across the board. This would not fix the current issues (at least without substantial work), but going forward would allow for creators (or publishers/rights holders) such as this to designate a work as crawl-able for AI. A robots.txt just for Claude.
  - ungut 1mo ago
    Doesn't really matter. The incentive structure to steal clearly exists, so why would they even go through the trouble?
  - ungut 1mo ago
    According to the US Chamber of commerce the US exports over 270 billion dollars. https://www.uschamber.com/intellectual-property/unlocking-creativity-copyright-report https://www.uschamber.com/intellectual-property/unlocking-cr...
- ungut 1mo ago
  Pretty easy to assertain that they don't acquire them legally due to the plethora of evidence and court cases against them. No copyright holder would be sueing them if they knew they sold the works in the first place. I always wonder why y'all feel the need for these impressive mental gymnastics. You can use the models /and/ think they are trained unethically. Living through the ambiguity without abandoning your ideals completely is a valuable skill these days.
  - yorwba 1mo ago
    I'm not aware of any successful accusations against OpenAI for illegally obtaining copyrighted material, in contrast to Anthropic, who settled for $3000 per work and then still had to buy legal copies to keep using them (likely for much less). Instead, the ongoing lawsuits focus on the idea that AI training involves making additional copies, for which they would need a copyright license instead of just one legal copy.
    - ungut 1mo ago
      You are kind of right, but you also did not look very hard. They deleted huge datasets in anticipation of lawsuits, at least that much is known. Of course plaintiffs were unable to depose their internal lawyers (who apparently know why they were frantically deleted) due to 'attourney client priviledge' further refusing to provide any kind of transparency. But yeah, I guess they are they are better at covering their tracks and destroying evidence. Also, as one more example, I find it hard to believe that their models could generate 'Studio Ghibli' style images without training on the movies. There is no licensing deal between them. I think the real issues here are two-fold: Firstly, Copyright is very ill equipped to handle these cases. Just because the model is tuned not to output the exact training data does not mean that compressing mostly-copyrighted datasets into a proprietary model is ethical, fair or /should/ be allowed, simply because they might destroy entire livelihoods. If you take those copyrighted works away you are left with, in OpenAIs own words, a cute little experiment. Secondly, there is absolutely no transparency. Datasets are easily deleted and its impossible to tell what the models have been trained on, especially after fine tuning. Moreover, only the biggest most successfull works would be easily identifiable without the fine tuned model. Once again, sticking it to the little man.
deadbunny 1mo ago
What tosh. It's copywrited material, they pay to access it like everyone else.

by ungut 1mo ago

yorwba 1mo ago

You mean this one https://committees.parliament.uk/writtenevidence/126981/pdf/ https://committees.parliament.uk/writtenevidence/126981/pdf/ where they write "it would be impossible to train today’s leading AI models without using copyrighted materials"? That doesn't mean they have to download those materials illegally. For a billion dollars, you can easily buy one legal copy of each book in Anna's Archive and still have some cash left over to run a whole-of-internet scraping operation.

spwa4 1mo ago
I'm pretty sure we would know if they did that. And we don't. Plus this is not legal in the EU (and Canada, and ... let's just say the entire rest of the world, and accept that I'll be wrong for one or two smaller countries). Doesn't that matter? Or is only Mistral disallowed from training on copyrighted materials? Je veux ma chaton fat, goddamit!
- yorwba 1mo ago
  AI companies legally acquiring books have indeed been in the news: https://news.ycombinator.com/item?id=49330742 https://news.ycombinator.com/item?id=49330742 And where are you getting the idea that Mistral doesn't train on copyrighted data? There's not a lot of code written by people who've been dead for more than 70 years, but somehow Mistral has been able to release coding models anyway.
  - spwa4 1mo ago
    But they have been training on copyrighted data since GPT-2 at least. 2019, and that's when it came out, so before that of course.
    - yorwba 1mo ago
      GPT-2 was trained using data scraped from the web (https://cdn.openai.com/better-language-models/language_models_are_unsupervised_multitask_learners.pdf https://cdn.openai.com/better-language-models/language_model... section 2.1), i.e. copyrighted data provided free of charge to anyone with an internet connection.
      - spwa4 1mo ago
        You mean very likely the Anna's archive torrent dump because it's MUCH better quality than the general internet and beyond a certain amount of input data (which is a lot, but much less than the internet) the only thing that matters in training is the quality of the data, to the point that now many labs have thousands of people just making and improving essentially school exercises full time? Hell, I know that for one "lab" (kindof AI lab) since 2020 or so has determined wikipedia quality is dropping fast. It was already dropping slowly before that, but now it's getting bad.
        yorwba 1mo ago
        No, I mean the WebText corpus whose construction from 45 million Reddit post with at least 3 karma is described in section 2.1 of the PDF I linked. They did remove all Wikipedia documents. Anna's Archive didn't exist in 2019.
tygon 1mo ago
I wonder if anyone has run the numbers on what the actual cost, both in cash and logistical headache, contacting so many copyright holders would be. That seems like quite the feat to calculate.
- yorwba 1mo ago
  There's an established network of intermediaries that can supply a large variety of books for a few dollars apiece, so no need to contact copyright holders directly.
  - tygon 1mo ago
    This is very true. As someone with quite the experience with materials published under Penguin, Scholastic, etc. you effectively have a "dictionary attack" on the matter, rather than true "brute force," but that still leaves quite a list to compile to send to each and is easier for larger titles than smaller ones. I wonder how that leads to a bias in what materials get used for training. You are not getting many local self-published books this way. It is almost like we need a "for use for training" agreement across the board. This would not fix the current issues (at least without substantial work), but going forward would allow for creators (or publishers/rights holders) such as this to designate a work as crawl-able for AI. A robots.txt just for Claude.
- ungut 1mo ago
  Doesn't really matter. The incentive structure to steal clearly exists, so why would they even go through the trouble?
- ungut 1mo ago
  According to the US Chamber of commerce the US exports over 270 billion dollars. https://www.uschamber.com/intellectual-property/unlocking-creativity-copyright-report https://www.uschamber.com/intellectual-property/unlocking-cr...
ungut 1mo ago
Pretty easy to assertain that they don't acquire them legally due to the plethora of evidence and court cases against them. No copyright holder would be sueing them if they knew they sold the works in the first place. I always wonder why y'all feel the need for these impressive mental gymnastics. You can use the models /and/ think they are trained unethically. Living through the ambiguity without abandoning your ideals completely is a valuable skill these days.
- yorwba 1mo ago
  I'm not aware of any successful accusations against OpenAI for illegally obtaining copyrighted material, in contrast to Anthropic, who settled for $3000 per work and then still had to buy legal copies to keep using them (likely for much less). Instead, the ongoing lawsuits focus on the idea that AI training involves making additional copies, for which they would need a copyright license instead of just one legal copy.
  - ungut 1mo ago
    You are kind of right, but you also did not look very hard. They deleted huge datasets in anticipation of lawsuits, at least that much is known. Of course plaintiffs were unable to depose their internal lawyers (who apparently know why they were frantically deleted) due to 'attourney client priviledge' further refusing to provide any kind of transparency. But yeah, I guess they are they are better at covering their tracks and destroying evidence. Also, as one more example, I find it hard to believe that their models could generate 'Studio Ghibli' style images without training on the movies. There is no licensing deal between them. I think the real issues here are two-fold: Firstly, Copyright is very ill equipped to handle these cases. Just because the model is tuned not to output the exact training data does not mean that compressing mostly-copyrighted datasets into a proprietary model is ethical, fair or /should/ be allowed, simply because they might destroy entire livelihoods. If you take those copyrighted works away you are left with, in OpenAIs own words, a cute little experiment. Secondly, there is absolutely no transparency. Datasets are easily deleted and its impossible to tell what the models have been trained on, especially after fine tuning. Moreover, only the biggest most successfull works would be easily identifiable without the fine tuned model. Once again, sticking it to the little man.

deadbunny 1mo ago

What tosh. It's copywrited material, they pay to access it like everyone else.