5 ms·

> they're copyright laundering systems. You put a few images (books, etc.) in, mix it around, and "poof" there disappears the copyright This also describes the

> they're copyright laundering systems. You put a few images (books, etc.) in, mix it around, and "poof" there disappears the copyright This also describes the process of going to university.

visarga 3y ago
Why not? This would fix the regurgitation issue and possibly be more palatable to copyright holders. The LLM trained on both copyrighted and open works can generate training data for a second LLM that never sees copyrighted works, achieving perfect separation of idea from expression. Copyright is supposed to protect expression but not ideas. Ideas are free, LLMs should be allowed to learn ideas. This can also filter out bad quality data and PII, or expand on parts of the distribution with less samples. It would be "dataset engineering" as opposed to "model engineering". A recent paper (TinyStories) shows you can train a LM entirely with synthetic data and even evaluate it with another LLM. They could make a model 1000x smaller that still has fluent English - child level English more specifically, but well articulated, and even capable of reasoning. https://arxiv.org/abs/2305.07759 https://arxiv.org/abs/2305.07759 I believe in the future LLMs will train on 100% synthetic data from previous generations of LLM. The benefit will be increased quality, more privacy, less copyright infringement and smaller more efficient models, like TinyStories. Only the first few generations of LLM need to scrape the whole internet, the next generations will have better data.
mjburgess 3y ago
> This also describes the process of going to university. I do not recall the lesson wherein I sampled from thousands of prior exam papers to generate my answer to the one I sat. Without any prior exam papers at all I could, indeed, sit the exam. The algorithm which animals follow to be able to do such great things is very expensive: acuring actual skill, technique, knowledge and competence with the world. The algorithm statistical AI follows is simple: hoover up all digitised human history and sample from it. This is entirely dependent on humans not doing that. We have people doing that all the time: writing books, blogs, videos, github, etc. 99.5% of anything we want chatgpt for exists -- we just don't have the audacity to steal it. Now we can!
- justrealist 3y ago
  If you think ChatGPT is dependent on exam training data to answer exam questions, you have a misunderstanding of how LLMs work.
  - mjburgess 3y ago
    You'll have to show me a single case of an LLM being trained on data which excludes exams and continues to answer exam questions. You'll note that ChatGPT performance drops off a cliff on many exams released after 2021. It's a trick.
    - whimsicalism 3y ago
      All of the GPT-4 exams it were evaluated on were not present in its training data. But honestly, if you don't get it now, I can't hope to convince.
      - mjburgess 3y ago
        It's always sampling from an existing distribution of relevant data points -- that's necessarily how its working. If you want to claim the sample set is only mildy similar to exam questions -- so be it, that may be true. Or if you want to claim that its sampling method is attentive to structural associations in its sample set, so that its not lifting from "identical distributions" -- so be it. So long as those "structural associations" are givens, and the data "givens", the process is just sampling from a domain of human effort without expending any of a similar kind. If there had been no internet, ChatGPT would be a dumb mute -- because it has no capacity to generate data; it does not develop actual conceptualisations of the world -- it samples from the data shadows created by people. To produce useful data requires expounding a tremendous effort -- growing an animal to cope with the world. It is this which is being laundered, unpaid and unacknowledged, through LLMs. Whilst star-trek-huffing loons claim this stuff is doing the opposite -- a ideological delusion which benefits all those whose bank accounts are increased by the lie that "ChatGPT wrote this". If we were prepared to price the data commons which has been created over the last 20-30 years of the internet, by everyone, its not hard to think training ChatGPT would cost a trillion. How much labour went into creating that digital resource, and by how many, etc.?
croes 3y ago
The difference is the ease of use and the mass of output.
aeturnum 3y ago
I see this comment all the time in this conversation and it strikes me as being in very bad faith. I would love to understand what you're trying to get at. Because - first, in a legal sense, it seems wrong? You cannot memorize a book and then reproduce it from memory to disappear the copyright. The law on AI & copyright is very young (and I am not an expert), but artistic expression cases that involve copying (i.e. warhol) have rested on the idea that the copying is transformative in some way - but this seems in direct opposition to the idea that machines cannot create works themselves. I.e. the legal understanding that protects AI from needing to respect copyright also suggests it should not be understood to be traditionally transformative. On a less legal level, an AI ingesting a huge amount of content and adjusting a neural network is very different from a student going to university in most ways. There are also similarities! But in a general way the direct comparison does not line up particularly well. So many complaints about Transformer-based AI exist on levels outside of the precise "agent doing the learning" (including this one).

> they're copyright laundering systems. You put a few images (books, etc.) in, mix it around, and "poof" there disappears the copyright This also describes the process of going to university.

visarga 3y ago

Why not? This would fix the regurgitation issue and possibly be more palatable to copyright holders. The LLM trained on both copyrighted and open works can generate training data for a second LLM that never sees copyrighted works, achieving perfect separation of idea from expression. Copyright is supposed to protect expression but not ideas. Ideas are free, LLMs should be allowed to learn ideas. This can also filter out bad quality data and PII, or expand on parts of the distribution with less samples. It would be "dataset engineering" as opposed to "model engineering". A recent paper (TinyStories) shows you can train a LM entirely with synthetic data and even evaluate it with another LLM. They could make a model 1000x smaller that still has fluent English - child level English more specifically, but well articulated, and even capable of reasoning. https://arxiv.org/abs/2305.07759 https://arxiv.org/abs/2305.07759 I believe in the future LLMs will train on 100% synthetic data from previous generations of LLM. The benefit will be increased quality, more privacy, less copyright infringement and smaller more efficient models, like TinyStories. Only the first few generations of LLM need to scrape the whole internet, the next generations will have better data.

mjburgess 3y ago

> This also describes the process of going to university. I do not recall the lesson wherein I sampled from thousands of prior exam papers to generate my answer to the one I sat. Without any prior exam papers at all I could, indeed, sit the exam. The algorithm which animals follow to be able to do such great things is very expensive: acuring actual skill, technique, knowledge and competence with the world. The algorithm statistical AI follows is simple: hoover up all digitised human history and sample from it. This is entirely dependent on humans not doing that. We have people doing that all the time: writing books, blogs, videos, github, etc. 99.5% of anything we want chatgpt for exists -- we just don't have the audacity to steal it. Now we can!

justrealist 3y ago
If you think ChatGPT is dependent on exam training data to answer exam questions, you have a misunderstanding of how LLMs work.
- mjburgess 3y ago
  You'll have to show me a single case of an LLM being trained on data which excludes exams and continues to answer exam questions. You'll note that ChatGPT performance drops off a cliff on many exams released after 2021. It's a trick.
  - whimsicalism 3y ago
    All of the GPT-4 exams it were evaluated on were not present in its training data. But honestly, if you don't get it now, I can't hope to convince.
    - mjburgess 3y ago
      It's always sampling from an existing distribution of relevant data points -- that's necessarily how its working. If you want to claim the sample set is only mildy similar to exam questions -- so be it, that may be true. Or if you want to claim that its sampling method is attentive to structural associations in its sample set, so that its not lifting from "identical distributions" -- so be it. So long as those "structural associations" are givens, and the data "givens", the process is just sampling from a domain of human effort without expending any of a similar kind. If there had been no internet, ChatGPT would be a dumb mute -- because it has no capacity to generate data; it does not develop actual conceptualisations of the world -- it samples from the data shadows created by people. To produce useful data requires expounding a tremendous effort -- growing an animal to cope with the world. It is this which is being laundered, unpaid and unacknowledged, through LLMs. Whilst star-trek-huffing loons claim this stuff is doing the opposite -- a ideological delusion which benefits all those whose bank accounts are increased by the lie that "ChatGPT wrote this". If we were prepared to price the data commons which has been created over the last 20-30 years of the internet, by everyone, its not hard to think training ChatGPT would cost a trillion. How much labour went into creating that digital resource, and by how many, etc.?

croes 3y ago

The difference is the ease of use and the mass of output.

aeturnum 3y ago

I see this comment all the time in this conversation and it strikes me as being in very bad faith. I would love to understand what you're trying to get at. Because - first, in a legal sense, it seems wrong? You cannot memorize a book and then reproduce it from memory to disappear the copyright. The law on AI & copyright is very young (and I am not an expert), but artistic expression cases that involve copying (i.e. warhol) have rested on the idea that the copying is transformative in some way - but this seems in direct opposition to the idea that machines cannot create works themselves. I.e. the legal understanding that protects AI from needing to respect copyright also suggests it should not be understood to be traditionally transformative. On a less legal level, an AI ingesting a huge amount of content and adjusting a neural network is very different from a student going to university in most ways. There are also similarities! But in a general way the direct comparison does not line up particularly well. So many complaints about Transformer-based AI exist on levels outside of the precise "agent doing the learning" (including this one).