6 ms·

> That's literally built into their corporate rules for how to take investment money, and when those rules were written they were criticised because people didn

> That's literally built into their corporate rules for how to take investment money, and when those rules were written they were criticised because people didn't think they'd ever grow enough for it to matter. How is OpenAI compensating the owners of IP they trained their models on? Or is that not what you mean? It's certainly how I read the part of the GP comment you quoted.

ben_w 3y ago
So far, looks like funding a UBI study. As the IP owners are approximately "everyone" in law, UBI is kinda the only way to compensate all the IP owners. https://openai.com/our-structure https://openai.com/our-structure
- edc117 3y ago
  Good catch, hadn't seen that. So the researchers, shareholders, and leadership of OpenAI will be happy to give up being ridiculously wealthy so they can be only moderately wealthy, and everyone else gets a basic income? I'm also just skeptical of UBI in general, I suppose - 'free' money tends to just inflate everything to account for it, and it still won't address scarcity issues for limited physical assets like land/property. I've love to be wrong about both of these things.
  - ben_w 3y ago
    I agree UBI can have that problem, but I think it can be avoided if e.g. the government owns the means of production. There are still risks in this scenario.
- lp0_on_fire 3y ago
  > As the IP owners are approximately "everyone" in law That makes no sense. If I write a book by myself, post part of it on my website and OpenAI ingests part of it - how does that make anyone besides me myself and I an "owner" of the IP?
  - ben_w 3y ago
    I don't understand why you're confused, but I think it's linguistics. If you write a book by yourself and post parts on your website and they ingest it, you are the copyright holder of that specific IP, and when I post this specific comment to Hacker News I am the copyright holder of this specific IP. In aggregate you and I together are the copyright holders of that book sample and this post, and I don't know any other way of formulating that sentence, though it sounds like you think I'm trying to claim ownership of your hypothetical book while also giving you IP ownership of this post? But that's not my intent.
    - lp0_on_fire 3y ago
      I don't think you're trying to claim ownership. It sounded like you were suggesting that the only recourse for OpenAI would be to fund a UBI program as a form of payment instead of directly paying the people who own the IP it ingested?
      - MacsHeadroom 3y ago
        They ingested the entirety of the internet. Everyone who has ever written anything, including our (implicitly copyrighted) HN comments and letters written 400 years ago, which is online was used to train GPT-4.
      - ben_w 3y ago
        Yes, I'm saying that because there's (currently) no way to even tell how much the model was improved by my comments on HN vs. an equal number of tokens that came from e.g. nytimes.com; furthermore, to the extent that it is even capable of causing economic losses to IP holders, I think this necessarily requires the model to be actually good and not just a bad mimic[0] or prone to whimsy[1] and that this economic damage will occur equally to all IP holders regardless of whether or not their IP was used in training. For both of these reasons independently, I currently think UBI is the only possible fair outcome. [0] I find the phrase "stochastic parrot" to be ironic, as people repeat it mindlessly and with a distribution that could easily be described by a Markov model. [1] if the model is asked to produce something in the style of NYT, but everyone knows it may randomly insert a nonsense statement about President Trump's first visit to the Moon, that's not detracting from the value of buying a copy of the newspaper.
- madeofpalk 3y ago
  Until my check from OpenAI shows up, they're not.
- caconym_ 3y ago
  This is a load of bullshit and I sincerely hope you know that as well as I do. As a thought experiment, let's say I pirate enough ebooks to stock a virtual library roughly equivalent in scope to a large metropolitan library system, then put up a website where you can download these books for free. I make money on the ads I run on this website, etc. This is theft, but as "compensation" I put some percentage of my revenues into funding a UBI study that might, if we're lucky—in half a century or so, in a progressive, enlightened version of the future we are by no means guaranteed to realize—make a fractional contribution to the thrust of a successful UBI movement. Does that make what I'm doing okay? Should all those authors deprived of royalties on their work now, even deprived of publishing opportunities as legitimate sales collapse, understand my token contribution to UBI as fair compensation for what I'm taking from them? That to me is a joke, and the only difference between it and what OpenAI is doing is that OpenAI's product relies on a technical means of laundering intellectual property that seems tailor-made to dodge a body of existing copyright law designed by people who could not possibly have conceived of what modern genAI is capable of. We will see what our lawmakers and courts make of it now, but either way, making a promise to pay me back later does not justify you in taking all the cash out of my wallet without my consent. Nor, for that matter, does tearing it up and returning it to me in the form of a papier-mâché sculpture of Shrek's head protruding from the bowl of a "skibidi toilet".
  - ben_w 3y ago
    IMO that's a terrible thought experiment given the situation. LLMs do not store enough content or with enough accuracy to even close to a virtual library. Unlike, say, Google and the Wayback Machine, the former of which stores enough to show snippets from the pages it's presenting to you as search results (and they got sued for that in certain categories of result), and the latter is straight up an archive of all the sites it crawls. Furthermore, the "percentage" in question for OpenAI is "once we've paid off our investors, all of it goes to benefitting humanity one way or another" — the parent company is a not-for-profit. > Does that make what I'm doing okay? Should all those authors deprived of royalties on their work now, even deprived of publishing opportunities as legitimate sales collapse, understand my token contribution to UBI as fair compensation for what I'm taking from them? Here's a different question for you: If a generative AI is trained only on out-of-copyright novels and open-licensed modern works, and then still deprives everyone of all publishing opportunities forever as in this thought experiment it's better and cheaper than any human novelist, is that any more, or any less fair on literally any person on the planet? The outcomes are the same. I'm sure someone's already thought of making such a model, it's just a question of if they raised enough money to train such a model. > OpenAI's product relies on a technical means of laundering intellectual property that seems tailor-made to dodge a body of existing copyright law designed by people who could not possibly have conceived of what modern genAI is capable of You may have noticed from the version number that they're on versions 3 and 4. When version 2 came out in 2019, they said: """Our model, called GPT-2 (a successor to GPT), was trained simply to predict the next word in 40GB of Internet text. Due to our concerns about malicious applications of the technology, we are not releasing the trained model.""" They were mocked for this: https://slate.com/technology/2019/02/openai-gpt2-text-generating-algorithm-ai-dangerous.html https://slate.com/technology/2019/02/openai-gpt2-text-genera... Even here: https://news.ycombinator.com/item?id=21306542 https://news.ycombinator.com/item?id=21306542 Indeed, they are still mocked for suggesting their models may carry any risk at all. Of any kind. There are plenty of people who want to rush forward with this and think OpenAI are needlessly slow and cautious. You may also have noticed their CEO gave testimony in the US Congress, and that the people asking him questions were surprised he said (to paraphrase) "regulate us specifically — not the open source models, they're not good enough yet — us". To the extent that any GenAI can pose an economic threat to a creative job, it has to be better than a human in that same job. For now, IMO, they're assistant-level, not economic-threat-level. And when they get to economic-threat-level (which in fairness could be next month or next year), they'll be that threat even if none of your IP ever entered their training runs.
- kelnos 3y ago
  That... doesn't seem sufficient, or legal, or (if legal) ethical. You can't just "compensate" people for using their copyrighted works via whatever means you've decided is fair. I think funding UBI studies and lobbying for that sort of thing is a public good, but is entirely unrelated to -- and does not make up for -- wholesale copyright infringement.
  - ben_w 3y ago
    I take no position at all about legality, if scraping is or is not legal and if LLMs are or are not "fair use" is quite beyond my limited grasp of international copyright law. But moral/ethical, and/or sufficient compensation? IMO (and it is just opinion), the damage GenAI does to IP value happens when, and only when, an AI is good enough to make some human unemployable, and that happens to all e.g. novelists around the same time even those whose IP was explicitly excluded from training the model. So, twist question: is it fair to pay a UBI to people who refuse to contribute to some future AI that does end up making us all redundant? (My answer is "yes, it's fair to pay to all even if they contributed nothing, this is a terrible place for schadenfreude"). Conversely, mediocre regurgitation of half-remembered patterns that mimic the style of a famous author cannot cause any more harm when done by AI than when done by fan fiction. Right now these models are pretty bad at creative fiction, pretty good at writing code, so I expect this to impact us before novelists, despite the flood of mediocre AI books reported in various places. Other damage can happen independently of IP value damage, like fully automated propaganda, but that seems like it's not a place where compensation would go to a copyright holder in the first place. Sufficient compensation? If AI works out, nobody will have any economic advantage, and UBI is the only fair and ethical option I am yet aware of for that. It's what happens between here and there that's messy.

ben_w 3y ago

So far, looks like funding a UBI study. As the IP owners are approximately "everyone" in law, UBI is kinda the only way to compensate all the IP owners. https://openai.com/our-structure https://openai.com/our-structure

edc117 3y ago
Good catch, hadn't seen that. So the researchers, shareholders, and leadership of OpenAI will be happy to give up being ridiculously wealthy so they can be only moderately wealthy, and everyone else gets a basic income? I'm also just skeptical of UBI in general, I suppose - 'free' money tends to just inflate everything to account for it, and it still won't address scarcity issues for limited physical assets like land/property. I've love to be wrong about both of these things.
- ben_w 3y ago
  I agree UBI can have that problem, but I think it can be avoided if e.g. the government owns the means of production. There are still risks in this scenario.
lp0_on_fire 3y ago
> As the IP owners are approximately "everyone" in law That makes no sense. If I write a book by myself, post part of it on my website and OpenAI ingests part of it - how does that make anyone besides me myself and I an "owner" of the IP?
- ben_w 3y ago
  I don't understand why you're confused, but I think it's linguistics. If you write a book by yourself and post parts on your website and they ingest it, you are the copyright holder of that specific IP, and when I post this specific comment to Hacker News I am the copyright holder of this specific IP. In aggregate you and I together are the copyright holders of that book sample and this post, and I don't know any other way of formulating that sentence, though it sounds like you think I'm trying to claim ownership of your hypothetical book while also giving you IP ownership of this post? But that's not my intent.
  - lp0_on_fire 3y ago
    I don't think you're trying to claim ownership. It sounded like you were suggesting that the only recourse for OpenAI would be to fund a UBI program as a form of payment instead of directly paying the people who own the IP it ingested?
    - MacsHeadroom 3y ago
      They ingested the entirety of the internet. Everyone who has ever written anything, including our (implicitly copyrighted) HN comments and letters written 400 years ago, which is online was used to train GPT-4.
    - ben_w 3y ago
      Yes, I'm saying that because there's (currently) no way to even tell how much the model was improved by my comments on HN vs. an equal number of tokens that came from e.g. nytimes.com; furthermore, to the extent that it is even capable of causing economic losses to IP holders, I think this necessarily requires the model to be actually good and not just a bad mimic[0] or prone to whimsy[1] and that this economic damage will occur equally to all IP holders regardless of whether or not their IP was used in training. For both of these reasons independently, I currently think UBI is the only possible fair outcome. [0] I find the phrase "stochastic parrot" to be ironic, as people repeat it mindlessly and with a distribution that could easily be described by a Markov model. [1] if the model is asked to produce something in the style of NYT, but everyone knows it may randomly insert a nonsense statement about President Trump's first visit to the Moon, that's not detracting from the value of buying a copy of the newspaper.
madeofpalk 3y ago
Until my check from OpenAI shows up, they're not.
caconym_ 3y ago
This is a load of bullshit and I sincerely hope you know that as well as I do. As a thought experiment, let's say I pirate enough ebooks to stock a virtual library roughly equivalent in scope to a large metropolitan library system, then put up a website where you can download these books for free. I make money on the ads I run on this website, etc. This is theft, but as "compensation" I put some percentage of my revenues into funding a UBI study that might, if we're lucky—in half a century or so, in a progressive, enlightened version of the future we are by no means guaranteed to realize—make a fractional contribution to the thrust of a successful UBI movement. Does that make what I'm doing okay? Should all those authors deprived of royalties on their work now, even deprived of publishing opportunities as legitimate sales collapse, understand my token contribution to UBI as fair compensation for what I'm taking from them? That to me is a joke, and the only difference between it and what OpenAI is doing is that OpenAI's product relies on a technical means of laundering intellectual property that seems tailor-made to dodge a body of existing copyright law designed by people who could not possibly have conceived of what modern genAI is capable of. We will see what our lawmakers and courts make of it now, but either way, making a promise to pay me back later does not justify you in taking all the cash out of my wallet without my consent. Nor, for that matter, does tearing it up and returning it to me in the form of a papier-mâché sculpture of Shrek's head protruding from the bowl of a "skibidi toilet".
- ben_w 3y ago
  IMO that's a terrible thought experiment given the situation. LLMs do not store enough content or with enough accuracy to even close to a virtual library. Unlike, say, Google and the Wayback Machine, the former of which stores enough to show snippets from the pages it's presenting to you as search results (and they got sued for that in certain categories of result), and the latter is straight up an archive of all the sites it crawls. Furthermore, the "percentage" in question for OpenAI is "once we've paid off our investors, all of it goes to benefitting humanity one way or another" — the parent company is a not-for-profit. > Does that make what I'm doing okay? Should all those authors deprived of royalties on their work now, even deprived of publishing opportunities as legitimate sales collapse, understand my token contribution to UBI as fair compensation for what I'm taking from them? Here's a different question for you: If a generative AI is trained only on out-of-copyright novels and open-licensed modern works, and then still deprives everyone of all publishing opportunities forever as in this thought experiment it's better and cheaper than any human novelist, is that any more, or any less fair on literally any person on the planet? The outcomes are the same. I'm sure someone's already thought of making such a model, it's just a question of if they raised enough money to train such a model. > OpenAI's product relies on a technical means of laundering intellectual property that seems tailor-made to dodge a body of existing copyright law designed by people who could not possibly have conceived of what modern genAI is capable of You may have noticed from the version number that they're on versions 3 and 4. When version 2 came out in 2019, they said: """Our model, called GPT-2 (a successor to GPT), was trained simply to predict the next word in 40GB of Internet text. Due to our concerns about malicious applications of the technology, we are not releasing the trained model.""" They were mocked for this: https://slate.com/technology/2019/02/openai-gpt2-text-generating-algorithm-ai-dangerous.html https://slate.com/technology/2019/02/openai-gpt2-text-genera... Even here: https://news.ycombinator.com/item?id=21306542 https://news.ycombinator.com/item?id=21306542 Indeed, they are still mocked for suggesting their models may carry any risk at all. Of any kind. There are plenty of people who want to rush forward with this and think OpenAI are needlessly slow and cautious. You may also have noticed their CEO gave testimony in the US Congress, and that the people asking him questions were surprised he said (to paraphrase) "regulate us specifically — not the open source models, they're not good enough yet — us". To the extent that any GenAI can pose an economic threat to a creative job, it has to be better than a human in that same job. For now, IMO, they're assistant-level, not economic-threat-level. And when they get to economic-threat-level (which in fairness could be next month or next year), they'll be that threat even if none of your IP ever entered their training runs.
kelnos 3y ago
That... doesn't seem sufficient, or legal, or (if legal) ethical. You can't just "compensate" people for using their copyrighted works via whatever means you've decided is fair. I think funding UBI studies and lobbying for that sort of thing is a public good, but is entirely unrelated to -- and does not make up for -- wholesale copyright infringement.
- ben_w 3y ago
  I take no position at all about legality, if scraping is or is not legal and if LLMs are or are not "fair use" is quite beyond my limited grasp of international copyright law. But moral/ethical, and/or sufficient compensation? IMO (and it is just opinion), the damage GenAI does to IP value happens when, and only when, an AI is good enough to make some human unemployable, and that happens to all e.g. novelists around the same time even those whose IP was explicitly excluded from training the model. So, twist question: is it fair to pay a UBI to people who refuse to contribute to some future AI that does end up making us all redundant? (My answer is "yes, it's fair to pay to all even if they contributed nothing, this is a terrible place for schadenfreude"). Conversely, mediocre regurgitation of half-remembered patterns that mimic the style of a famous author cannot cause any more harm when done by AI than when done by fan fiction. Right now these models are pretty bad at creative fiction, pretty good at writing code, so I expect this to impact us before novelists, despite the flood of mediocre AI books reported in various places. Other damage can happen independently of IP value damage, like fully automated propaganda, but that seems like it's not a place where compensation would go to a copyright holder in the first place. Sufficient compensation? If AI works out, nobody will have any economic advantage, and UBI is the only fair and ethical option I am yet aware of for that. It's what happens between here and there that's messy.