4 ms·

Is NAI related to SD in any way?

shadowgovt 4y ago
Different trained model, which was extracted from its creator via unauthorized system access and is illicitly available via BitTorrent. The drama appears to have started because an open-source developer who wrote a web frontend to control SD adapted that frontend to control the hacked model also, which has offended SD's CEO (because of the general principle of "Don't help software pirates"). The additional drama includes that said open-source author has looked at the leaked code and concluded that it stole some of his open source work in the way it tunes text parameters.
- ShamelessC 4y ago
  Damn, Emad seems sort of clueless? Hard rules about software piracy like that feel very 90's and totally unnecessary. Just avoid explicitly condoning any projects and move past it! It will be old news in like 2 weeks at the current rate of things.
  - shadowgovt 4y ago
    Given the dude spent half a million dollars on training SD, I wouldn't be surprised if even though he chose to open-source the trained model, he has strong opinions on whether people should have the right to choose to open-source such things vs. having third-party crackers breach their systems and publish for them.
    - nullc 4y ago
      Perhaps, but the model weights themselves are currently understood to be uncopyrightable, and it's pretty inconceivable that the model could become copyrightable without becoming a derivative work of the training data. Unless these AI companies want google and facebook to be literally the only companies in the world that can train large scale machine learning (by using their TOS to get licenses from their users) they should tread carefully. In this particular case the leaked code apparently exposed that the proprietary codebase was also using the OSS developer's work without attribution.
      - shadowgovt 4y ago
        If anything, in the absence of copyrightability, the only protection is trade secrecy and I'd expect Emad to be even deeper in the opinion space of "We cut off the oxygen (systematically speaking) of those who would steal trained ML data." There's an interesting anecdote around how stand-up comedians protect jokes against theft, given how weak copyright is on jokes: it's keying cars, poisoning drinks (generally non-fatally, but it's hard to have a good night on stage when your lower GI tract wants to be elsewhere), and never-work-in-this-town-again agreements. We put these protections into the law because the alternative isn't no protection; it's people-take-it-into-their-own-hands protection.
        pclmulqdq 4y ago
        This method only works when the people you are trying to attack and/or blackball from an industry actually want to be in that industry. Lots of people will just use these ML models without trying to participate in the "ML community" the same way my wife and I tell each other jokes from comedians without trying to be comedians ourselves. The users of these models and the developers are fundamentally different.
      - mensetmanusman 4y ago
        “ it's pretty inconceivable that the model could become copyrightable” If it costs $10 million to find information/weightings/etc., our current legal system would consider that intellectual property which might not be copyrightable but would be considered IP theft if stolen.
        nullc 4y ago
        Yes, it could be a trade secret, but if the trade secrecy would still apply is extraordinarily fact specific. If they were negligent in handling it, e.g. left it on a publicly accessible share and some member of the public stumbled into it, then trade secret protection would likely be lost. If some employee violated their NDA and snuck it out-- well that would be a different matter. etc.
      - 4y ago
- minimaxir 4y ago
  > which has offended SD's CEO (because of the general principle of "Don't help software pirates"). There is likely a legal threat by NAI involved in the decision calculus here.
- neuronexmachina 4y ago
  > The additional drama includes that said open-source author has looked at the leaked code and concluded that it stole some of his open source work in the way it tunes text parameters. My understanding might be out of date, but as I recall it seemed that the code he thought was stolen from his open-source work actually originated from a third project that's MIT licensed.
  - Ukv 4y ago
    I think that's confusing between two issues in opposite directions: 1. Accusations that AUTOMATIC1111 (the web frontend developer) copied code from the NovelAI leak relating to the loading of hypernetworks 2. The leak revealing that Anlatan (the company behind NovelAI) had copied code from AUTOMATIC1111's repo (who, as above, Anlatan are accusing of copying from them) relating to the weighting of words. AUTOMATIC1111's repo does not have a permissive license to allow this The third party MIT-licensed code is relevant to #1. Some code AUTOMATIC1111 was accused of copying from the leak (https://i.imgur.com/r1AkvBG.png https://i.imgur.com/r1AkvBG.png) actually already appears in multiple older permissively-licensed public repos (https://github.com/lucidrains/perceiver-pytorch/blob/main/perceiver_pytorch/perceiver_pytorch.py https://github.com/lucidrains/perceiver-pytorch/blob/main/pe..., https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/attention.py https://github.com/CompVis/stable-diffusion/blob/main/ldm/mo...), one of which was credited in the readme by AUTOMATIC1111. For #2, the Anlatan CEO blamed it on an intern (https://i.imgur.com/BFjKG1V.png https://i.imgur.com/BFjKG1V.png). The leak shows that the offending code was committed by the CEO (https://i.imgur.com/aLiA2tr.png https://i.imgur.com/aLiA2tr.png), which doesn't necessarily rule it out originating from an intern (e.g: "send me the code over teams to review and I'll add it") but doesn't look great. From other examples I'd say AUTOMATIC1111 did get a bit sloppy in terms of not following clean-room design regarding the leak, but I'm inclined to give some leeway to a solo developer making a hugely popular public tool for free.

Different trained model, which was extracted from its creator via unauthorized system access and is illicitly available via BitTorrent. The drama appears to have started because an open-source developer who wrote a web frontend to control SD adapted that frontend to control the hacked model also, which has offended SD's CEO (because of the general principle of "Don't help software pirates"). The additional drama includes that said open-source author has looked at the leaked code and concluded that it stole some of his open source work in the way it tunes text parameters.

ShamelessC 4y ago
Damn, Emad seems sort of clueless? Hard rules about software piracy like that feel very 90's and totally unnecessary. Just avoid explicitly condoning any projects and move past it! It will be old news in like 2 weeks at the current rate of things.
- shadowgovt 4y ago
  Given the dude spent half a million dollars on training SD, I wouldn't be surprised if even though he chose to open-source the trained model, he has strong opinions on whether people should have the right to choose to open-source such things vs. having third-party crackers breach their systems and publish for them.
  - nullc 4y ago
    Perhaps, but the model weights themselves are currently understood to be uncopyrightable, and it's pretty inconceivable that the model could become copyrightable without becoming a derivative work of the training data. Unless these AI companies want google and facebook to be literally the only companies in the world that can train large scale machine learning (by using their TOS to get licenses from their users) they should tread carefully. In this particular case the leaked code apparently exposed that the proprietary codebase was also using the OSS developer's work without attribution.
    - shadowgovt 4y ago
      If anything, in the absence of copyrightability, the only protection is trade secrecy and I'd expect Emad to be even deeper in the opinion space of "We cut off the oxygen (systematically speaking) of those who would steal trained ML data." There's an interesting anecdote around how stand-up comedians protect jokes against theft, given how weak copyright is on jokes: it's keying cars, poisoning drinks (generally non-fatally, but it's hard to have a good night on stage when your lower GI tract wants to be elsewhere), and never-work-in-this-town-again agreements. We put these protections into the law because the alternative isn't no protection; it's people-take-it-into-their-own-hands protection.
      - pclmulqdq 4y ago
        This method only works when the people you are trying to attack and/or blackball from an industry actually want to be in that industry. Lots of people will just use these ML models without trying to participate in the "ML community" the same way my wife and I tell each other jokes from comedians without trying to be comedians ourselves. The users of these models and the developers are fundamentally different.
    - mensetmanusman 4y ago
      “ it's pretty inconceivable that the model could become copyrightable” If it costs $10 million to find information/weightings/etc., our current legal system would consider that intellectual property which might not be copyrightable but would be considered IP theft if stolen.
      - nullc 4y ago
        Yes, it could be a trade secret, but if the trade secrecy would still apply is extraordinarily fact specific. If they were negligent in handling it, e.g. left it on a publicly accessible share and some member of the public stumbled into it, then trade secret protection would likely be lost. If some employee violated their NDA and snuck it out-- well that would be a different matter. etc.
    - 4y ago
minimaxir 4y ago
> which has offended SD's CEO (because of the general principle of "Don't help software pirates"). There is likely a legal threat by NAI involved in the decision calculus here.
neuronexmachina 4y ago
> The additional drama includes that said open-source author has looked at the leaked code and concluded that it stole some of his open source work in the way it tunes text parameters. My understanding might be out of date, but as I recall it seemed that the code he thought was stolen from his open-source work actually originated from a third project that's MIT licensed.
- Ukv 4y ago
  I think that's confusing between two issues in opposite directions: 1. Accusations that AUTOMATIC1111 (the web frontend developer) copied code from the NovelAI leak relating to the loading of hypernetworks 2. The leak revealing that Anlatan (the company behind NovelAI) had copied code from AUTOMATIC1111's repo (who, as above, Anlatan are accusing of copying from them) relating to the weighting of words. AUTOMATIC1111's repo does not have a permissive license to allow this The third party MIT-licensed code is relevant to #1. Some code AUTOMATIC1111 was accused of copying from the leak (https://i.imgur.com/r1AkvBG.png https://i.imgur.com/r1AkvBG.png) actually already appears in multiple older permissively-licensed public repos (https://github.com/lucidrains/perceiver-pytorch/blob/main/perceiver_pytorch/perceiver_pytorch.py https://github.com/lucidrains/perceiver-pytorch/blob/main/pe..., https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/attention.py https://github.com/CompVis/stable-diffusion/blob/main/ldm/mo...), one of which was credited in the readme by AUTOMATIC1111. For #2, the Anlatan CEO blamed it on an intern (https://i.imgur.com/BFjKG1V.png https://i.imgur.com/BFjKG1V.png). The leak shows that the offending code was committed by the CEO (https://i.imgur.com/aLiA2tr.png https://i.imgur.com/aLiA2tr.png), which doesn't necessarily rule it out originating from an intern (e.g: "send me the code over teams to review and I'll add it") but doesn't look great. From other examples I'd say AUTOMATIC1111 did get a bit sloppy in terms of not following clean-room design regarding the leak, but I'm inclined to give some leeway to a solo developer making a hugely popular public tool for free.