4 ms·

Thousands of lawyers and billions to spend.

Der_Einzige 2y ago
Again, how do you reliably prove that I used your model if I do a bunch of tricks to "hide" this? i.e. high temperature, exotic samplers (i.e. typicality sampling), using lora/soft prompts/representation engineering, using it as part of a chain on top of other AI, etc I don't care if every lawyer on earth is hired by Meta. Show me evidence that any particular LLM can be trivially fingerprinted based on its outputs. No, that "red token green token" paper on watermarking (https://arxiv.org/pdf/2301.10226 https://arxiv.org/pdf/2301.10226) is not an example of this because of how trivial it is to defeat. Edit: I can't reply to the comment saying "Subpoena" but this commentator seems to think that using any LLM at all is grounds for a court to issue a Subpoena requiring you to disclose which LLM you're using. If this actually happened, you'd see a massive chilling effect. Also, what stops someone from silently replacing the model with a non infringing one the moment someone starts asking questions? I'm pretty sure that most courts aren't capable of getting expert testimony which is good enough to deduce that I silently swapped out my blarg_3.1 model which was made using a 1/3 llama3 merge with something else with gloop_1.5 which is no longer infringing. Like seriously, I again ask, given the idea of courts with warrants and Subpoena's, why should I care about meta's licensing? Edit2: If you're afraid of an employee leaking this info, don't tell your employees. Good thing clever model merging leaves no traces if you delete metadata!
- hodgesrm 2y ago
  Sounds as if you might not have dealt with corporate litigation. There's no need for courts to issue subpoenas. Lawyers can issue them too, and for practically anything. If you don't respond they can issue a motion to compel in court, which if granted means you are legally required to respond with the requested documents. The real problem is that all of this starts to get very expensive if you want to fight it. Meanwhile, all those lawyers need is to grab your text messages, emails, and slack messages. There's no need to look at code. If you have been up to something it's likely to come out pretty quickly. p.s., Hiding your tracks after a subpoena is issued is a good way to end up in jail, at least in the US.
- JumpCrisscross 2y ago
  > how do you reliably prove that I used your model if I do a bunch of tricks to "hide" this? Subpoenas.
  - Der_Einzige 2y ago
    If you live in a glass house, you won't start throwing stones. https://www.forbes.com/sites/alexandralevine/2023/12/20/stable-diffusion-child-sexual-abuse-material-stanford-internet-observatory/ https://www.forbes.com/sites/alexandralevine/2023/12/20/stab... https://www.airforcetimes.com/news/your-air-force/2024/03/26/another-airman-probed-by-fbi-for-allegedly-leaking-intel-on-discord/ https://www.airforcetimes.com/news/your-air-force/2024/03/26... (i.e. as source of where LLMs get their classified data from) If you scrape a large enough part of the internet, you're naturally going to get extremely illegal training data that you won't effectively filter out. I guarantee you that at least a tiny bit of highly classified information was not filtered out of most LLM training data (it wasn't found during the filter step), and it's quite remarkable that this and the above revelations have not led to anyone being Subpoena'd or related in regards to it. So no, I think that folks will be literally the opposite of litigious on this issue. You want to play that game Zuck? Let's see what happens when I hire my researchers to find the dirt on your models dataset. We will see then who "settles out of court".
- ensignavenger 2y ago
  An employee (or a hacker) leaking the fact, maybe even leaking details, is all it takes to get the bloodhounds called on you and now you are subject to discovery. Sure, you can lie, but if you are doing anything large enough to get attention, you are exposing yourself to possible liability.

Again, how do you reliably prove that I used your model if I do a bunch of tricks to "hide" this? i.e. high temperature, exotic samplers (i.e. typicality sampling), using lora/soft prompts/representation engineering, using it as part of a chain on top of other AI, etc I don't care if every lawyer on earth is hired by Meta. Show me evidence that any particular LLM can be trivially fingerprinted based on its outputs. No, that "red token green token" paper on watermarking (https://arxiv.org/pdf/2301.10226 https://arxiv.org/pdf/2301.10226) is not an example of this because of how trivial it is to defeat. Edit: I can't reply to the comment saying "Subpoena" but this commentator seems to think that using any LLM at all is grounds for a court to issue a Subpoena requiring you to disclose which LLM you're using. If this actually happened, you'd see a massive chilling effect. Also, what stops someone from silently replacing the model with a non infringing one the moment someone starts asking questions? I'm pretty sure that most courts aren't capable of getting expert testimony which is good enough to deduce that I silently swapped out my blarg_3.1 model which was made using a 1/3 llama3 merge with something else with gloop_1.5 which is no longer infringing. Like seriously, I again ask, given the idea of courts with warrants and Subpoena's, why should I care about meta's licensing? Edit2: If you're afraid of an employee leaking this info, don't tell your employees. Good thing clever model merging leaves no traces if you delete metadata!

hodgesrm 2y ago
Sounds as if you might not have dealt with corporate litigation. There's no need for courts to issue subpoenas. Lawyers can issue them too, and for practically anything. If you don't respond they can issue a motion to compel in court, which if granted means you are legally required to respond with the requested documents. The real problem is that all of this starts to get very expensive if you want to fight it. Meanwhile, all those lawyers need is to grab your text messages, emails, and slack messages. There's no need to look at code. If you have been up to something it's likely to come out pretty quickly. p.s., Hiding your tracks after a subpoena is issued is a good way to end up in jail, at least in the US.
JumpCrisscross 2y ago
> how do you reliably prove that I used your model if I do a bunch of tricks to "hide" this? Subpoenas.
- Der_Einzige 2y ago
  If you live in a glass house, you won't start throwing stones. https://www.forbes.com/sites/alexandralevine/2023/12/20/stable-diffusion-child-sexual-abuse-material-stanford-internet-observatory/ https://www.forbes.com/sites/alexandralevine/2023/12/20/stab... https://www.airforcetimes.com/news/your-air-force/2024/03/26/another-airman-probed-by-fbi-for-allegedly-leaking-intel-on-discord/ https://www.airforcetimes.com/news/your-air-force/2024/03/26... (i.e. as source of where LLMs get their classified data from) If you scrape a large enough part of the internet, you're naturally going to get extremely illegal training data that you won't effectively filter out. I guarantee you that at least a tiny bit of highly classified information was not filtered out of most LLM training data (it wasn't found during the filter step), and it's quite remarkable that this and the above revelations have not led to anyone being Subpoena'd or related in regards to it. So no, I think that folks will be literally the opposite of litigious on this issue. You want to play that game Zuck? Let's see what happens when I hire my researchers to find the dirt on your models dataset. We will see then who "settles out of court".
ensignavenger 2y ago
An employee (or a hacker) leaking the fact, maybe even leaking details, is all it takes to get the bloodhounds called on you and now you are subject to discovery. Sure, you can lie, but if you are doing anything large enough to get attention, you are exposing yourself to possible liability.