I saw this tweet from someone at the DoD discussing how stupid it was that the USA governement cannot use Llama2 technology but the CCP can. I searched up the use policy and it seems correct. What do you all think??
Llama2 is... not amazing. It's good for small things and it's great that you can do those small things locally. But this honestly feels like an attempt at publicity. "It's so good, DoD shouldn't use it!"
It broadly prohibits "military, warfare, nuclear industries or applications, espionage, use for materials or activities that are subject to the International Traffic Arms Regulations (ITAR) maintained by the United States Department of State." There is no other specific mention of the DoD.
The US can absolutely use it, if they want. US law permits them to violate the license; Facebook can sue to be paid costs ("recovery of his reasonable and entire compensation for such use and manufacture") afterwards, but they can't really prevent its use.
Based on what exactly? The linked article prohibits activities, not entities. Unless the CCP doesn’t use it for spying and military uses and the DoD wants to, this makes no sense.
Believe it or not the differences between the CCP and DOD are not nall black and white.
Take for instance the CCP nuclear policy.
China deliberately has zero nukes ready for launch. Retaliation will ultimately take up to a week as nuclear warheads need to placed on missiles as they are deliberately not attached. This is 100 percent by design giving china time to prevent global obliteration via mistaken retaliation. Additionally
Meanwhile the US has nukes fully ready and armed with deliberate intention to deliver retaliation before enemy nukes even hit their targets. There is no <no first use policy> here.
It's definitely a different and more complicated story for other things outside of war. But Facebook is simply stopping usage related to war.
This makes sense to me. The CCP can't legally be stopped from not using it anyway.
Nathaniel Gleicher, Head of Security Policy at Meta said this
"We have decided not to make Llama 2 available for defense use, as we believe that the risks of misuse outweigh the benefits. We are committed to ensuring that our technology is used in a responsible and ethical manner, and we believe that this decision is in line with that commitment."
Is there a different link that shows they are prohibiting all use by the Department of Defense? The linked page is their acceptable use policy. From that policy it appears it is fine for the Department of Defense to utilize it, as long as it is not for one of the prohibited reasons.
>> 2. Engage in, promote, incite, facilitate, or assist in the planning or development of activities that present a risk of death or bodily harm to individuals, including use of Llama 2 related to the following:
>> a. Military, warfare, nuclear industries or applications, espionage, use for materials or activities that are subject to the International Traffic Arms Regulations (ITAR) maintained by the United States Department of State
The key piece here (IMO) is: "assist in the planning or development of activities that present a risk of death or bodily harm to individuals". There's plenty of DoD activities that are a few steps away from actual conflict engagement. Is helping an engineer move HESCO barriers too close to "activities that present a risk of death or bodily harm to individuals"? What about helping a team learn how to use a weapon or even any piece of equipment within the DoD? Who decides how many degrees from 'these activities' are ok?
In practice you're correct, but the title implies—and the original poster elaborated on this in their comment [0]—that the DoD is banned specifically, while other militaries (specifically the CCP according to OP) are authorized. That is inaccurate—if any military is wholly banned then all militaries are.
Private companies over the military and sovereign nations, isn’t there a famous Zuckerberg quote about that?
Ah yes…
“Companies over countries,” he told me once, as we discussed a blog post about Facebook’s goals. “If you want to change the world, the best thing to do is to build a company,” he added.
I assume they are referring to: "a. Military, warfare, nuclear industries or applications, espionage, use for materials or activities that are subject to the International Traffic Arms Regulations (ITAR) maintained by the United States Department of State..."
Since that's the only reference to "military" I can find. Regardless, if we assume that the US Dept of State does not list the US military on its prohibited list, this would not be an issue.
Even if they do "prohibit" this elsewhere, it would not in fact prevent the US military from using it if they were able to say (to the satisfaction of a US judge) that they had a compelling national security reason to do so.
And if all of that failed, and for some reason the US military wanted to use Llama2, they would just do it anyway, and neither FB nor the judiciary would stop them.
Now, whether or not it would be a good idea for the US DoD to use Llama2, is a different question.
You are correct. The author of this drivel has apparently never read a terms of use policy before. This is pretty much boilerplate policy dating back to the 1980's, if not earlier.
This title is editorialized and wrong—there is nothing in this policy that mentions the DoD. I assume the poster misunderstood this line in the prohibited uses:
> Military, warfare, nuclear industries or applications, espionage, use for materials or activities that are subject to the International Traffic Arms Regulations (ITAR) maintained by the United States Department of State
This is the only mention of any US department, and it doesn't ban them, it references a list of banned activities that is maintained by the US State Department. This means two things:
1. The DoD could use LLaMA 2 for any non-prohibited purpose.
2. No other military is allowed to use LLaMA 2 for military purposes.
A better title would be "Meta prohibits use of LLaMA 2 for militaries", but that would be less exciting because it's a pretty common provision in many terms of use.
> A better title would be "Meta prohibits use of LLaMA 2 for militaries", but that would be less exciting because it's a pretty common provision in many terms of use.
I also don't see the point of such a provision. If an enemy military uses it, good luck taking them to court - if you are at war with a country, they probably won't be paying up any damages.
If your own countries military wants to use it, then the military can force you to hand the tech over if they really don't want to pay for it.
The term is perhaps of use for neutral-country militaries?
My assumption is that like most of these "prohibited uses" it's there to cover Meta's liability for any of these use cases. It's not like someone who's going to use LLaMA 2 for criminal activity will be dissuaded by the ToS either, but it gives Meta a fig leaf to remove their own culpability for whatever ensues.
No company is stupid enough to give out its tech for free to the government when there are billions of dollars worth of contracts for the taking. Doubly so when it comes to the military, which has unlimited money and no accountability in spending it.
Multiple people have posted the same snippet already, and I think it's fairly standard boilerplate in big tech legal documents. For example, I think I remember reading the same line or similar in iTunes' TOS many years ago.
But this stuff would probably get ignored by any army around the world. And realistically if the US military had to seriously use any of these tools to change the outcome of a conflict, are we going to pretend that this line of legalese is going to prevent them from doing so?
This whole page is going to get ignored by anyone who's actually interested in using LLaMA 2 for anything violent or illegal, and their lawyers know it. All of these terms must have a different purpose than actual effective prohibition.
There's lots of ways to read this. The way I read it, Meta is trying to keep Llama from getting "poisoned" with ITAR-controlled data.
Why? Because, oh boy, once a word or sentence or paragraph is classified as ITAR technical data, that little nugget is a bomb waiting to go off in your system. ITAR-classified data can't so much as look at a wire that's outside of the borders of the United States, which, somewhat obviously, throws a cinder block into most systems designs for cloud stuff. And if it's one of these newfangled LLMs, there's no way to take the system apart to get the nugget out. That is a LOT of compute wasted.
There's a reason that GovCloud and other ITAR clouds are awesomely crippled versions of their original incarnations - they have to audit every scrap of every environment and tool available with a fine-f#@king-tooth-comb. A lot of software vendors I know gave up on their GovCloud ports, and just send a team to do the install on-prem. It turned out to be a lot less work than maintaining a special GovCloud build that you have zero idea will work at rollout. If the customer is ok with Docker on-prem, this is an even easier decision. GovCloud storage and emails, yeah, that's fine. Anything more complicated than that, oof - it's going to be its own completely forked build, or else you're going to be going nuts with flags and such.
Now, ok, not going to name names, but there are some vendors providing "AI" in the ITAR space, but, IMHO, I'm extremely skeptical that they have gone through the right channels. I would tread with extreme caution here. Their public faces are quite familiar from some other DoD IT fiascoes from the last few years, and on their DoD LLM thing they've been, at best, disingenuous when questioned about the nature of the waivers they've "acquired". At worst they're outright lying. They seem to be working through informal channels most of the time with highly-placed connections, but that's a very brittle way to do this sort of business. You probably know who I'm talking about here. Anyway, state of nature right now, the safe way for experimenting with LLM in defense is still on-prem.
39 comments
[ 4.3 ms ] story [ 94.8 ms ] threadhttps://www.law.cornell.edu/uscode/text/28/1498
Take for instance the CCP nuclear policy.
China deliberately has zero nukes ready for launch. Retaliation will ultimately take up to a week as nuclear warheads need to placed on missiles as they are deliberately not attached. This is 100 percent by design giving china time to prevent global obliteration via mistaken retaliation. Additionally
https://en.m.wikipedia.org/wiki/No_first_use
Meanwhile the US has nukes fully ready and armed with deliberate intention to deliver retaliation before enemy nukes even hit their targets. There is no <no first use policy> here.
It's definitely a different and more complicated story for other things outside of war. But Facebook is simply stopping usage related to war.
This makes sense to me. The CCP can't legally be stopped from not using it anyway.
"We have decided not to make Llama 2 available for defense use, as we believe that the risks of misuse outweigh the benefits. We are committed to ensuring that our technology is used in a responsible and ethical manner, and we believe that this decision is in line with that commitment."
https://about.fb.com/news/2021/12/metas-adversarial-threat-r...
>> a. Military, warfare, nuclear industries or applications, espionage, use for materials or activities that are subject to the International Traffic Arms Regulations (ITAR) maintained by the United States Department of State
The key piece here (IMO) is: "assist in the planning or development of activities that present a risk of death or bodily harm to individuals". There's plenty of DoD activities that are a few steps away from actual conflict engagement. Is helping an engineer move HESCO barriers too close to "activities that present a risk of death or bodily harm to individuals"? What about helping a team learn how to use a weapon or even any piece of equipment within the DoD? Who decides how many degrees from 'these activities' are ok?
[0] https://news.ycombinator.com/item?id=37174278
Ah yes…
“Companies over countries,” he told me once, as we discussed a blog post about Facebook’s goals. “If you want to change the world, the best thing to do is to build a company,” he added.
So he is being consistent (perhaps foolishly).
Since that's the only reference to "military" I can find. Regardless, if we assume that the US Dept of State does not list the US military on its prohibited list, this would not be an issue.
Even if they do "prohibit" this elsewhere, it would not in fact prevent the US military from using it if they were able to say (to the satisfaction of a US judge) that they had a compelling national security reason to do so.
And if all of that failed, and for some reason the US military wanted to use Llama2, they would just do it anyway, and neither FB nor the judiciary would stop them.
Now, whether or not it would be a good idea for the US DoD to use Llama2, is a different question.
> Military, warfare, nuclear industries or applications, espionage, use for materials or activities that are subject to the International Traffic Arms Regulations (ITAR) maintained by the United States Department of State
This is the only mention of any US department, and it doesn't ban them, it references a list of banned activities that is maintained by the US State Department. This means two things:
1. The DoD could use LLaMA 2 for any non-prohibited purpose.
2. No other military is allowed to use LLaMA 2 for military purposes.
A better title would be "Meta prohibits use of LLaMA 2 for militaries", but that would be less exciting because it's a pretty common provision in many terms of use.
I also don't see the point of such a provision. If an enemy military uses it, good luck taking them to court - if you are at war with a country, they probably won't be paying up any damages.
If your own countries military wants to use it, then the military can force you to hand the tech over if they really don't want to pay for it.
The term is perhaps of use for neutral-country militaries?
The current title here is editorial and opinion
But this stuff would probably get ignored by any army around the world. And realistically if the US military had to seriously use any of these tools to change the outcome of a conflict, are we going to pretend that this line of legalese is going to prevent them from doing so?
Why? Because, oh boy, once a word or sentence or paragraph is classified as ITAR technical data, that little nugget is a bomb waiting to go off in your system. ITAR-classified data can't so much as look at a wire that's outside of the borders of the United States, which, somewhat obviously, throws a cinder block into most systems designs for cloud stuff. And if it's one of these newfangled LLMs, there's no way to take the system apart to get the nugget out. That is a LOT of compute wasted.
There's a reason that GovCloud and other ITAR clouds are awesomely crippled versions of their original incarnations - they have to audit every scrap of every environment and tool available with a fine-f#@king-tooth-comb. A lot of software vendors I know gave up on their GovCloud ports, and just send a team to do the install on-prem. It turned out to be a lot less work than maintaining a special GovCloud build that you have zero idea will work at rollout. If the customer is ok with Docker on-prem, this is an even easier decision. GovCloud storage and emails, yeah, that's fine. Anything more complicated than that, oof - it's going to be its own completely forked build, or else you're going to be going nuts with flags and such.
Now, ok, not going to name names, but there are some vendors providing "AI" in the ITAR space, but, IMHO, I'm extremely skeptical that they have gone through the right channels. I would tread with extreme caution here. Their public faces are quite familiar from some other DoD IT fiascoes from the last few years, and on their DoD LLM thing they've been, at best, disingenuous when questioned about the nature of the waivers they've "acquired". At worst they're outright lying. They seem to be working through informal channels most of the time with highly-placed connections, but that's a very brittle way to do this sort of business. You probably know who I'm talking about here. Anyway, state of nature right now, the safe way for experimenting with LLM in defense is still on-prem.