I'm curious: Have you seen indications that major militaries and politicians believe AGI, rather than special purpose ML for military purposes, is important for national security? I'm really not sure whether this is true, or whether military and political leaders think it's true.
There is a difference between sharing the tech hype, and risk management. Why would our political and military leadership not be interested in this sort of tech in the modern world? If it doesn’t work out, then it doesn’t work out, but if it does, then they’ll want in on it. Aside from that there is the mass surveillance angle on it. We recently had a nice scandal of sorts here in Denmark where the chief of our secret military or whatever you’d call this was arrested by our secret police because he may or may not have shared secrets about how the US spies on us. It was something that even included charges against our former minister of defence possibly leaking things, something which could have seen him twelve years in prison. Luckily our courts saw it as a political matter and refused to let it run in closed proceedings which led to charges being dropped.
The matter of the leaks were very “Snowdeny” in that it’s possibly that parts of our own government and our secret police share all Danish internet traffic with the NSA, who then in tern share information with our secret police. Which meant that our secret police could do surveillance on us as citizens through a legal loophole, as they aren’t allowed to do they directly, but are allowed to share surveillance information with the NSA. Part of this information comes from the giant American tech companies as well. Despite their promises to not share the data they keep for you. I know it’s sort of crackpot sounding, but between echelon, Snowden and the ridiculous amounts of scandals, I think it’s safe to assume that the American military wants in on LLMs and monitor all the inputs people put into ChatGPT and similar. So for that reason alone they’d want in on things.
Then there is how the war in Ukraine has shown how cheap drones are vital in modern warfare, and right now, they need to be manually controlled. But what if they didn’t? Maybe that’s not obtainable, but maybe it is.
Then there is all the other reasons you and I can’t think of. So even if they don’t believe it’s eventually going to lead to an AGI, or whatever else the hype wants, they’re still going to be interested in technology that’s already used by so many people and organisations around the globe.
I'm sure they're interested in it, but I'm uncertain that they view it as a promising and critical enough capability to push for a higher priority when weighed against other interests.
For instance, neither of your examples - surveillance or automated drones - has anything to do with AGI. They don't need LLMs to do mass digital surveillance; they already do that and were doing it for decades before LLMs were a twinkle in anyone's eye. Sure, they'll try to tap into the user data generated by chatgpt etc. (and likely succeed), but that's not a different capability than what they're already doing. And automating drones - which, by the way, this is not future technology as you seem to imply, it's here today - is a special purpose ML system, that maybe benefits from incorporating an LLM somewhere, but certainly isn't pinging the chatgpt api!
But sure, you're exactly right at the end, I have no idea whether they see other angles on this that are new and promising. That's why I asked the question, I'm very curious whether there are any real indications thus far that militaries think the big public LLM models will be useful enough to them that they'll want to put a thumb on the scale to favor the companies running them over the companies that make their bucks on copyrighted content.
Wiretapping vast amounts of data on the internet is quite cool, but actually sifting through all that data is the really difficult part. Right now intelligence services are probably looking at lots of false positives and lots of dots they can't connect because the evidence is just too dispersed for a human or a current-generation system to make sense. LLMs could enable them to make the analysis more targeted and effective.
But for all we know intelligence services could be using LLMs for years now, since they are usually a few years ahead of everybody else in many regards :-)
This is not the new capability that LLMs have pioneered. It's true that it is difficult to sift out signal from the noise of a vast data trove, but it is difficult in a way that people have been getting extremely good at since the late 90s. What you're describing is a Google-level capability, and that's truly a very complex thing not to be downplayed. But it's a capability that we've had and been honing for decades now.
I'm sure language models and transformer techniques will be (or more likely: already are) an important part of the contemporary systems that do this stuff. But I'm skeptical that they care much about GPT-4 itself (or other general models).
I'm not skeptical about whether they think it is useful and an important capability to incorporate ML techniques into their systems, I'm unsure how much utility they see in general (the "G" in AGI) models.
> LLMs could enable them to make the analysis more targeted and effective.
How? I'm not trying to be combative, I genuinely am curious if you have an idea how these things could be usefully applied to that problem. In my experience working in the information security space, approximate techniques (neural nets, etc.) haven't gotten much traction. Deterministic detection rules are how we approach the problem of finding the needle in the hay pile. So if you have a concrete idea here that could represent an advancement in this field.
I guess my next question is how many needles do you find and how sharp are they? Detection rules would filter out most of the noise, then something like an LLM would do a post filter for intent analysis to rank relative risks for human intelligence to look at.
I suspect this would disincentivize operators to take care in the way they write their detection rules, and the nondeterminism of the LLM would then result in false negatives. So the rate of growth of the needles set would increase, and the analysts would be getting lower quality information mediated by the LLM.
In a world where false negatives--i.e. failing to detect a sharp needle--are the worst possible failure mode, approximations need to be handled with exceeding care.
I'd guess leaders are thinking more in terms of national capacity to create more advanced technologies than geopolitical adversaries. If US policy shakes out in a way that protects copyright holders at the expense of AI innovation, I think it's apparent that the end result will be that our rivals will both violate copyright and beat us to building widespread expertise.
> I'd guess leaders are thinking more in terms of national capacity to create more advanced technologies than geopolitical adversaries.
I think there's a strong argument that they should be thinking in those terms, but I'm a lot less convinced that they do usually think in that way.
Or more charitably, they have the responsibility to balance current interests against future interests. And this isn't just a tricky thing for democracies, dictators also have to strike this same balance, just with different trade offs.
But in this case, for the US, it honestly isn't clear to me that policy makers should favor the AI side of this tussle. I think culture has been among the, if not the very, most important export of the US for nearly a century, and I think favorable copyright treatment has been at least part of the story with that.
Maybe that whole landscape is different now in a way that makes that whole model obsolete, but I think it's an open question at least.
What it seems to me from the milieu of everything I've read and heard (that is: I can't cite examples, this is an aggregate developed from hundreds of articles and podcasts etc.) is that there is already an "AI" arms race underway, but that it has more to do with specialized ML systems than with consumer LLMs.
But I'm not really in the loop, and maybe OpenAI really is more important to the US DoD than Disney (as a stand-in for big copyright-based businesses generally) is to the politicians they donate to. But I dunno! That's why I asked the question :)
I would be more intrigued by the national security angle of this if copyright holders were going after, say, Palantir. But I just don't know how important they see these language models as being, or how interested they are in OpenAI's mission to discover AGI.
It mostly doesn't matter if the military wants specialist systems, in the long run generalist systems tend to win in power and adaptability.
Some of this may be a misunderstanding of what modern militaries do, if they are shooting guns there's already been some level of failure. Massive amounts of war gaming, sentiment analysis, and propagandizing occur, see the RAND Corporation for more details on the military development of algorithms and artificial intelligence.
Yeah this makes sense. Maybe RAND publications will indeed give me some insight into my question.
But I also buy that there is a lot of overlap between military work and any other kind of white collar work, which LLMs are definitely useful (but not revolutionary) for.
US media is a huge cultural influence. It takes a ridiculous amount of mindspace globaly. However, with youtube&tiktok this seems to be changing - the most important influence is not from Hollywood but ”random” youtubers. So, ”content” is waning in influence for sure, unlike hardcore national security things like US dollar, the carrier fleet or ballistic nukes. Or AI.
On Youtube creators who are native to the anglosphere still have a big advantage. TikTok is really the big equalizer. With AI voices being the norm, nobody cares about your accent.
Guess it depends on what you mean by "advantage"... English-language channels are a dime a decillion. But if you started posting content in Tagalog you'd find yourself gaining traction with an audience that doesn't have as many alternatives.
(1) Do you think "developing AGI" a realistic, achievable goal? If so, what evidence do you see that we're making progress on the problem of "general" intelligence? Specifically, what does any of that have to do with Large Language Models?
(2) Are there any "national security" applications of Large Language Models that you're aware of?
It seems to me that it would be a very difficult case to make that the national security impact from allowing the rule of law to erode would be somehow outmatched by the (speculative) wager that somehow LLMs have some relevance to the national security. It would be an even harder case to make that any of this has something to do with "general" intelligence.
If you manage to put a bunch of listening devices at a place you're moderately interested in, a cafeteria at an enemy base for example, you might end up with literally hundreds of hours of conversations, most of them completely uninteresting, but a few that might possibly contain nuggets of information of the utmost importance. Listening to all these conversations requires resources. This is even more difficult if the people there speak in jargon, in their own language, and nobody but an expert in the subject can determine which conversation snippets are significant.
If you have good LLMs, you can run all your recordings through extremely high-quality speech recognition and then use something like Chat GPT for summarization, classification, finding all mentions of the nuclear reactor in <place> etc. Same goes for satellite image analysis.
I think we'd need to see these things get a lot more reliable for them to be viable in this use case. This seems like a "leaky net" as opposed to some more deterministic strategy (e.g. grepping large lists of keywords, or parallelizing the task over thousands of human analysts). When you're looking for a needle in a haystack you need to inspect every leaf and stalk.
So should we put copyright through the shredder on the wager that somehow generative techniques will find applications for mass surveillance?
As for 1, pass an image to a multimodal LLM and simply ask 'what is going on in this image'. Robot LLM models are already turning this in to actionable data they of which they can interact with the world. As in you can send a Robot into a room it has not been before and tell it "bring back a sock, a blue one not a red one" and get an actionable response with a higher degree of success. This takes some degree of general intelligence (though maybe not human level).
Well the real test of all this stuff is "what can I use it for?". And I can sort my own socks, so that's not super compelling ;). More seriously, the real world is complex.
Let's say I want to replace the forklift operator at my local lumberyard with a robot forklift that can ostensibly outperform a human employee. Even if there is some magical AI program which could theoretically drive the forklift around, identify boards by their dimensions, species, dryness, location, etc., there's a whole bunch of sensory problems that a human body solves easily that are super hard to solve in the environment of a lumber yard. There's dust, rain, snow, mud--so if you're relying on cameras how will you keep them clean? You can't visually determine how dry a board is, you have to put a moisture meter on it and read the result. My point is, even if you have a "brain" capable of driving the forklift you still have a massively complex robotics problem to solve in order to automate just the forklift. And we haven't even begun to replace the other things the operator does in addition to driving the forklift. He can climb out of the forklift and adjust the forks, move boards by hand, affect repairs on equipment, communicate with other equipment operators, customers, etc.
Good luck replacing him in a cost-effective manner.
This is an issue of 'mechanical intelligence' being hundreds of millions of years old and 'higher intelligence' being pretty new on the evolutionary spectrum.
And the AGI will keep you around as a dexterous 'robot' while supervising your thoughts to make sure you're keeping in line I guess, while day after day cranking out more capable robots in which to replace you with eventually.
How do you unplug the power on your iPhone? You can't even take the batteries out. But ya, if you assume it will take massive amounts of power to run an AI in the future its easy to see your logical error here.
Developing AGI, as an abstract idea, is a matter of national security. That doesn't mean people are willing to accept the real-world consequences of it. Especially when it could affect them financially.
Additionally, I'm not even sure the US is capable of having national priorities at the moment. The Congress has become incapable of making decisions. While the executive and the judiciary branches have stepped up to compensate, they tend to handle each issue separately without any general direction.
Apple could buy most of the NYT, RIAA and MPAA companies combined with petty cash. The big ones are Disney and Sony with a combined market cap about 250b. Microsoft alone is worth over 10 times that.
MPAA and RIAA are not joint-stock companies, they are specialised trade unions to enforce intellectual property. They have no shares to acquire. If you mean acquiring the individual members, that would encounter an enforcement from your favourite antitrust enforcement commission.
Honestly I've always wondered what would happen (and how much the entertainment world would change) if a company like Apple, Google, Microsoft, etc did just that. Or heck, if it turns out you need the rights to train LLMs and its easier to do that with public domain stuff, they just flat out bought half the entertainment industry and assigned everything to the public domain. Every Disney work every for example.
No-one is going to buy a major media company and then throw the rights into the public domain. What they would do is buy the rights and then sue all competitors in the GenAI space.
In the US, this isn't possible. There is no legal mechanism for putting things into the public domain outside of the expiration of the term of copyright. The best you can do is to promise not to enforce your copyright.
