Paul Scharre explains the global AI arms race

America's defense strength comes from the people who build it, welders, machinists, and others who produce the equipment our military relies on. And while demand is rising, the number of skilled workers hasn't. But the GE Aerospace Foundation and the company are stepping up. The foundation is launching a $30 million program to train 10,000 workers in advanced manufacturing. The program will expand instructor ranks, provide modern equipment, update courses to match industry credentials, and lower cost barriers, allowing graduates to step directly onto the line or into the hanger. These efforts are helping grow the skilled workforce the defense industrial base needs. Learn more at gearospace.com/workforce. On this hour 200th episode of the podcast, my colleague Patrick Tucker and I welcome back our very first guest on the show. Paul Sharre is the executive vice president at the Washington based think tank, the Center for a New American Security. He's also a former Army Ranger, and the author of several books, including four battlegrounds, power in the age of artificial intelligence, and Army of none, autonomous weapons in the future of war. Two years ago, Time Magazine named Paul one of the 100 most influential people in artificial intelligence. Paul Sharre, it's been seven years since we first spoke. Welcome back to Defense One Radio. Well, thanks for having me back and congratulations on such a successful 200 episodes. Well, thank you. Yes, it's been quite a lot. We've done a lot of very formats, and I'm kind of proud of the work that we have done. So thank you for that. I really appreciate it. And thanks for being a part of the very first day. All right, so hey, actually going back to that time, back in 2018, to the extent that we at Defense One talked about artificial intelligence, it was largely through the lens of a, you know, pretty quiet pentagon program at the time, a program with Google known as Project Maven. And as I understood it, the AI helped scan terabytes of drone and surveillance footage over the Middle East, for example, to assist in locating targets the US military could strike during the war on ISIS. Now one of the things I'd like to do and our discussion today is to tell the story of how the technology has evolved since then. So to get us started, what more do you think it would be helpful to note about the understanding and use of AI in its infancy back during the first term of President Trump, and you know, perhaps even earlier, if you like, we're basically just getting started and then we'll kind of form a contrast with where we are now. Yeah, so I think like traveling backwards the time to these ancient times of 2018, what was just the technology was different, not just that it was not as capable, but the AI systems that people were using at the time were qualitatively very different than large language models. And the deep learning revolution had kicked off around 2012 with some progress on image net, image classification, data set and challenge, using deep neural nets, so deep learning, using neural networks, to train neural networks on originally images to then build classifiers to identify objects, to identify that there are objects and then classify what those are. And that kind of brand of AI classifiers was really the bulk of AI systems, what you would call narrow AI, that we saw in that time frame. And the huge sea chains that we've had over the last several years since then has been running originally with GPT-2, which I think among the AI nerds, drew a lot of attention, less so, you know, the global scale that CHAT GPT did. A few years later, the development of large language models and increasingly general purpose systems. And it's not just that, okay, well, language models can generate text and that's interesting and useful. It's this really significant expansion of the scope of what AI systems can do from doing like one thing and doing it sometimes even better than people, whether it's identifying images or think back in the 90s, you know, something like IBM's Deep Blue, which didn't use deep learning, but was able to play chess better than human speed, Gary Kasparov. To today, much more general purpose systems that could do a whole wide range of tasks. And you know, we can get into a whole thing of like, well, how do you compare them against humans? And that's an interesting discussion, but it is a huge qualitative difference from, say, seven years ago. I want to kind of transition now to, of course, your most recent book and that really kind of begins to set the stage for the bigger issues that animate the way that I think our audience interacts with this issue. But so in 2023, you published your book for Battlegrounds about the AI tech race, which is kind of currently playing out largely between the US and China. So what can you share from that book and what have you learned since about how we understand China is using this technology and its various applications? Yeah. So I think, you know, what motivated me in writing this book was discussions that it started under way in Washington. Honestly, around the 2017, 2018, 2019 timeframe, I kicked off in a big way with the National Security Commission on AI. I'm about to say to you, okay, the US and China on this AI competition, we have to win. Which yeah, everybody watching the degrees on and generating a lot of ideas coming out of different quarters of think tanks, government and commissions and other things about what we need to do. I wanted to take a step back and just try to understand what are we competing in? Like how what would it mean to compete in AI? What is the nature of that competition? And a lot of folks have compared AI to another industrial revolution. Okay, if you buy that comparison, I think there's some validity to that. We saw that the industrial revolution led to major shifts in global power as some countries were able to capitalize on technology, sooner and then translate that into not just economic but also military power, but also that the key metrics of power changed. Whole and steel became key inputs of national power and manufacturing capacity became a key measure of national power. Oil became a geos Strategic Resource. The countries will identify wars over. And so I wanted to understand what was the nature of that competition in the age of AI? What are those things that we should be competing over? And at least this is tapered off a little bit, but there was a period of time where like this idea of data is a new world was all in vogue. And all these articles in the economists of the places, data is a new world. And like you got to you got to wave about nine months later of all these articles of data is not the new world. Data is not like oil. It's like, obviously, it's not the same thing. But oil is valuable is data valuable. Yeah, like what's the similarities with the differences, but I wanted to take a step back I think more broadly. And so the data is a key component of this competition, but clearly computing hardware is one. And that's one that we've seen up heat up significantly since I published the book. In fact, the export controls I have chips were coming down as the book was like in the final stages of publication and much to the publishers dismay. I was trying to jam in updates and texted to new updates of US export controls, which might have seemed arcane. But I was like, this is really important. And that's become a big centerpiece of competition. And then also talking about the competition of talent, which is of course really critical. And then in institutions and basically the mechanisms inside the US military inside the PLA to take this technology, it's really coming out of the commercial sector and pull it into the defense sector and then transit that to military advantage. I remember this. This was at the beginning of the period where big data was like a hype term. And my publisher back in 2014 was like, everyone was talking about big data. And I'm like, yes, but you have to talk about basically the three things that are AI, which is just compute data and math. And the math, there are, I mean, people coming up with new interpretations of ensemble models all the time. But it's really just derived from basis, just really the alternative to basic regression statistics that allows you to constantly update. And all of these, and every new algorithm has that as it's like, the data is, well, this is the thing where maybe life comes in. This is the entire point of it, who had a activity across the Middle East, where the United States was collecting just hundreds and hundreds of hours of video footage. And that would go to analysts and analysts would sit there and try to do like pattern of life and figure out which of the many people there were surveilings, the one who's going to put 90 days somewhere. And that is very difficult and takes a long time and takes a lot of human resources. So there was a ticking queue and sort of thing that was created to deal with the data. And that was later turned into big. And the compute thing is also really critical now. This is the dimension that AI competition is taking in terms of like a material competition. for stuff that exists on this planet. But I also think that the compute issue is one that we should, the North American has a very particular sort of stake in it that is different from the way that great big consumer facing foundation models understand compute. So they're building up massive data centers or getting lots of subsidies. They're trading like at $100 billion between themselves. They're mostly unlike data center stuff. And that's because of a premise that more compute and more data will help you just scale up the performance of the foundation models. And while they did scale up faster than many people appreciated faster than even more's law, there's some suggested that now that scaling, just of that particular way of developing AI through large language models and stuff has plateaued. But you're not seeing that in the same way. And you also see that there is a finance supply like GPUs. But Google has endured part of the reason there was a massive sellout for the video probably reason Peter Tilltsall about all of his stuff is because Google has come up with an alternative. So there's going to be innovations there that keep pace that turn that into a new fight every day. But the thing about the compute aspect of AI is that you have a situation right now where there's really large public companies, these foundation models that provide a service that millions of people all around the world are constantly using. They're kind of sucking up a lot of the attention. They kind of are crowding out a entirely different schools of research areas and available GPUs. They could go to a different version of AI. Different-- that's very critical for the military. But the public doesn't really use it all. It's like a small AI. It's like on platform not necessarily connected to big enterprise cloud, et cetera, et cetera, AI, which I call small AI. And to me, it's something that right now, I don't see the kind of dealing with as well as they should. Because actual warfare is going to lead a combination of both of these types of involved, poking way in on this, because these are much more premium and expert on it than I am. But future warfare is going to require both of these types of AI. We're going to need distributed high autonomy that is not constantly connected to a cloud, that exists on platform C3P, that's your C3P OAI. And you're going to need massive decision assistance that are also working through big problems of logistics and supply all of the time. We're working scenarios for what's dynamically changing on the battlefield. Adalizing all of the dead is collected off about everything on the battlefield and turning that into useful updates. And at some point through a gentica AI, actually like orders. You need them both. And the great big, now Silicon Valley megafauna, foundation model companies are kind of sucking some of the innovation space. They're sucking that the researchers are kind of determining the entire field of AI research around them to a certain extent. And they're sucking up a lot of chips. And that's actually a problem. But I think that the Pentagon has to deal with, because there's done a lot of consumer space right now. There's not a massive market for a small AI that they should be. There will be hopefully in the future. No, I mean, I think this is such a critical point that one that you're making, which is that both of these futures are real. And you'll hear this all the time. Like you were saying, from companies, oh, the future is big models and ever larger and larger and larger and more compute intensive models. And that's the future of AI. That's what we're hearing from the leading labs. And then a lot of other companies that can't compete in that space. No, no, no, the future small models. It's too expensive. And you can't do that. And like, the reality is both of those things are true. There's some fascinating work done by the AI research group Epoch that's tracked a lot of the quantitative measurements in both scaling frontier models, but also in the sort of miniaturization, if you will, of models into more efficiency. And not only is there this exponential growth in the amount of compute being used in frontier models, and not just in training, but of course, increasingly in the inference side as well, as we're saying more compute used for inference. But there's also this pattern that we've seeing, which was up in the data, of exponential improvements in the efficiency of models. So once a model comes out, so like, GPT-4 hits the shelf. OK, so at the time, that was the most advanced system. Then what we've seen in the data is nine months later, there'll be a language model released by that company or somebody else that is able to achieve the same level of performance at half the amount of computing power. And then nine months later, after that, another half. So like, the amount of computing power needed to achieve the same level of performance is having every nine months. I need if those numbers move around a little bit. The point is you're seeing both of these trends. And so like, the frontier of AI is needing more and more data and compute. But at any given moment, the amount of computing need for certain level of capabilities drops really fast, makes it much more accessible. And so if you play that forward, you can envision a world where there might be, if this trend continues at the frontier, there's really expensive models to train and to run that are doing really exquisite types of intelligence, maybe in scientific research that are incredibly valuable. But on the other hand, we might see the commodification of sort of like routine data data intelligence. It's good enough for a lot of what people are doing where the cost of that kind of intelligence drops pretty close to zero. Much like if I get on Google search, like I don't have to pay for that, like that cost of just accessing Google search information on the internet is like really, really well. And so I think both of those things are real. And then there's a question of, OK, for an organization like US military, like how do you harness all of those? Because you want it all. And you want to feed it to warfighters and then let people play with it. And I'd figure out how do we turn that to warfighting advantage. Yeah. And I think it also speaks to right now the business model of some of these Android companies. And when you want to understand what an AI bubble looks like, it looks like companies that are based their current valuation on reaching a sort of performance point and having lots of like data-based agreements and new like data center agreements in play. And that's the way they justify their current valuation before they IPO or whatever. When in fact, yeah, to your point, like especially as the trend moves as it obviously is towards open source models, then that becomes-- I think you'll see all of those of you are just talking about really accelerate, which is probably a very good thing for the United States military. But does suggest that there's some of the most innovative players right now really are in a position where there's sort of obliged to fight against actual innovation? And yet eventually, you have to in order to keep going sort of about that. So that's when OpenAI began to-- after DeepSec came out, which was basically a open source sort of rip off of them, this is when they launched into an entirely new kind of direction where they were going to part of the things to say and be in national labs. And another partner with the military to do exactly this, like custom make-- give them weights, give the military, give Los Alamos and say, be in national labs, the weights to their models, and then work with people on site that have very particular problems that are very complex. The US military is a great example of that. And then work with them to run all of that on whatever that customer brings in terms of an energy capability, energy source, and the available data center. So they're actually making their big AI smaller too now. But it's only as small as the military. So I see it's a fascinating space right now. But you see that a lot of the way we think about some of this stuff at this moment, especially in terms of the valuation of some of these companies, is so transitory. It's moving on just so quickly. Can I wonder point out just two recent headlines that is to this point, and I'll let Paul or whoever wants to go from here. But their headlines that I saw in the last two weeks from late November in the economist, investors expect AI used to soar, but that's not happening. And just last week in futurism, the number of people using AI at work is suddenly falling. So I find those sort of trends kind of interesting, of course, in light of our conversations here. And of course, we've got two different paths that we're talking about-- military and civilian. And these are, of course, civilian applications. But I find those fascinating. Yeah. Yeah. I remember back when in 2002, everyone was like, God, the internet is nothing but a weird site for you to buy second-hand, pesdispensers and stuff. This hype thing is crazy right now. I'm just-- I really get it bullshit. I'm not even going with it. And yeah, it turns out that people actually did figure out new stuff to use it for. But what I like about what the B&D is doing right now is they, under research and engineering, they are moving out with new-- basically, just tools that are available to everybody so that people can figure out how to build a relationship with AI on a very, like, day-to-day person and person to machine thing. And that's kind of unique to the US military's approach to AI that gives me-- and I hopefully the free world's reassurance about the way the US is developed and uses AI military constructs as opposed to China or Russia on a fixed date. Yeah. I mean, I think, like, if you pull a thread on what you were talking about here about this rapid pull oferation of the technology, whether it's deep seek or-- or some other kinds of models, you know, like Boba also releasing models on China. To me, one of the takeaways is that even if US labs stay right at the forefront of this technology, Chinese labs are like single-digit months behind. And whether that's like two months or nine months, I'm not sure that at the end of the day, it translates that lead alone, transit to a mean, familiar advantage. If the Pentagon is, share it, I believe, five years behind the frontier, right? And so my takeaway is, yeah, great that we have US companies in the lead and we want to be building out infrastructure and the energy supported, and that's all good, but we've got to find ways to rapidly translate that technology into some kind of actual military advantage. And that, you know, putting it in a hands of warfighters is really great. I'm glad to see the military moving out on that, but it requires this, either, I mean, one of the things that's very clear looking at historical examples of disruptive technologies is giving people the widget is not enough. That it requires people to figure out how to use it effectively, it requires people to transform their military operations around that, oftentimes in ways that are culturally uncomfortable, and that's going to be hard. And I think that's a longer term process, and it's one that, I think we should be realistic about the fact that we are competing on a pretty level playing field against the PLA. And like in order to win that competition, we're gonna have to find ways to actually be innovative and bring in the tech and be disrupted inside the US military in ways that, you know, sometimes gonna break channing people are gonna like, but that's what's gonna take to maintain a military advantage. America's defense strength comes from the people who build it, welders, machinists, and others who produce the equipment are military relies on. And while demand is rising, the number of skilled workers hasn't, but the GE Aerospace Foundation and the company are stepping up. - So let me, if I may piggyback off that into the one setting where it is very clear, there is both in terms of like the small kind of on platform AI. And it's also even at this point, like larger uses of AI, and that is the opinion battle field, where they, Ukraine and now Estonia are moving out on a new concept of like government that employs agente AI. But they're basically redesigning aspects of all government processes realizing that civil defense and like, you know, the resilience of normal day to day life, this kind of function is like an essential thing that's an aspect of security, especially if you're under attack. And so they're going to use large scale and agente AI to like do government service in a lot of ways, which is not a hard military thing, but these connected to security. It's not something we could do in this country because people have really bad feelings or sentiment towards government and government, like, but they have in the Soviet and the Ukraine a very different relationship with the government. And I think that's fascinating, but on the other end, the like, here's a good example. Like last couple weeks ago, the Pentagon and SETCOM announced that they had deployed a squadron of Shaheed kind of knockoff drones. And these are, you know, high-moveable, basically one way to tackle it, but they're moving. They're actually, they have on board an intelligence and they're, we wouldn't use them, not just for strike, but for ISR. But the fact they were now doing like copycats of Iranian equipment. We wouldn't even have access to, if some like a bunch of guys in the battlefield of Ukraine hadn't picked that thing up, then reversed engineered it themselves and then created a specific drone that's specific for like defeating Shaheeds. And then they gave this to us. And now, you know, we're crowning, we're crowning as though this is the hallmark of innovation that we copied in Iranian design. So like to get to your point about why it's so important to actually have real use cases, real deployments, not just, you know, play school in like relationship with AI. How do you, when you look at how it is used right now in the context of your brain, particularly by your brainians, is the US paying attention to those lessons well enough? How do they better exploit that relationship in a way that is favorable to both of you, how do you actually bring value to the relationship by the little exploits for this bad word for it? And where do you see all this like weird geopolitical mess that are a problem that we're conversation that we're having right now about Ukraine? And we don't have to get too much in the politics, but are we in danger of losing or damaging relationship that is truly critical for us to actually innovate in this field? - Okay, well, there's a lot to unpack there. I mean, let me take at least the first chunk of that and then we can circle back on it. I mean, I think I think sharing a reaction to this point about us basically doing a knockoff of these Iranian drones, which is on the one level. Okay, great, we're learning from others and we should be willing to take less to think we can find them, but also, like, come on. I mean, we seeded the advantage here. And the whole, I'm glad that we see the US military and the blogger kind of US defense ecosystem waking up to the potentials of drone warfare based on what we've seen in Ukraine. That's better than people completely ignoring it. But we're way too reactive. Way too reactive. All of the things that we're seeing unfold were not only foreseeable, they were foreseen by many people, including yourself, and other extra by Coroids, many other people who were riding on this like a decade ago and talking about the potentials of drone warfare. And I remember these conversations 12, 13 years ago, where you'd have a couple of voices in the wilderness saying like drones in the future, this is coming and hearing from senior military officials, just this really narrow conception of drone. Well, you know, in order to do drones, you need satellites, not everybody, it's like, that's how the Air Force does it. Does that grow? Well, Army did it with their Gragel drones back then. Like, that's one way of doing it. You know, you know, all this intel analyst doing the pet on the back end. Like, no, that's one way of doing it. That is not the only way of doing it. And I think you see that repeated, like I see the repeated again in conversations within the US defense ecosystem about Ukraine. And I don't want to sound too cranky about this, but you don't get it, you know what I mean? It kind of makes me mad, right? Because it's like, because at the end of the day, we're putting our military advantage risk. What do I hear people say is that, well, like drones are a big deal when you get these drones. And then the response all here is like, well, you're not gonna fly these quadcopters across the Pacific. Look, obviously not. Obviously not. Okay, then you might fly them across the Taiwan Strait. You could have, you could put really small drones on Taiwan. Okay, but also the wake up call should be that we are, the system, the Pentagon, the services, you know, chunks of industry are thinking about the future of warfare far too narrowly. And we need to broaden our aperture for how we think about this technology and what it's capable of doing. And we need to lean in to the possibilities here. And it may not be that particular platform that's relevant in future worst maybe. But the idea that like a triple mast doesn't work, like no, it does. We can clearly see that it does. Drones are gonna be valuable in the future. We wanna be at the forefront of that. And I think we need to be much more willing to explore capabilities and concepts of operation that are outside the comfort zone for the services. And Newfranc is an example of they have to do it by virtue of necessity. And that's part of the reason why the relationship is like really important. But one of the other things that Newfranc sort of teaches is that there are ways to apply, I'm sure a large foundation model to this problem or not. There's no such thing as air dominance over Newfranc. There's no air superiority, which is why it's stopped. It's why it's like in this world where one knows and drones did that, just the massive consumer drones did that. And that's because Newfranc figured out how to 3D print them close to the frontline. But Russia also has a massive supply chain from China that allows them to continue with that. And that is resulted in this word stalemate where there's no air superiority. So what are the things when we look at the specific, specifically, in terms of how you would apply AI to that theater, taking everything that we've learned about the specifics of the Ukrainian battlefield. How do we then apply those in a different context to create long range attributable stripe, which is what we're trying to do and that they come trying to do that under a program. But there's other things that they, that maybe they're not doing. What are your thoughts? Yeah, I mean, so actually I want to take a moment, just take a beat on one issue that I think is really key that comes out of Ukraine war, which is not about a specific, it's about time. So when I was in Ukraine last year and I met with their drone operators and drone developers and the head of their drones and their armed forces, one of the things that I repeatedly heard from them was the importance of them compressing this time cycle between war fighters that's happening on the front lines and the drone developers. And we spoke with the head of their drones for their armed forces and it was very, I mean, we're all lost, they were pretty critical of US friends. Yeah, they think it's junk. They think our stuff is junk. They're like, we will take them because we will shoot them at the adversary, but this is not exactly the best drone I've ever seen. Right, they didn't like it. And so my estimate is what's wrong with the US drones? And the feedback was fundamentally that the US for a variety of reasons is not able to innovate on time cycles that are relevant enough. And so what I heard from them, you know, example of, okay, US company finally gets approval to get something out to the battlefield in Ukraine. They use it, you know, it doesn't work. It's some reason the frequencies are getting jammed, whatever. There's obviously a lot of electronic work going on in the front lines. They get the feedback, take it back, 12 months before the company's got to iterate on it and then jump through a bunch of infernists to the industry, a bunch of hoops in government to do testing the value of the way she didn't get it approved to set it back overseas. And that's just not fast enough to be useful. And I talked to drone developers who were chatting in real time with the companies that were using their technology in the front lines and making changes right there in their warehouses that they're building the technology. And so like, that's, how do we translate that idea to our system for future conflicts, whether that's a conflict against China or other ones, how do we compress these time cycles of development and make it easier for us to adapt on the fly? And that's going to be really critical both when we're preparing for future conflicts, both of which were in a work, you know, no plans for our first contact. If you want to meet some stuff that's going to work, some of it's not, the enemy's going to be adaptive. We're going to have to adapt on the fly quickly. And we did a lot of this actually in a rocket of GANIS stand in ID defeat in, you know, ISR, task force, MRAP task force. But one of the things that strikes me is in all of these examples, we had to go outside the traditional system with the right standalone organizations, task forces, in many cases that reported directly to the Secretary of Defense in order to move it at time cycle of relevance. And so we're going to have to find ways to shake up our system in order to move faster to stay competitive with the way the technologies be being in future conflicts. And they're trying to do so that right. There's this bunch of new directions that the Pentagon has taken. And there's the memo that requires preference for consumer, a dual use software, which means that you have to find stuff that's over there. There's one that pushes a lot of authority, particularly for drones down to the Italian level to like, to the acquisition executive is actually a guy that does shooting or at least they have a lot more input. Yeah, there's the most recent one. What else can they be doing that they should be doing now in order to speed up those time cycles? And do you think we were talking about so I'll leave with that and then I'll get back to like the AI question. Yeah, I do think that this leadership team in a Pentagon to the credit is very invested in shaking things up and moving faster, and clearing on some of the red tape. I think that there's a big piece of this that's got to happen on Capitol Hill as well. And Congress has got to be supportive. Some of that has to do with just like things like getting less control out of the appropriators and giving the Department of Defense or whatever call it these days, you know, flexibility to spend money very fast and be flexible on how they how they move money around. I think that's really important. It also requires something of a shift in mentality of, you know, when we buy something, how long are we planning on buying this for? And if it depends a lot on what we're investing in, the Army several years ago bought a new handgun. Okay, if you're buying a new handgun in the Army, that's pretty stable technology. We can buy the handgun and keep it for a very long time. We can be somewhat deliberative in the process. But if you're talking about cheap, expendable drones, we need to be willing to acknowledge we might invest in this system or electronic warfare systems, other ones and five years might be obsolete. And that's okay. Now that might change then your strategy that it might be more that I'm going to buy things in small batches and field them to units to experiment with, take it out in the field and, you know, sort of bacon options to scale production quickly. If it turns out, hey, we go to war or we see that anticipating and we need to do it. It's sort of a different way of thinking about how long you're going to buy something just like, and if you buy a car, you might plan to keep it for a long period of time. If you buy a iPhone, you know, I keep the thing for 10 years realistically, right? That's going to be too outdated. And so I think that's that's going to have important implications for how the military actually spends. So I'm on the, I wanted to be back to the other thing that the US lady has was going to play a part in the US maybe creating something like parody and hopefully dominance in the event of like a large scale future conflict that obviously won both drones and innovating the drones in the front line. But that innovation cycle as the entire point of some of these AI models of all using it's supposed to be cut like, you know, you're new to loop at work. That's supposed to be helping you do that like scale up yourself and your talents so that you can do things faster. And when you think about not everybody is going to want to be on an island in the Pacific and then trying to connect to a cloud to use the military version of chat to duty. There might be room for some of that. But a distributed kind of personalized version of a large scale foundation. What we today reexperiences chat GPT scaled under that person scaled down to that that indian has to figure out this new drone problem is we wrote about probably quickly. That's a potential game changer is that something we have unique advantage on or is this something that's also going to be very rapidly democratized in terms of in terms of use because that kind of puts a lot of trust in the NCO. And that's not something that the Chinese or the Russian military historically done. Yeah, I mean, I think there's two potential I'd say it's potential sources of advantage for the United States, but we have to after harness them and use it in certain way. So one is actually just in the computing hardware itself in the chips. We do have a huge advantage here over China. You know, partly because of Nvidia, but ultimately it's because of actually TSMC and the fact that they rely on US software and US tool link equipment to make these chips that are all made into timeline. Which is like actually with a real choke point lies on the semiconductor manufacturing equipment side. And I think that's one really critical that we're making the right moves to sustain that advantage over the long term. And so while there's a lot of debates around which chips should we allow export to China, I think that's actually quite important of a conversation now. The most important thing is on the manufacturing equipment side and sustaining those export controls to China so that they are dependent on chips externally. And I think that's that's like absolutely super critical. And then there's a question of like, well, how much compute should we allow them to have? I think we want to probably the most sensible position would be to be exporting chips that are just a little bit better than what they can make domestically to help suppress the domestic market inside China. But there's a pretty healthy gap between what China has and what the US has so that it can seize on that advantage for the military standpoint. And when the military's got to be moving really fast in terms of finding ways to use this computing power because if they're like five or 10 years behind the leading edge of terms of computing power, then it probably doesn't really matter. But then the other side is you know, on the cultural piece, I think that is potentially a huge source of advantage for the United States. We're already in a place where we delegate a lot of this authority to junior officers and MCOs. I think what's particularly powerful here is it's not just baked into the culture of the US military relative to the PLA, but it's actually baked into the culture of like the American system. In a way that's much more entrepreneurial and much more just freedom for people to innovate and be creative that inside China. That's not the same that you can't have particularly a tax side, a lot of innovation. We've seen actually the last 25 years that's very possible in China, but there's also a very strong authoritarian streak in the country that we've seen under Xi, the CCP cracking down on. That's you know, in some ways from a geo-strategic advantage standpoint, I think probably actually the US advantage because I think in the long run that's going to have effects to hindered China. So how do we actually capitalize on that? Right, so we've got to be putting that technology in the hands of our war fighters, our junior officers or MCOs and allowing the to grab ahold of it and then use it effectively. So I got kind of two questions that pick up on points that the UN Patrick had flagged. And the first, Paul, you've been working at a DC for a while. Do you, and you have mentioned Congress a time or two in this discussion? Do you worry that US lawmakers are behind the curve in any particularly concerning ways when it comes to this, you know, this AI kind of tech race we've been talking about? Well, I think it's fair to say that we don't have the tech literacy that we need in Congress in order to stay on top of this technology. Now, in fairness, the technology we've been super quick, lots of places people don't have the tech literacy. I think there's probably too many companies where companies are trying to figure out what is this AI and how do I use it and, you know, how do we use it effectively. And I've been super impressed by the level of energy that we've seen out of Congress since really the chat GPT moment a few years ago, where we had lawmakers, you know, holding hearings and inviting tech leaders to come in and get smart on this quickly. And in a lot of my personal interactions, I mean, there's a lot of members and staffers that are actually, we're clearly deeply knowledgeable. It's hard, of course, is translating that to lawmaking and regulation or, you know, what is the role of Congress effectively? We see debates, for example, around preemption right now. And states like California moving forward on regulation and the absence of federal regulation, I don't think anyone believes, for example, that having 50 different regulatory regimes in the United States is a good thing. Nobody wants that. That would really kill innovation. On the other hand, that's not the alternative being proposed. Right now we see the federal government doing nothing in terms of regulating this technology. And I'm just not sure politically that's going to fly with the American people who were worried about their jobs, their worried about surveillance, tech, the world about child safety. There are a lot of concerns that people have that are actually pretty valid in pretty legitimate, that we've got to work through as a society. And Congress probably is a role to play in some of these issues in delivering funding for things that can be really critical for sustaining national advantage, whether that's in computing hardware or military spending or RID, but also for managing what if the technology delivers and everything that is proponent say it's going to do is going to be probably good amount of societal disruption. And we're going to have to find ways to navigate that as a society to make sure that A is actually beneficial for everyone. Do you think there's anything useful to be said about the race to build new data centers all across the country? I was just looking at a map today from a source called DC Byte. It showed Virginia had the densest cluster of these and there were lots of others in Georgia, Ohio, Texas, and other states in the southern US. So I'm kind of curious, you know, what if at all, what's been the conversation around this topic at your level? Well, you know, it's funny. I think that no matter what AI is doing, people are critical. So right now, there's so much spending on AI and the build out of data centers that the concern is, is there a bubble and this is overbilling and there's not going to be demand for this? Now, look, maybe that's true. You know, I'm not sure to be honest with you. I think some of the circular funding to deals as just like a layperson, like a little squirrely, but what do I know? I would observe certainly that with people comparing things to the dot com bubble and when that person dot com crash, like, yeah, that definitely happened, but the internet turned out to be real. And so I do think over the long run, whether there's a bubble or not, whether there's a crash or not, that making a long term bet on AI is probably pretty sound. This technology is quite likely to continue to mature. There's just sort of a more near-term question of, okay, will some of these AI companies that are scaling up very rapidly and taking on, you know, making a lot of spending like open AI, are they going to be able to generate enough revenue to satisfy investors? I'm not sure about that. You know, on the one hand, you could look at it and say, well, the math doesn't end up, and lots of people have kind of, you know, tried to do these calculations. And there are, and there's a lot of investors that people, boy, wish they had gotten in on the ground floor of Netflix and Amazon and Facebook. And, you know, I think that is driving a lot of this investment, the potential of the technology. I will say that it's interesting to hear this conversation right now about, oh, we're building too many data centers. When a year ago, what I was certainly hearing from all of the AI experts worried about maintaining use advantages, we're not building enough data centers. China is going to outbuild us. And so I would like for me to think about national advantage. It's hard to know exactly how to calibrate this. I do think it is valid that computing hardware is going to be the key thing that makes the difference between the US and China in terms of national advantage. And you can think about computing hardware and data centers is sort of the equivalent, like a comparison here, not identical, but a comparison to manufacturing capacity during the DelShrub Revolution. And just like you could, we saw actually the US and others retool, um, any factory plants during World War II to start cranking out, um, military types of production that similarly, you know, having data centers, having computing hardware, having chips will be really critical for national advantage, both in the, um, sense of advancing AI in the sense of it. It, applying AI across society for economic productivity and, um, growing industry. And then if necessary, in certain areas, applying that to military advantage, or the national security application, say like intelligence or cybersecurity. And so I'm actually pretty happy that we have this massive build out. I was worried a year ago that we wouldn't see this level of growth. Obviously if this all comes crashing down, that would be great, um, you know, from an economy standpoint or from an asset competitive standpoint. But I'm, it's, it's an interesting shift from even like three years ago or so, you would ask for it saying like, it's companies are talking about AGI and, and but the market isn't reacting as though AGI is coming around the corner and they should be. And now we see the markets reacting. And then people don't like that. Um, but I think, you know, AI is coming and I'm glad I actually to see companies investors responding to the potential of this technology. Yeah. Well, we've mentioned, uh, Ukraine and we've talked of course, um, about Russia, China. It's kind of a big picture question. Um, but I do wonder, how do you think AI can help democracies or do you think it's a bit of an unhelpful distraction to try to even link, uh, those two concepts at all? I'm not sure. I mean, I think you could imagine things that would be like troubling if you got real extreme concentrations of power in companies, um, that are, that are like to a larger scale than the kind of winner tick all dynamics that we already see in different aspects of digital tech. Um, you know, when operating systems and handsets and social media platforms, for example, AI, so if you ended up where there's just like one or two companies that are dominant, the dwarfs in the scale far beyond what we're seeing today, that probably would be worrying from a concentration of power standpoint. Um, a lot of the things that I've seen or worried about or see potential advantages of our much more second order, um, I think certainly the types of surveillance tech that we're seeing China, Pioneer and X-porter on the world are troubling for democracy. Um, but I think we've seen in the United States that we haven't just blindly used AI, um, you know, throughout the country so far to like created news or fail in state, which is encouraging. Like there's actually, you know, around things like public cameras and AI surveillance in that context, pretty pretty healthy and robust, um, you know, debate inside the United States and sometimes the grassroots movement against, for example, um, you know, public surveillance tech using AI coming out of, you know, cities and states. I think that give and take is really important to get responses that are good from a democratic standpoint and having a dialogue between individuals and companies and the government at the federal level and state level and local level and grassroots movements and civil society and media like that's messy, but that's going to get a better outcome in the long run. Obviously, there are some wealth cards out there like labor market disruption that could change the game significantly and maybe it'll turn out that AI will be more like other types of technologies in the past, um, like the shift from, uh, farms to factories and okay, sure, like all the labor that was previously working in agriculture got displaced by the industrial revolution, but there are new jobs were created and we'll be fine, they'll be true for AI or maybe it'll be different. The industrial revolution did poor horses have a business and they've grown up like horses and they're just really good reasons for humans to perform labor in the economy and that's a thing where like, okay, we'd have to navigate that as a society. Um, and so, um, I think that's a great thing. And so I'm not sure, but I think it's, what's gonna be important is having ways society to work through these challenges to ensure that we're, the benefits of this technology, however the materialized, are gonna be broadly beneficial and not just benefit if you. - Yeah. All right, my very final question. And again, thank you for taking so much time to talk to us, by the way. - Of course. - On a departing and much lighter note, are there any films or TV shows and recent years that for you have addressed issues related to AI in compelling or perhaps even hilariously misleading ways? I'm curious. - Oh, so my little hobby horse of this is that I think that the way that we often talk about AI is too anthropocentric. It's too much centered on the idea of a human-like type of AI system. Even conversations around AGI is like, "Why do we think that AI is moving in the direction of something that looks like human intelligence?" My suspicion is that what we're likely to find. I think we're seeing elements of this already as AI improves is that intelligence is not like a staircase where you move from ants to dogs to chip anties to humans and AI moves up the staircase. But much more like this vast space of possible intelligences of the humans are one that was specialized for a certain type of evolutionary niche that we fit into and in AI is probing the space and that even large language models that generate text that is human-like because we're trying it that way, under the hood the way they're functioning is quite different and quite alien. And so I sort of don't like things that have AI systems or robots that sort of feed into this anthropocentrism. And I love things that make AI seem kind of alien and weird. One of my favorite, I'm gonna spoil something here. So I'm like, "Well, well, well, well, well." One of my favorite is the movie X Machina, where like the whole premise is, you see this robot and you know it's a robot and can't it pass the touring test and be convincing someone that it's like a person which is kind of intriguing. But there's this great moment to end where I think it was Ava, it was the name, where the robot character basically like turns on a dime and turns against the humans. And I think what I like about that is not only is it very well presented, it's very chilling, but it really drives home the idea that our mental models for intelligence and how we deal with other human beings are not only not necessarily gonna work out like well for AI, but might really set us up for failure in a way that's quite catastrophic because we have all of this internal machinery that allows you and I to have this conversation, us to meet strangers on the street, interact with people in all of the social settings that we have in our lives. And it works out super great, but those mental models that we use, if we apply them to AI and we sort of anthropomorphize the machine might be really quite dangerous. And I think that that's, we've seen, I mean, not that we've seen large language models behave in a silly and that way, but we've seen glimmers of that with large language models where they'll do strange, bizarre things that I think allow us to see that beneath this thin, veneer of human like presentation, what's going on under the surface is quite alien and weird. And I think that's an idea that we want to hold on to as its technology continues to evolve. - Well, Paul Sharre is the executive vice president at Center for a New American Security in Washington, Paul, thank you so much for speaking with us. - Thank you, thanks for having me back on the show. Very exciting. - Well, that's it for us this episode. Thanks for listening. And until next time. - America's defense strength comes from the people who build it, welders, machinists, and others who produce the equipment are military relies on. Learn more at gearospacet.com/workforce.