This transcript is from a CSIS event hosted on May 1, 2025. Watch the full video here.
Navin Girishankar: Welcome, everybody. Welcome. I’m Navin Girishankar. I’m the president of the Economic Security and Technology Department here at CSIS. Really delighted to invite all of you for this important leadership conversation on U.S. Leadership in AI. We’re focused on economic security. And you can’t really have a serious conversation on economic security if you’re not talking about U.S. leadership on AI. And you can’t have a serious conversation on U.S. leadership in AI if you’re not talking about our ecosystem, our relationship with allies, what we’re doing at home across the full AI stack from frontier models all the way to the energy needs.
And so we are really delighted and honored to have Scale AI founder and CEO Alexandr Wang with us today. We are delighted to have him here for this leadership conversation. And he’s going to be taking – the person who will take us through that conversation with him is Greg Allen, who’s the head and director of our Wadhwani AI Center. And there’s no better person in our building to have this conversation with Alexandr.
Alexandr needs no introduction. I remind you – excuse me, because I’m on the other side of 50, so I’ll say this. Which is, that I’m reminded of the famous Lincoln saying that it’s not the years in your life but the life in your years. And really, Alexandr is, like, a phenomenal leader for our country across what Scale AI does with government, private sector, multinationals. And so really will give us a unique perspective with this incredible footprint. And he’s been a leader really in sharpening the focus and making sure we’re aware of the core issues of national security and technology leadership when it comes to AI.
So I’m really looking forward to this conversation. So with that, I’m going to invite Greg and Alexandr up here on stage. And, really, thank you and welcome to CSIS. (Applause.)
Gregory C. Allen: Wow, Alexandr, thank you so much for coming to CSIS.
Alexandr Wang: Yeah, of course.
Mr. Allen: So we already heard a little bit about this from Navin. I think you’re ubiquitously known in the AI industry and you’re a mere 99 percent known in the AI policy world. But for those who’ve been sleeping under a rock, I think there’s just some things about your story that are pretty remarkable and worth calling out. So, as we all know, you’re the founder and CEO of Scale AI, which is described as a company accelerating AI development by delivering expert-level data and technology solutions to leading AI labs, multinational enterprises, and governments.
And this is a company that you founded in 2016 as a 19-year-old MIT student with the vision of providing critical data and infrastructure needed for complex AI projects. So now Scale AI from that sort of humble origin you’ve got more than 900 employees, you’ve got hundreds of customers across a very diverse set of industries, and you’re really one of the linchpins of the global AI ecosystem but many times by enabling other people who have really, really strong brands that we all know today.
So that is a pretty great story. So you started in 2019 and then only six years later by the time you’re 2025 – sorry, by the time you’re 25 you made your debut on the Forbes billionaire list and Scale AI is valued at more than $14 billion.
So that’s a pretty good six years, you know, from ’19 to ’25. I had some hits myself but not on that scale. And so I just want you to kind of tell that story. How did you get Scale AI to where it is now? How do you start from startup to one of the linchpins of the AI ecosystem?
Mr. Wang: Yeah. So it starts, you know, kind of where I grew up. I grew up in Los Alamos, New Mexico. My parents were both physicists who worked at the national lab and it was – you know, it’s one of the few places –
Mr. Allen: Los Alamos of nuclear fame and Oppenheimer fame, right?
Mr. Wang: Exactly, yeah, of the Manhattan Project, of Oppenheimer. You know, in many meetings this week in D.C. people refer to AI as a Manhattan Project-level effort and I can always say, oh, yeah, I actually grew up there.
Mr. Allen: Yeah. (Laughter.)
Mr. Wang: But that was – I mean, I think I was – I really grew up with a lot of education and, I would say, purpose around national security and technology and sort of the intersection of the two because of that.
And then when I was – so then I went to MIT after growing up in Los Alamos and I studied artificial intelligence, and this was right around the time frame when DeepMind came out with some of its initial breakthroughs around AlphaGo – you know, number-one Go player.
And sort of it felt very visceral to me that AI, you know, if it continued this, like, incredible pace, which even at that point – you know, deep learning started in, roughly, 2010 and this was around 2015.
In those five years deep learning went from something extremely rudimentary, could not – you know, was not really capable of anything all the way to, you know, number-one Go player, great image recognition capabilities.
Mr. Allen: So at the time you’re sort of starting your professional journey the turning point is already obvious to you, right, related to AI, and that’s, like, the starting point is, like, OK, we are at this critical moment for AI. That’s the wave I want to ride.
Mr. Wang: Yeah. It felt very clear that AI was going to be critically important, and I remember I would sometimes tell people at the time, hey, I think AI is going to be the most important thing in our lifetime and then, you know, they kind of laugh. And it was a hard thing to say with a straight face at the time but now I think you can say it a bit more – with a bit more of a straight face.
Mr. Allen: Yeah.
Mr. Wang: But so, yeah, AI is clearly going to be very important and if you decompose the problem of AI, you know, you can boil it down to: algorithms, so the sort of exact neural network architectures and whatnot that are being used; computational power, which boils down to chips and power and datacenters and all that stuff; and, lastly, data, which is really the raw material for intelligence and sort of the lifeblood of these models and where they learn all their stuff from.
And as you sort of looked around at the ecosystem there were, you know, companies like OpenAI and DeepMind focused on algorithms. There were companies like Nvidia and others focused on the computational problem and there was a dearth of people focusing on the data problem.
And so that was the sort of origin of Scale back in 2016 was, really, it’s very clear that if I believe AI is going to be one of the most important technologies in my lifetime and there’s this sort of gaping hole of necessary infrastructure that’s necessary to support this industry, you know, someone’s got to do it and it may as well – I may as well try.
And so that was the original – the sort of genesis moment and then we started by working in the autonomous vehicle industry. You know, we worked with folks like Waymo, Toyota, General Motors, and many others, and this was – you know, at this point, back in 2016-2017, the biggest dream we could dream about with AI was self-driving cars, which was very exciting but –
Mr. Allen: Yeah, I remember. Like, it’s, like, 2016. That’s when the then CEO of Ford said that, like, in 2021 you will get in the car, go to sleep, and arrive at your destination completely unharmed and safely. So this was, like, a very hot moment for autonomous cars and you’re a part of that wave – that initial excitement.
Mr. Wang: Yes, exactly. And, ultimately, we’ve been successful, right? Like, Waymo, one of our partners –
Mr. Allen: Now they’ve achieved that dream, a little bit less than the Ford CEO was thinking it would happen, but now it’s real. You really can do that.
Mr. Wang: Yeah, exactly. I mean, and it’s incredible. But obviously our ambitions with AI continue to grow exponentially over time.
Mr. Allen: And just for folks who are thinking about, like, the AI value chain, the AI stack, I think folks mostly know what an OpenAI does. They know what a Waymo does. But – so, like, when you sign a contract with Scale AI, what are the services that you’re getting? Which of those are software as a service? Which of those are, like, human touch, kind of white-glove service? You know, what is Scale AI providing? And I suspect it’s a pretty diverse portfolio of stuff right now, but to the extent you can sort of summarize.
Mr. Wang: Yeah. So we really think about it as twofold. One is data infrastructure and the other are technological solutions. And so, on the data-infrastructure side, really it’s about – you know, we call it our data foundry. But just like you have a chip foundry that spits out chips, you need a data foundry that spits out very high-quality expert data that can fuel these models’ capabilities. And this is scale – I mean, even in the timeframe that we’ve been working on the problem, I mean, the complexity and richness of frontier data has just grown astronomically. And it’s been – you know, it has to scale with the capabilities of the models.
Mr. Allen: Yeah. So when you’re starting out and focusing on autonomous vehicles, that might be, like, I need 50,000 images of stoplights, and I need them from all angles and I need them from different lightings and weather conditions and all these kind of things. So it’s like when you’re curating the data set, you’re actually programming the AI system. And what you’re basically saying to a company like a Waymo or like a Cruise is we are A+-level, you know, data-set curators and also data-infrastructure managers. Let us handle that for you and we’ll help you make a great autonomous car. Is that kind of the pitch at the time?
Mr. Wang: Yes, yes. And it evolved towards much more data generation. And, you know, the interesting problems in the self-driving world, which actually have very strong parallels with current problems in defense and national security –
Mr. Allen: Sure.
Mr. Wang: – the core problem is, you know, the problem of what’s called sensor fusion. You know, you have a lidar sensor. You’ve got a radar sensor. You’ve got a GPS. You’ve got cameras. How do you actually fuse all the disparate signals?
Mr. Allen: So that the whole is more than the sum of the parts.
Mr. Wang: Exactly.
Mr. Allen: Yeah.
Mr. Wang: Yeah, yeah. And that was, I think, one of the – that problem was one of the primary hurdles that had to be crossed before we have, obviously, safe autonomous vehicles.
Mr. Allen: And you crossed it before other people did, which put you in an incredibly strong position to grow to all these other now portfolios. So can you talk a little bit about the jump between – and I’m sure there’s steps in between – but between the autonomous-vehicle era of Scale AI, which I realize you still do all this work, but now this generative-AI era, where you really reoriented so much of what the company is doing around the generative-AI revolution.
Mr. Wang: Yeah. So starting in about – so if you look at the arcs of AI, so this sort of like autonomous vehicle or, broadly speaking, computer-vision-driven sort of trajectory, you know, it continues even to this day. But then starting about 2017, 2018, 2019, we saw interesting things happen with language models. And 2017 was a transformative paper; 2018 was the first GPT; 2019 was GPT-2, and so on and so forth. And we were partnered with OpenAI in the very early days of this. We worked with them on GPT-2 back in the day.
And it was this incredibly – you know, it’s funny to reflect back on it, because from a research perspective, this trajectory of transformative GPT to GPT-2 to GPT-3, it was very boring from a research perspective, because all it amounted to was you just threw more data and more compute at the exact same algorithm, and then it just got way smarter. And that’s a very unsatisfying answer for a researcher, because, you know, you want to believe that there’s some tricks or there’s some –
Mr. Allen: There’s some algorithm to crack, some, like, secret sauce. But really it’s just the same recipe, make the pizza twice as big and bake at the same temperature, and bam.
Mr. Wang: Exactly. Exactly. So more data, more compute. And so we saw these incredible scaling laws, which obviously has defined so much of the modern AI landscape is the existence of these scaling laws. And we worked closely with OpenAI and then later the entire breadth of the AI industry, you know, including Google and Meta and many others, on this incredible trajectory of large language models and now more agentic models and reinforcement learning.
It’s been – I mean, it really has been one of the most incredible things to witness. And I think, you know, if you had asked me in 2016, or even 2015, hey, in one decade where do you think AI capabilities would be, it was very difficult to predict that they would come this far. You know, even at the time we would often – you know, myself and other researchers – would talk about, oh, you know, when do we think that AI models will be better than the best competitive coders and the best competitive mathematicians? And, you know, I think the timelines were, like, thirty years, forty years, fifty years. But it turns out it was less than a decade.
Mr. Allen: Yeah. And you were a competitive mathematician.
Mr. Wang: (Laughs.) Yes.
Mr. Allen: So I realize, you know, this maybe is, like, a sore subject, but how do you personally feel as you see AI thriving, succeeding at the stuff that you spent a lot of your high school, you know, competing at? Is it like the pride of a father looking at your little robot baby? Or is it like – you know, you’re, like, I could still take them. How are you feeling?
Mr. Wang: Definitely can’t still take them. (Laughter.) I think it really causes you to perceive the rate of technology development in a very – with a very sober lens, which is that these models are becoming extremely capable over time. And that trajectory of capability, you know, just – you have to pay attention to it. And I think many people from many – you know, many, many different fields or areas of study I think of have experienced something similar, which is just the pace of progress is absolutely astonishing.
Mr. Allen: Yeah.
Mr. Wang: Yeah.
Mr. Allen: So Scale AI – thank you for taking us through the Scale AI story. And there’s a connection here between your personal history. And I remember, of course, in, I think it was 2018 timeframe, when Google pulled out of Project Maven. And I was, you know, a think tank wonk even then. And I was writing, hey, this is a mistake and AI researchers should be helping with military work. And it was a pretty unpopular view at the time. But you and I have something in common, which is we took the unpopular view. And you went even further, which is you actually moved your company to it. And I think for folks who don’t understand how remarkable this is, in the timeframe that we’re talking about – which is the – you know, the early 2020s, the late 20-teens, there were sort of two views on working with the DOD.
Number one is, this is a terrible idea because the DOD is a terrible customer and you’re going to lose a lot of money. And the other point was, this is a terrible idea because the entire, you know, AI tech ecosystem hates the military and all your best people are going to quit. So that was the conventional wisdom at the time that you, you know, running this, not at the time juggernaut startup. At the time, it was a little bit more of a humble startup. And you said, no, I’m actually – this actually matters to me. I’m going to put this company to work for the U.S. national security community. So talk me a little bit through that decision and how you think it’s gone.
Mr. Wang: Yeah, so before – so, before I – so maybe I’ll take a step back. One core belief I have is that over time rationality prevails. And over any time –
Mr. Allen: (Laughs.) That’s a good one.
Mr. Wang: Yeah. Over any time horizon, eventually the rational viewpoint will prevail. And this was, I think, a case of that, where, when I thought about this problem it was just so obvious that we’re at the brink of this incredibly powerful new technology. The applications towards national security are very obvious. And it was – it was going to be imperative for the United States to remain ahead. And, you know, this came off the heels of – I went on a – I went on a trip to China. One of our investors organized this trip. We visited the Chinese AI companies at the time. A hundred percent of them were focused on facial recognition and surveillance. And it was very clear – and they were very good.
And so it was very clear that there was – there was going to be this moment where the intersection between AI and national security was obvious to everybody in the world. And at that moment, it was clear that the United States would need to have the highest quality human capital as well as the best companies focused on this problem. And certainly there was – there was skepticism from our investors, there was skepticism from our employees, and there were employees who left the company as a result of these decisions. But from my standpoint, it was sort of, you know, knowing – well, my parents work on what – you know, the history of the place where I grew up, it felt just – and the impact that Los Alamos and the Manhattan Project has had on Pax Americana and the global order – like, it felt so clear that that great AI technology was going to need to be applied to the national security problem set. And that we had, frankly, a moral imperative, I think, to support it.
Mr. Allen: That’s great. And so I’d like you to talk a little bit about the work that you’re actually doing for the national security community as a defense contractor. So you’ve got contracts with the Office of the Chief Digital and AI Officer, which is – where I’m an alumni of that organization. You’ve also got other kinds of things going on across the DOD, across the broader national security community. What kind of stuff are involved in – I mean, I realize some of this is classified so you can’t talk about everything, but what would you sort of say is the impact of Scale AI on the national security ecosystem so far? Because you’re involved in some pretty big projects, some pretty big contracts.
Mr. Wang: Yeah. So one of the pieces of work that recently was announced that we can talk about is our work with DIU and INDOPACOM around Thunderforge. And the effort is really to utilize AI and AI agents in the military planning and operational planning process. And some of this actually came out of some work that we did with CSIS –
Mr. Allen: Yes! And I should give a shout-out to my colleague Ben Jensen, who has been a part of that work, and it’s been a great partnership.
Mr. Wang: Yes, exactly. We’re – I think our teams together originally saw the potential for AI within wargaming efforts, and I think that – you know, we’ve seen the ambition of that idea just expand towards operational planning, military planning, wargaming simulation, and encompassing this whole loop. And it really is part of what we believe is a critical element of the future of warfare, which is towards what we call agentic warfare, which is, you know, there are a number of processes today where they’re extremely manually intensive – and that’s not necessarily the problem. The problem is speed. You know, we need to be able to act as quickly as possible if we imagine what future conflicts will look like and what the future of warfare might entail.
And so we’ve built this system in partnership with DIU and many great members of the military to leverage agentic systems towards sort of the massive acceleration of this process. And it’s one of the things we’re very excited about.
We also have done a lot of work with CDAO. You know, one of the things that we’re working on with them most recently actually was a test and evaluation framework for AI systems. And I think this is – this gets to the heart of I think some of the critical components of AI deployments in a largescale way within national security. I think there are a lot of people who have fears – very reasonable fears, I think, or concerns that, you know, we should not put unfit or immature AI systems in positions of great importance or power.
Mr. Allen: Yeah.
Mr. Wang: And it’s I think a very fair –
Mr. Allen: Important things should be reliable things, as a general rule, right? (Laughs.)
Mr. Wang: Yes, exactly. And I think for whatever reason the AI debate partially because it’s so – it is so new and it’s so central to I think the discourse, I think sometimes the dialogue will go in, like, two opposite directions – one group of people saying, oh, we just – we should never deploy AI systems, they can’t be trusted; and another group saying, we need AI systems everywhere without any abandon, and we just – you know, they have to – we have to accelerate everywhere across the moon. And I think the – our viewpoint is somewhere in the middle, which is, we should move very quickly to deploy AI systems throughout the DOD, throughout the national security community, but we should do so in a thoughtful way with the right test and evaluation to ensure that these systems are robust.
Mr. Allen: Totally, totally agree. Very easy to agree with that statement. As you said, one hopes that rationality will prevail, right, and that that’s obviously the right answer.
You mentioned Thunderforge, which is a very interesting program, and so I want to kind of ask you, what do you feel like you learned at Scale about war planning – either the gaps in conventional war planning or the opportunities from applying LOMs with that Thunderforge program? And I guess a follow-up would sort of be, well, we have LOMs but these things are available openly on the internet, there’s opensource versions of them, so how do you think a capability like this relates to sort of strategic advantage in the long haul, given that a lot of these capabilities are increasingly commoditized, increasingly democratized, et cetera?
Mr. Wang: Yeah. This is – these are some important questions. Yes. So I think first off, our view – you know, one of the incredible things about large language models and the ability to utilize them in military planning, operational planning, is just the degree of knowledge that they have, the knowledge and experience they have. You know, one of the things that we often talk about is, you know, the difference between – you know, right now we have incredibly talented humans with decades of experience doing the military planning process, and AI agents are – have literally thousands of years of context built into them, and knowledge, that can act, you know, near instantaneously. And so, you know, it would be a fool’s errand not to leverage them.
And so I think a huge amount of the value from our perspective is, like, how do you leverage the incredible amount of both their speed, of course, but also just the amount of context and prior knowledge and capability that they have to integrate into our military planning process.
You know, the second question that you asked is a very hot topic now, of course, because open-source models, especially those coming out of China, have moved incredibly quickly. And, you know, DeepSeek and more recently Qwen from Alibaba are world-class models. You know, it’s impossible to describe it as anything but that. And it does raise these questions. Is there strategic advantage in, you know, the models that we have in the U.S. versus the models that they have? And, you know, is the advantage in the implementation or is the advantage somehow in the feedback loop, or where does all that exist?
You know, my belief is that – or not even my belief – the CCP has had an AI master plan, and they’ve talked about this – I think you could date back to 2017, 2018, they began talking about these sort of grand plans to win on AI on a longer time scale. And some of the things that they’ve acknowledged as part of that plan are something we haven’t even talked about; you know, the need to win on data, the need to win on compute, the need to win on these sort of implementation and proliferation of this technology throughout their military.
You know, they explicitly say, you know – this goes back to 2018, so it was actually quite prescient from them – that they view AI and other advanced technologies to be leapfrog technologies –
Mr. Allen: Yes.
Mr. Wang: – where, if they overinvest in AI compared to the West, then they have an ability to leapfrog our defense capabilities. So their strategic picture, I think, is very clear. And I think they’ve had a similar picture since 2017, 2018. And DeepSeek and Qwen and these models are sort of on the smooth continuum of that strategic view.
I think what’s necessary for us within the – you know, for the DOD is, like, how do we accelerate our deployment so that we can begin leveraging these feedback loops within our own deployments of AI for national security to accelerate our advantage? And one of the things that – you know, I testified on the Hill a few weeks ago for the House Energy and Commerce Committee. One of the things I talked about was this concept of data dominance, right?
If you think about the raw material for intelligence for all these models is their data, then the obvious conclusion is the way that you will win is through data dominance, more and better and more relevant data than your adversaries. And the key elements of that ultimately, as we’ve seen in autonomous vehicles or even in chatbots more recently or even historically with algorithmic feeds in the large companies, is that that boils down to can you implement the feedback loops that will actually enable these technologies to improve at an exponential rate.
Mr. Allen: Yeah. So you talked about adoption and the speed of adoption, the speed of iteration, being one of the sort of critical sources of advantage in AI. And I think that certainly applies to the DOD context, where I think scale, it’s fair to say, is like trying to provide the kinds of infrastructure and services that make it easier for every single program of record to try and incorporate AI in meaningful ways, that sort of thing. That part of the story, I think, makes sense to me.
I’m curious, you know, one of the things that we – that I thought a lot about when I was in the DOD was all the ways that the DOD made my life hard, right; where I was thinking if I wanted to build this AI and I would encounter, you know, not just the ways that AI is hard because AI is a tough technology in some ways, but all the ways that it is uniquely hard in the DOD context.
And I think scale is very interesting in this regard, because you have at this point a pretty sizable defense-portfolio business. But you also have a very, very sizable commercial-portfolio business. And so if you’re thinking – if you were, like, advising this administration, if you were advising the DOD, if you were advising Congress on, like, what are the ways that DOD can make it easier for those who are looking to meaningfully and significantly adopt AI for national-security advantage, what can we do to make their life easier? What can we do to make it closer to that, really, speed of adoption that you’re able to achieve in the commercial sector?
Mr. Wang: Yeah. You know, there’s many answers to this question, obviously, but I’ll try my hand at a few key components.
I think one of the first ones is just, you know, access to infrastructure, whether it be cloud infrastructure, data infrastructure, algorithmic infrastructure. You know, if you compare the platforms that a new startup out of Silicon Valley that they can build off of and the sort of the shoulders of giants that they can rest upon, and you compare that versus a new program within the DOD that wants to get started doing some AI stuff, it is night and day. And so an incredible – a necessary prerequisite to enable the same level of innovation within the DOD and the national security community as in Silicon Valley or the startup community or the commercial sector, is we need – you know, the DOD needs to take focus towards enabling enterprise-wide infrastructure to support all these new efforts.
Because, you know, a new startup can just integrate with, you know, however many APIs that they want, and then they’re off to the races. And in the DOD, you know, it’s each independent component that you want to integrate in is this whole complex conversation and battle. And, you know, defense procurement doesn’t make it any easier. And it’s extremely complicated. And so one of the things that we believe is necessary is our enterprise-wide infrastructure components. You know, we need enterprise-wide cloud infrastructure. Obviously, that’s been a priority and we’ve made good progress there. You need enterprise-wide data infrastructure. You need enterprise-wide model infrastructure. You need – you need many components of, sort of, this stack more broadly.
And then I think what’s necessary is, well, certainly, I think – well, the DOD has made a lot of progress in creating new acquisition pathways that enable the department to move more quickly. These are not – many of them are not fully widespread practices yet. And so I think there’s sort of – you know, that might be more, frankly, even an adoption question around many of the newer acquisition methods that have been – that have been invented. And then I think there’s a – there’s a fundamental question or requirement around ambition.
And, you know, one of the things that – you know, one of the things that we’ve seen as it comes to – when it comes to the DOD and AI is, fundamentally, are we thinking big about the ways in which AI technology could be absolutely transformational to our efforts? You know, what are the ways – you know, in our – in our parlance, like, how can – what does agentic warfare enable in terms of lethality, or deterrence, or real capabilities that would – that would be, you know, our own leapfrogs on our own technology stack?
Mr. Allen: Yeah. Are you making an analogy here between, like, sustaining innovation and disruptive innovation? Or are you taking this a different way?
Mr. Wang: Yeah, no, I think that’s – I think that’s a good example. I mean, I think that the – you know, one of the reasons why Silicon Valley and San Francisco is such a disruptive ecosystem is that the level of ambition is entirely unbridled. And, you know, the concepts that people aim to build with AI are very big, are very expansive, sound totally ridiculous at the moment. You know, take OpenAI. You know, the concept that we’re going to build AGI sounded ridiculous when it got started. And today, doesn’t sound so crazy.
Mr. Allen: Yeah. Pretty mainstream at this point.
Mr. Wang: Yeah. And the pace of technology – you know, particularly the underlying pace of AI improvement is so dramatic that you will ultimately, you know, in some small number of years, be able to achieve your ambitious – your ambitious ideas. And so one of the key things is we need the department and members of – people within the department who are empowered to think through what are those hyper-ambitious AI concepts that we believe would have meaningful deterrent effect versus our near-peer adversaries? And we need to just start building them.
Mr. Allen: Yeah. So I don’t want to give short shrift to all the other government agencies out there, because I know you do work with the intelligence community, you do work with the federal civilian agencies. You know, I’d asked what can the DOD do to make their life less hard when it comes to adopting AI. Is there anything that stands out to you that you want to say about the non-national security government agencies, should be thinking about when they’re thinking about adopting AI?
Mr. Wang: Yeah. Just the same – in the same thread as agentic warfare, is really agentic government. Like, how can – you know, really the question, I think, ultimately boils down to, how can you leverage AI agents, how can you leverage, you know, these new powerful models towards totally redesigning government services, government processes towards, you know, the most efficient and most effective versions of themselves? And I think there’s –similarly, there’s real opportunities to think quite ambitiously about, you know, what is – what is the optimal government of the future look like? And how can not only each agency, but inter-agencies, like, what is – what are the best models of collaboration? How can AI agents be leveraged to turn many of our government processes from human is the loop to human is on the loop processes?
Mr. Allen: Yeah. I mean, there’s – in what you’re saying an analogy that comes to mind for me is a lot of, like, digital government is we went from paper to PDFs and not, like, we went from paper to, like, a software workflow, whereby this stuff all takes place in, like, you know, databases and online, et cetera. And I think there’s a(n) equivalent version of that with AI, right, which is you don’t want just the AI transition to be we do the exact same thing we always do but now there’s AI there somewhere; you actually want to reimagine what can be done – how could we make this 10 times more efficient, a hundred times more efficient.
Mr. Wang: Right. Exactly. And I think that the opportunity for those levels of improvement, like, making government processes truly a hundred times better that’s on the table with the technology, and so that’s really what I think we should aim for.
Mr. Allen: So as we leave the DOD conversation we’re going to stay in national security and I want to think about China because you on January 21st sent a letter to the newly elected President Donald Trump on the importance of winning the AI war with China.
So what do you really see as the stakes of technology competition with China and where do you think we are in the story after DeepSeek, after everything else that’s been going on?
Mr. Wang: Yeah. As I mentioned, I mean, I think that the CCP has laid the stakes out more clearly than anybody else, which is they believe that it’s critical to military dominance and in the same thread also critical to economic dominance and I think, you know, at a very simplistic level that is what we’re playing for.
We’re playing for, you know, economic strength and military strength, and because of the compounding nature of the technology, the degree to which, like, you know, when you have AI systems that are improving quickly they will improve at just this dramatic rate there is some latent network effect here, which could mean that there might be some winner-take-all sort of dynamic.
And so those are the stakes. I think it’s quite clear and I think, you know, now – you know, when I was saying this stuff back in 2022, 2023, it was viewed as quite hawkish but now I think these are somewhat mainstream.
Mr. Allen: Yeah. It’s really been shocking how fast that turn is. As you said, rationality winning out. There’s also more data to support the point, I think. It’s just becoming more obvious.
So, you know, you recently testified before the House Committee on Energy and Commerce. So what do you think about the intersection of the AI issue and all of its energy stakes? And, you know, when I think about China I think about a country that if they want a coal power plant to go from empty field to coal power plant in 12 months they can do that. So what do you think about the AI and energy issue and its relationship to national competition?
Mr. Wang: You know, if you were to draw out – if you looked at the graphs of what does Chinese power production look like versus U.S. power production the Chinese graph is straight up and to the right, enlarges –
Mr. Allen: Just, like, gigawatts of energy generation capacity, right? Yeah.
Mr. Wang: Yeah. Total of – total number of gigawatts of energy in the country. It’s straight up and to the right, in large part because they don’t shut down their coal power plants, and the U.S. curve looks, roughly, flat.
And there’s a lot of great innovation happening there, actually. Like, a lot of it is replacement of fossil fuel generation towards green generation. But the aggregate number is relatively flat, and if you think about everything we’ve been talking about, what does AI capabilities boil down to?
Well, it boils down to algorithms. It boils down to computational capacity and capability, and boils down to data. This computational capacity fundamentally depends on incredible amounts of energy being produced.
And so, you know, we have the National Energy Dominance Council. We have many efforts aimed towards ensuring that the United States reaccelerates its energy production and we’re able to sort of meet the challenge of the AI race.
But I think it is a fundamental prerequisite of the United States winning.
Mr. Allen: Yeah. And another, you know, area where in the past you’ve taken a pretty strong position is on, you know, the policy toolkit around export controls or investment restrictions or all these other kind of things. So what are your thoughts about U.S. AI and semiconductor technology, export controls, vis-à-vis China?
Mr. Wang: Yeah. No, this is a complicated topic and a very topical one right now and, you know, it’s – the DeepSeek and Qwen examples are quite instructive which is, you know, these are models where the Chinese labs had many fewer computational resources than we did but they are certainly on par with frontier U.S. models. And so it’s clear that the answer to the question of who will win in AI is not – it’s not simply just who has more computational capacity.
Mr. Allen: It’s not a one-variable equation.
Mr. Wang: Exactly. It’s not a one-variable. It’s a three-variable equation. It’s, like, algorithmic advances play a large part. You know, they had a lot of algorithmic advances. And data plays a large part. And they have a more orchestrated economy around data. And so we need to think about – I think one of the things we need to think about, as a country, is, like, what is the first – I think in the past, I think, we’ve potentially simplified the problem too much towards just saying, hey, if we win on compute, we win globally. And I don’t think it’s quite that simple. I think it’s a complicated situation.
And then also we need to ask the question, what does our leverage look like when it comes to – when it comes to compute? And, you know, there’s a recent analysis done by SemiAnalysis around the new Huawei chips. And the gist is, at a system level, you know, it’s something along the lines of 2X the computational capability of the sort of leading Nvidia system, but requires 5X the amount of power. And so – which is almost a perfect trade for them. You know, they’ll trade power efficiency for computational capability any day.
Mr. Allen: Because if you – if you can build power at an incredibly fast rate, then all that just comes down to is dollars or, you know, RMB. And if they’re willing to spend anything, then you can make that trade.
Mr. Wang: And they clearly are. I mean, there’s large – you know, on the heels of Stargate, the Bank of China announced large scale programs to support computational infrastructure in the country. You know, the PLA has over – you know, last year, had over 81 contracts with AI companies, large language model companies in China. So that the degree – the capacity which they are at, fundamentally, an all-systems-go, you know, race dynamic moment is – you know, that is just the reality.
Mr. Allen: Yeah. So, switching gears a bit, before, you know, when we were talking about defense AI systems in mission-critical use cases, you know, you mentioned the importance of test and evaluation. And Scale AI definitely has some capabilities that are relevant in that. And I think what’s interesting is that sort of same family of capabilities, you’re also providing a version of that to commercial companies who are looking for – maybe they wouldn’t use necessarily the phrase, test and evaluation, which is very government, you know, idiom. But when they think about safety testing, when they think about reliability testing, Scale AI has stuff that’s of interest to them.
So I’m interested in your capabilities here. And I’m especially interested because in February, you know, the U.S. AI Safety Institute announced the selection of Scale AI as its first third party evaluator accessed to assess AI models. So what exactly is under the hood of this partnership with the AI Safety Institute? And how’s it going so far?
Mr. Wang: Yeah. So I think, first off, to some of the stuff I mentioned in the past, you know, I think – again, I think we believe fundamentally that AI technology is very powerful. It should be deployed very broadly. We should have the ability to understand the performance characteristics of the AI models that we deploy to just ensure that, you know, we’re doing it intelligently, and we’re doing it – doing it well. So, that’s one of the reasons. I mean, it’s certainly true for the DOD problem set or the national security problem set, because, you know, we’re dealing with very powerful capabilities that that must be managed properly. But it’s also important for, you know, every American enterprise, particularly ones in critical industries like, you know, financial services, or health care, or telco providers, or pharmaceutical companies, or – you know, the list goes on and on.
So, that’s really – you know, I think our view is that there needs to be extremely high-end capabilities around testing, monitoring, understanding the performance of AI systems. And that is a – effectively a prerequisite to what we’ve been talking about, which is a largescale adoption wave and implementation wave of AI across the economy and across the DOD.
Mr. Allen: Can you give us a sense of what the state of the art in technology is here? I mean, I think folks who are familiar with this field will remember folks like, you know, Blake Lemoine, who was a Google engineer testing the capability, and decided that it was sentient because he’d had a bunch of conversations with it. So, literally, the testing was human being uses model, you know, runs tests at the speed at which the individual can type. And that was the state of the art of testing at one point. And I feel like we’ve come a long way, but I don’t think there’s widespread awareness of the ways in which we’ve come a long way. So just from, like, a technology-enabling toolkit, what is out there right now for test and evaluation when it comes to large language models? Or even when it comes to, you know, stuff like autonomous vehicles?
Mr. Wang: Yeah. I would say – I would say that the two critical components are really – the two main branches are – the first is effectively large-scale scenario-based testing. So, you know, one of the things that we do – both for the open-source community, as well as more privately – is create large-scale benchmarks, which effectively boil down to large-scale situation-based environments where you can run the model, and you can see how well it performs against a pre-specified task. And you can do that – you know, you that in simplistic scenarios, and you can do it in very advanced scenarios.
You can have – sort of on the simplistic side, you can just sort of see if the model can answer a math question. On a very complicated end, you can – you can give it a complex legal review task. And it has to, you know, be able to look up things on the internet. It has to be able to be able to look through all your files. It has to be able to look through all these critical components and get – and ultimately arrive at it a correct answer.
Mr. Allen: So instead of, like, one Google engineer typing out the questions, now you’ve got this gauntlet of – I’m just making this up – like, a million questions that are constantly evolving, that’s on an automated basis, that’s very scalable and fast.
Mr. Wang: Exactly.
Mr. Allen: That’s one of the tools?
Mr. Wang: Think about running it through, like, a million simulated environments, and seeing how it performs in all million simulated environments. And then you augment that with large-scale live production monitoring capabilities. So, you know, this is the general truth of AI systems. They will encounter edge cases in the wild. And they will encounter new scenarios in the wild, because the reality is our world is incredibly chaotic and encompasses a lot of problem – surprise. And so you need large-scale monitoring capabilities to identify anomalies that come up in production traffic to be able to properly address them in production.
Mr. Allen: That’s great. So the toolkit is getting better all the time, and you’re a part of that story.
So also, in your testimony the other week before Congress, you talked about the importance of the network of global AI safety institutes. So, you know, in the United States we have the AI Safety Institute. In the UK, they have the AI Security Institute. I think there’s 11 of these government institutions around the world right now. And you were saying that this network that we’ve created is important. And I just want you to elaborate on, like, what is the role of this network, both in ensuring, you know, reliability, robustness, security of AI systems, but also of, you know, being a vehicle for U.S. leadership.
Mr. Wang: Yeah. I think these AISIs sometimes get a strange reputation, because I think people view them as, you know, these cells of doomers in every country that sort of, you know, is worried about the end of the world. I don’t think it’s anything like that. I think there’s sort of a – what they’ve evolved into, and what they are currently – continue to evolve into is a testing capability for, you know, this global network of countries. And over time, the interest to develop, you know, what are the reasonable standards for AI performance? What are the reasonable standards for how AI should look, how it should perform, how it should behave? And how do we ensure that it’s not only safe, but also performant and capable?
And so – you know, and really, truly, the U.S. AI Safety Institute, I mean, has the ability to lead globally. And I think we should view that as an opportunity, not just to, I think, you know, talk about strictly safety, but as an opportunity to really set standards in a in a global way. Because we have this global network. These other countries – you know, many of these other countries don’t have anywhere near the resources that we do. Don’t have anywhere near the capacity and the research capability and the sort of technical capability that we have in the United States. And so we have the ability to really set real standards globally in a way that will shape how AI is deployed in all of these countries.
Mr. Allen: Yeah. And, I mean, I’ve met the staff of the AI safety institutes of Japan, of Korea, of a bunch of other countries. They’re standards people. These are the standards people. They’re not the AI doomers of Japan. They are the standards people – the technology standards people. But for those who, you know, might hear this phrase, “standards,” and “we can set standards,” like, why does that matter? Why is it a useful thing for the United States to be engaged in, is helping lead on AI standards of this kind?
Mr. Wang: Yeah. I think there’s a few – there’s a few directions to look at. I think first, just from a pure economic standpoint, it’s very valuable for American technology to be as widely distributed as possible throughout the globe. And there’s a lot of – there’s a lot of economic value that will come from that. And, frankly, global economic value that will come from that – this process of exporting American technologies more globally.
I think the second piece of it is also, from our way of life and our values perspective, we have the ability to ensure that democratic values and, you know, many of the values that we have here in the United States are exported through the technology more globally. I think it’s impossible to look at AI as not a technology that’s deeply intertwined with the system of government and the ideology of the developers of the technology. And so it’s not only I think there’s a clear sort of – there’s two ways to look at the technology. One is you could have sort of like a pure economic lens in terms of, you know, having – helping to set the rails for this incredibly powerful and important technology wave, but also the ideological view, which is how do we help ensure democratic values prevail.
Mr. Allen: That’s amazing.
So we’re now in May, but in March Scale AI submitted its response to the White House request for information on the AI Action Plan. So what are the steps that you’re hoping this administration – you’re advising this administration to take when it comes to AI policy? I’m sure some of the things that we’ve touched on already, but what else would you highlight?
Mr. Wang: Yeah. I think – I think it mirrors a lot of what I discussed in my hearing with House Energy and Commerce a few weeks back. But you know, first off, as I mentioned, like, we are in this critical juncture, and the Chinese Communist Party has a clear plan. You know, they have – they have – there is a lot of intent and a lot of coordination there. And so we need to take this moment and boil it down to where are the areas where we must lead.
And one of the areas that I discussed was data dominance, which we talked about earlier. How do we ensure that the United States always leads in data?
Similarly, how do we have energy dominance so, you know, we’re able to ensure that we lead in energy and energy capacity?
You know, we need to make sure we get the sort of building blocks correct. One of the suggestions I made was towards building a national AI data reserve as part of an effort to sort of – you know, for the country and the government to view data as a strategic asset, fundamentally.
Mr. Allen: There’s this vague analogy to the national petroleum reserve, something like that.
Mr. Wang: Something like that.
Mr. Allen: Yeah.
Mr. Wang: Yeah. Data’s the new oil, as they say.
Mr. Allen: Yeah. (Laughs.) Yeah.
Mr. Wang: So that was one – that was one critical element.
You know, the other one is to really begin unleashing AI within government and to build towards an agentic government. You know, I think there we should expect that every agency throughout the government has real progress towards implementing AI across the board to actually drive real advanced capabilities.
And then, you know, the last one is really ensuring that we build the AI workforce for the future. You know, there was a recent executive order around AI education, and leveraging AI for education, and ensuring that we utilize it in this country. But fundamentally, you know, we’re going to – you know, it is very important that we prepare the country for this new technology wave. Like many other technology waves, there’s going to be incredible new opportunity that it creates. And we need to ensure that we as a country are prepared for it on the front foot of building this new workforce.
Mr. Allen: So it’s really interesting that you say that because, you know, I can’t count how many times I’ve heard AI analogized to Sputnik moment, right? DeepSeek was a Sputnik moment. AlphaGo was a Sputnik moment. ChatGPT was a Sputnik moment. And if you look back at the history of Sputnik, there was a bunch of different things that happened after Sputnik. Some of them were, you know, legendary moves. Creating NASA, good move. Creating DARPA, good move. But one that is often forgotten is the National Defense Education Act, where they said, holy heck, we are in a multidecadal science and technology competition with the Soviet Union, and we need to quadruple the number of engineers that this country produces. So, you know, to hear from you, right, that preparing our workforce for this AI revolution is just a really lovely, lovely note to end on.
So, as you can all imagine, Alexandr’s time is exceptionally precious. But – and I hope this is still true – he is offering to be exceptionally generous as well, and so we’re going to immediately after this have a reception outside. And I – as I said, his time is precious so I can’t promise he’ll be here terribly long, but I’m so glad that you will be here. And I hope all of you join us who have come to watch his remarks for the reception outside.
So, with that, please join me in thanking Alexandr Wang for coming to CSIS. (Applause.)
Mr. Wang: Thanks for having me. (Applause.)
(END.)
