The AI Arms Race: How Artificial Intelligence is Reshaping Intelligence, Warfare and Global Power

[00:00:00]Hello there, and welcome to the State Secrets podcast. I am Suzanne Kelly, CEO and publisher of State Secrets.

[00:00:04]It's great to have you here. I am very eager to introduce our guest today because we're talking about artificial intelligence and how it's really become the defining technology of our time. But beneath the headlines about sort of the chatbots, the image generators, there is a much bigger story here. AI is rapidly changing how intelligence is collected, analyzed, and acted on. It is reshaping literally military planning, accelerating scientific discovery as we know. It's transforming cyber operations, and really fueling a new strategic competition between the world's leading powers. And that's why I am so excited to welcome Chip Boucher.

[00:00:47]Chip is someone who has spent his entire career thinking about a lot of the challenges that AI is tackling and presenting. Um he is a former CIA officer. He now serves with the Special Competitive Studies Project, where your focus, Chip, is really on the intersection of emerging technologies, national security, and America's long-term strategic competitiveness. Is that fair?

[00:01:12]>> That is absolutely fair, and it's great to be with you again, Suzanne. Thanks for having me on the show.

[00:01:17]>> Thank you for being here. I I I really want to kind of cover the basics with you first because I've got some deep dive questions for you today when it comes to national security, but artificial intelligence, Chip, I think for a lot of people is still a bit of a mystery. Um they kind of understand it it might help you, uh you know, collect your grocery list. It It might help you decide which grocery store has the cheapest whatever you're looking for. It might give you the directions to get there. Like there are there are very basic things that AI has pulled from a lot of the technology that existed. Um how would you describe AI though just on sort of a basic level for how people are using it today, and then on a national security level, and what AI means for national security.

[00:02:05]>> I'll try to keep it kind of basic, but there are there are two big flavors of artificial intelligence that we are beginning to use at more larger and larger scale. The first is generative artificial intelligence, gen AI.

[00:02:24]These are the large language models and the chatbots that probably most of your listeners have already experienced some degree or another, and they are indeed the ones that can help you with your shopping list, your your menu planning, help you plan your vacation, and that sort of thing. Where you input either typing or voice, you know, a query, and it and it taps into its vast body of training and comes back with with some answers.

[00:02:55]And that that is one important area, and then the other that's uh sprouted on the scene really in the last 12 to 18 months is what is known as agentic AI, and these are artificial intelligence systems that take action at your direction or on your behalf in the world. So, not just answering a question, but actually following guidance that you provide it, taking steps. Those steps could be not just giving you recommendations about what restaurants to go to, but booking a reservation for you.

[00:03:34]Booking flight reservations for you.

[00:03:37]Filling out forms.

[00:03:39]All all sorts of things.

[00:03:41]That is called agentic AI.

[00:03:43]And they And the the US national security community, you know, has need for both, right? So, generative AI is definitely useful as the I see things about how AI systems can kind of run over its vast corpus of classified and unclassified information to gather insights and and generate generate insights. And then for agentic AI systems as well, they are potentially useful for the intelligence and national security security communities as they think about using AI to, you know, perform intelligence missions, perform national security missions.

[00:04:25]That that could include targeting. It it could it could include you know, other cyber activities.

[00:04:31]There's a a lot of potential uses out there.

[00:04:34]>> Yeah. I you know, I I'd love to kind of walk through some of this with you because I I think what we have seen, Shep, just in the past 16 months, 18 months, it is such a dramatic change in the quality of the information that's coming back as well. So, for example, I I think the phrase is hallucinations.

[00:04:55]When AI might get something wrong, it may have misunderstood a question or it may have included data that was out of date depending on where it's drawing from, the large language model that it's using.

[00:05:07]I'm curious. If you had to describe where we are now versus where we were when Deep Seek was released. That is, of course, the AI model that China put out that was a lot cheaper, a lot faster, but it wasn't really hailed as being that great at the time. Where are we now? How far have we come since we saw the first kind of inklings that there was going to be real competition with a United States national security adversary, China, >> Right.

[00:05:42]>> when it comes to AI.

[00:05:43]>> So, I think you're asking about sort of two things. One is, you know, what is the state of the art now and how good is it for AI systems and then how how good is China in comparison to the United States? And I'll I'll tackle those in turn. You're You're absolutely right. I mean, we are at a transformative moment, an inflection point, I think, in terms of the capability of AI systems. I think this really was driven home just this year with the release of Mythos by Anthropic, which demonstrated such a breadth of capability as to really capture the attention of the White House, of the Pentagon, and of the US intelligence community. Here was a a an AI model that was fully capable of performing very significant, very complex cyber missions all on its own. Okay?

[00:06:42]Um these models uh are showing themselves very, very useful on the battlefield in the Iran war in Ukraine in the fight against Russia's occupation uh of Ukraine for targeting, for computer vision, for uh campaign planning, for the management of autonomous systems. They are now so good as they are having real-world national security and and intelligence applications that professionals uh you know, like you and I sort of appreciate and and want to make use of. To the point where the White House, when Mythos was first released, right, put a put the clamps on it and didn't want it in the hands of uh of uh foreign persons until a a safer version could could be created.

[00:07:37]>> I I was just going to ask. You're like You're like my personal AI bot. You're You're so smart about all this stuff, but explain to us to what guardrails are in AI and why guardrails are so important.

[00:07:51]And throw a third question on there, the differences in how the United States looks at guardrails versus maybe the way China or Russia look at guardrails.

[00:08:01]>> So, guardrails are limitations that the model developer bakes into the algorithm when it when it creates an AI system.

[00:08:12]So, right now Anthropic created Fable 5 as a guardrailed version of Mythos. So, it it has many of the smarts that Myth Mythos does. It was built from the same stuff, but it has in place on it limitations. And those limitations are things like if you query it to help you design a bioweapon or, you know, tell me how to subvert democracy, it will not respond, right?

[00:08:46]It will tell you that that is out of bounds. You hit the guardrail and that it will either refuse to answer or it will revert you to a more simplistic model to give you a more simplistic answer.

[00:09:03]And this this is important in the competitive landscape because while Chinese models have some guardrails, they are guardrails meant to serve the Chinese Communist Party's interests, not necessarily what you and I would consider the appropriate guardrails.

[00:09:24]And they they tend to be lower and they tend to be kind of focused on preserving the myths that the CCP likes to perpetuate about itself, Chinese history, not talking about what actually happened at Tiananmen Square, things of that nature.

[00:09:43]So, that's you know, what what that issue is about.

[00:09:47]Meanwhile, China has been advancing its capabilities at a very rapid pace.

[00:09:53]They tend to follow a different technical approach. They call it open-source models.

[00:09:58]The models can be downloaded, the weights examined by by users. That's that's open-source.

[00:10:06]US frontier model developers are closed-source. You cannot do that with Claude or or with ChatGPT.

[00:10:15]But in Chinese models you can, and they aggressively make use of a technique called distillation in which a junior model learns from querying a a more capable model to see how it thinks and how it reacts.

[00:10:32]And that's what China has done to great effect to rapidly improve the intelligence, if you will, of their models by learning from Opus 4.8 or ChatGPT 5.5.

[00:10:47]And doing it at a at a much lower cost and a much less requirement for advanced compute capabilities.

[00:10:56]So, there's disputes about the ethics of that. All model developers use distillation to some form or another, but it's definitely a violation of the terms of service that Anthropic and OpenAI have in place. And the aggressive and pervasive way that Chinese model labs are pursuing distillation, it really speaks to a you know, like a an industrial espionage level attack on our model developers to learn from from our frontier systems so they can do it cheaper and faster.

[00:11:34]>> But no one can really stop them from doing this, right? It's And And And there lies the problem from a national security perspective as well. Let's talk a little bit about how you see AI impacting three particular areas, Chip.

[00:11:50]How do you see AI impacting war?

[00:11:53]How do you see it impacting intelligence?

[00:11:56]And how do you see it impacting the economy?

[00:12:00]>> So, we we have entered It's already happening in places like Ukraine, Iran, Gaza, you know, we're seeing it play out. We are we are we have entered the era of algorithmized warfare.

[00:12:14]Where adversaries will be relying much more on autonomous unmanned systems.

[00:12:24]And artificial intelligence both to to guide those systems. Um I think of drone swarms, for example.

[00:12:33]AI-enabled drone swarms that are able to find and fix and attack targets.

[00:12:40]But then also using AI for all the other requirements of managing a military campaign. So, planning, targeting, uh making sure your logistics train is in order and safe from attack.

[00:12:57]So, this is this is happening in in alongside it is the algorithmization, I guess, of the intelligence that goes along to support war fighting and and the national security missions beyond war fighting.

[00:13:14]So, it's it's going to just accelerate from here on out.

[00:13:20]>> Yeah.

[00:13:21]And And the impact that AI is having on intelligence, talk to us a little bit about that.

[00:13:26]>> There's two ways that artificial intelligence is impacting intelligence, and one is it's it's changing and expanding the list of issues and topics that the intelligence community needs to focus on. And that's that's a tough tough thing to swallow, but it's it's absolutely the case. We now have a stake in understanding how good Quinn 3.8, when it's released in a few days, how good it is. And we we need to know what is what is beyond 3.8. What's what's being cooked up in the labs in China. So, um not only in terms of the algorithms, but but the chips, the chip designs, uh that that China's is trying to create, the innovation power of foreign foreign nations is increasingly of consequence and importance because the the the struggle between nations uh has always had something to do with technological power, but a nation's ability to field a capable AI system, autonomous systems, robotics, field eventually quantum computing, these will these will determine the balance of power in the years ahead.

[00:14:47]So, it's a it's very much a target.

[00:14:50]And the second way in which AI matters is what tools the intelligence community is is making making use of.

[00:14:59]So, we at the Special Competitive Studies Project have done a couple of deep dive papers that have examined how AI can be used for strategic warning and other analytic missions.

[00:15:11]We have talked uh about how AI can be used on collection, right? I wrote a paper uh entitled the digital case officer. One could actually imagine that um you can collect foreign intelligence from human assets digitally. Right? You approach people on social media websites or via email or via text or a video call and you're able to actually manipulate them and get them to share classified information.

[00:15:45]And there's not a not a human case officer but a digital case officer and a system behind it that that does it. So, it's both a target and as a tool.

[00:15:55]>> It's crazy to think about what the future is going to look like. I I sometimes allow myself to try to think, you know, 50 years down the road and I'm not really sure I have a clear picture of 10 years down the road at this point.

[00:16:06]What do we know?

[00:16:08]>> [laughter] >> What do we know so far about Qwen 3.8?

[00:16:11]It's a Chinese model. Um what are you expecting? What do we know about it and what do you expect?

[00:16:15]>> Well, uh officially we we don't know a whole lot. Uh they haven't released uh performance statistics against benchmarks. They haven't even released some of the the details on uh the context window size and and the parameters that it uses. But, it's it's rumored in in some of the um you know, online chat fora that I read that it it's going to be very very capable, probably on par with close to uh Mythos Fable 5 um ChatGPT 5.6 soul.

[00:16:48]Um that that's part of the story. That's that's the capability side.

[00:16:53]The other side that's going to matter greatly is the cost. And And this is this is something that, you know, we should talk about a little bit and along with distillation, um you know, China is able to develop systems that users can use to create effectiveness but at a fraction of the cost, right? So, Fable 5 uh depending on how you how you how you look at it, the the token per token cost is about $50 for the output.

[00:17:23]>> And a token is like a very small unit of information.

[00:17:27]>> increment of what you get from a from a generative AI system. It's like a a part of a word or part of a sentence is an individual token or fraction of an image, okay, or sound sound bite.

[00:17:42]Um, it's it's the the the the basic increment of what a a generative AI produces.

[00:17:48]Uh, American models were built at great expense and the cost charged to the users is quite high. Chinese models, by comparison, are very very cheap. Um, you know, one one model out there is less than $5 per output token.

[00:18:04]So, a tenth of of the cost.

[00:18:07]So, surprise, surprise, more and more users, including in this country, are pivoting to make use of Chinese models because of the use costs being so much less.

[00:18:21]This plays to China's advantage as it tries to gain market share.

[00:18:27]>> Right. And the sort of obvious, I guess, elephant in the room there, too, Chip, is if more Americans are getting their information from a Chinese language model or a Chinese AI model, um, then they're being fed some information, maybe not all of it, right?

[00:18:44]>> You have the >> There there are very good reasons that um, non-Chinese companies, non-Chinese citizens, I would say even Chinese citizens need to be concerned about uh, Chinese developed models.

[00:18:58]Uh, they are cheaper. Uh, they are open source, which means you can bring them on prem and and and develop them, tweak them, um, in a ways that that suit you.

[00:19:09]But, [clears throat] we We entirely know what comes inside that Trojan horse. Are there uh backdoors that the Chinese have uh government has insisted that the labs put in place that exfil data or have they put in place guardrails that are not apparent to the user that you know, prevent getting a fullsome accurate truthful answer to all your questions?

[00:19:35]Um so there good reasons to be concerned about that. Um and uh so companies need to look uh uh closely uh before they import them.

[00:19:48]>> I love that you used the word the term Trojan horse, too, cuz having just seen The Odyssey, which I hope everyone in America gets to see Christopher Nolan's new movie. Um >> [laughter] >> there there are some things that have changed and some things that haven't in terms of the tactics and strategies, but the technology is a whole lot different.

[00:20:05]That's a great a great phrase to use to help folks, I think, understand what we're looking at here, sir. Um as we as we kind of close out this conversation, it probably needs to go on another hour, so I'm definitely going to have you back. But you're kind of setting the table for us and and just understanding the basics of AI. So I am looking forward to our next conversation. But as you're thinking right now about how quick the technology is um developing, about the different users who are able to develop it easier, cheaper, faster, what do you think is the number one concern and I'm sure there's more than one, but that the United States needs to be thinking about from a national security perspective as a lot of these new models are released? How can the US government how can uh democracy protect itself in this new technological age?

[00:20:57]>> So as I was alluding to earlier, when it we're in a different place now uh with regard to the capability of these AI systems to include Chinese developed systems.

[00:21:07]So think back just a few weeks ago, the White House was restricting access to Mythos. We still can't, as ordinary citizens, access Mythos. We can access Gable 5, but not the original system that that was released.

[00:21:23]Because of fears, justifiable fears, within the White House and the intelligence community about what Mythos could do. Well, now the world is going to gain access to Chinese developed models that are very close to Mythos capability.

[00:21:42]So, the protection already is is being eroded. And so, there are a couple consequences of this I I want to talk about. One is just in the realm of self-cybersecurity, uh if we if we haven't been there for the last couple years, um we really are in this place now of like near constant cyber warfare.

[00:22:05]Uh persistent.

[00:22:07]These models will be out there and uh adversaries, starting with China, starting with Iran, will be using them to creatively use AI for persistent AI-enabled attacks and disinformation attacks at a near constant rate. And they'll be very sophisticated systems. And really, the only way to defend against them is for defenders to also use Mythos quality uh highly capable AI systems to defend our systems at speed and at scale.

[00:22:40]Uh and the idea of of some sort of deterrence model working in the cyber realm, I think, is rapidly rapidly eroding.

[00:22:49]The The other thing to think about is [clears throat] uh we've always talked about uh comparing nation-state to nation-state in terms of its military power, its economic power, uh maybe its social capital, you know, strong of a political system does it have? AI power now counts, too.

[00:23:08]Uh, you know, this was shown in really stark terms in the Iran war when the Iranians figured it out pretty quickly that their data centers in, uh, Bahrain and UAE that were were being used by the Gulf countries, potentially their militaries, they attacked them.

[00:23:25]That infrastructure is now highly consequential.

[00:23:30]It needs defending. It needs understanding, you know, how we're going to field these systems and how we're going to protect them.

[00:23:36]And, uh, a nation's AI power is incredibly important to understand.

[00:23:42]And the And the balance of power, as a result, is going is going to shift. You know, very interestingly, a few weeks ago, the, uh, the head of Argentina, you know, announced this this new policy that AIs could gain personhood in Argentina. And and with you know, property rights, uh, making a bid for AI power by sort of lessening governance restrictions.

[00:24:10]Uh, other countries are inviting data centers. In meanwhile, in our country, there's a backlash against data centers.

[00:24:17]Not in my backyard, uh, projects are being canceled.

[00:24:22]And, um, you know, we've already talked about the the cost per use difference between what China's developing and what we're developing. So, this this this contest for AI supremacy is going to enter this new phase where before we could kind of pat ourselves on the back and say, "Our companies are developing the most advanced systems." Yes, they're pricey.

[00:24:47]It's going to cost hundreds of billion dollars to build out the data infrastructure, but it's all worth it because we will own the frontier of what is possible.

[00:24:57]That reality just got upended.

[00:24:59]So, what will matter now going forward is who can secure, deploy, and operationalize AI systems across the economy faster and better. And across the economy to include our national defense architectures.

[00:25:19]So, it's this diffusion is what they the technologists call it. How quickly this technology diffuses across the economy.

[00:25:28]That is the race that's going to really matter going forward.

[00:25:31]>> I think that's a great term all of us should know moving forward. Chip, thank you so much.

[00:25:37]I also want to just stress for folks, you know, you have done a lot of uh work on this at the Special Competitive Studies Project. Can folks go to your website SCSP uh to to find out to access all this?

[00:25:50]>> Absolutely.

[00:25:51]Uh www.scsp.ai.

[00:25:55]Uh we uh if you go to our work, you will you will see an intelligence tab and and see some of the things that we've written about AI, how it's being uh used in the intelligence community or should be used by the intelligence community, and uh as a target. And uh Suzanne, as always, thank you very very much for the invitation. I uh love The Cipher Brief, and everybody should be reading it.

[00:26:18]>> Well, you're very kind, Chip. We I always love talking to you. I always feel smarter I just for having the opportunity to ask a lot of my questions that, you know, I I I think really need context um folks to have to better understand what's happening and why it does or and how it does impact them. And we really appreciate you taking a few minutes today to help us figure that out. Thanks so much, Chip I share with the Special Competitive Studies Project.

[00:26:42]Great to see you.