Jake Sullivan谈如何应对AI不确定性及管理对华竞争

Hello, and welcome back to The Cognitive Revolution. Today, my guest is Jake Sullivan, former National Security Advisor to President Biden and currently the inaugural Kissinger Professor of the Practice of Statecraft and World Order at the Harvard Kennedy School.

This conversation was an exciting opportunity to unpack American AI policy and broader geopolitical strategy with someone who has shouldered intense decision-making responsibilities at the highest level of the U.S. government, and who answers critical questions—including about the nature of the threat that China poses to the United States and the nature of a stable equilibrium that the two countries might achieve—in the clearest and most concrete terms I’ve heard anywhere. I appreciated that tremendously.

After listening to and reflecting on this conversation, I would say that Jake embodies what I might call the very best of the establishment. While AI was a relatively small part of his overall portfolio for most of the 4 years he spent as National Security Advisor, he nevertheless took the topic seriously enough to meet with leaders from all of the frontier AI labs and to seek a wide range of expert opinions. Agree or disagree with his conclusions, there’s no question that he is deeply informed and that his approach combines rigorous pragmatism with refreshing restraint.

While he does take for granted that the United States’ global leadership position is good and ought to be upheld, he is not pushing a broader ideology, angling for regime change, or making complicated second-order or bank-shot arguments. On the contrary, at every turn he seems to really recognize the profound uncertainties associated with AI technology and, very simply, as the cliché goes, tries to figure out how to maximize the benefits while also minimizing the risks, both from the technology itself and from the possibility that government action could end up making things worse.

As National Security Advisor, this meant, on the one hand, assessing the national security establishment’s readiness with respect to AI and publishing a memorandum that called for greater urgency in adopting AI technology. On the other hand, it also meant initiating an AI risk dialogue with China.

At one point in this conversation, Jake described himself as boring for focusing so much on the base case and for thinking so linearly. But I would suggest that this style simply reflects 4 years spent in one of the least boring positions in the entire world, where his job was more often than not to identify the least bad of a set of unappealing options. In any case, when it comes to international relations, boring is probably, in most cases, a good thing.

Managed strategic competition with China, however messy and unsatisfying, may indeed be our best path forward. And if nothing else, it is really good to hear an explicit argument that this strategy will serve us better than either a Cold War-style AI arms race or an attempt at a grand bargain.

At the same time, I do think it’s important to be mindful of the limitations of this generally sensible approach. While Jake is not at all dismissive of short timelines or existential risks, and does take scenarios like AI 2027 seriously, the standard policy toolkits and international relations playbooks don’t have much on offer to address the most extreme AI possibilities. This does create a real gap exactly in the cases where the quality of executive leadership might matter most—either because we need leaders to show wisdom and restraint under crisis conditions, which I do think Jake would do well, or because we need them to inspire the public solidarity required to navigate what could be shockingly disruptive transitions.

After we stopped recording, Jake was kind enough to say that he’s a fan of the show, which, given his schedule and responsibilities, I found both extremely flattering and honestly a bit hard to believe. Still, I took the opportunity to ask him how I could make the feed more valuable.

His answer is something that the entire AI safety community should hear: Go deep and get as concrete as possible on the specific details of the downside scenarios that you are most worried about, so that the people in power can have the clearest possible threat models and also the most concrete possible plans to address them.

I know that many in the AI safety community are trying to do exactly this right now. But it seems to me that we can’t have too much of that kind of work, and I definitely plan to feature more of it on the podcast going forward.

For my part, if I could be so bold as to offer Jake a suggestion in return—and it does seem quite realistic to me that he might return to a position of power just as transformative AI is hitting hard, say, in 2029—it would be to spend more time developing a positive vision and story for the American people.

Boring may be best in many scenarios, but as I so often say these days, the scarcest resource is a positive vision for the future. There’s a good chance that the current or the next administration will be called on to do more than muddle through.

With that, I hope you enjoyed this window into the worldview, mindset, and thinking behind the United States’ approach to AI and national security with former National Security Advisor Jake Sullivan.

Jake Sullivan, former National Security Advisor to President Biden and currently the inaugural Kissinger Professor of the Practice of Statecraft and World Order, welcome to The Cognitive Revolution.

Thank you very much for having me.

I’m excited for this. We have a ton to cover, so I’m going to try to be brief in my questions, which is not normally my style, and give you the bulk of the airtime.

I thought we’d structure this by just setting the stage. I think so many AI conversations go a little bit awry because people don’t necessarily have the same expectations of, or even understanding of, where we are in the technology today, let alone where it’s going to go in the future.

So, just to give you a chance to share with the audience where you’re at: How AGI-pilled are you? I guess that means, how soon do you expect powerful AI, and how powerful do you expect it to be? Where do you put that in relation to the other national and global crises that we have? And, for bonus points, how would you compare yourself to other parts of the national federal government?

It’s a great question, and like most great questions, it doesn’t have a totally straightforward answer. How AGI-pilled am I? I believe that the possibility of transformative, powerful AI coming in the very near future—that is, in the next couple of years—is something we have to take very seriously.

I took it very seriously when I was National Security Advisor. You have to build policy and make strategic decisions on the basis that that is a distinct possibility. But do I believe it’s an inevitability? Are we certainly going to see it by 2027, 2028, or even 2030? No.

That’s because super-smart people make the case that it’s coming very fast, and other super-smart people make the case that there are actually a number of things that have to happen, and those things may take quite a long time before we get to something like AGI, powerful AI, or even ASI.

I build a planning assumption that this is distinctly possible. If it comes—and when it comes—it will be transformative across every dimension: economic, societal, and, yes, national security. It is on par with the most urgent and important national security issues there are because it touches everything, from the future of warfare to the future of nation-state competitiveness to the future of non-state actors and the threats they pose, and everything in between.

It is a tough epistemic environment, with the divergence of top thinkers being just so radical. That’s a frequent theme of this show.

What would you say have been the things that have most informed your thinking? Maybe you can weave in a little bit of where you think Congress is on this, where you think the military is, and where the agencies are, because they’re probably all over the place.

Has it been private conversations with leading developer CEOs? How much has just getting hands-on with the technology itself informed your thinking? How much are these manifesto-type documents impacting the Washington discourse? Where are people taking their cues from in Washington?

As National Security Advisor, I really tried to cast a very wide net. I met with the leaders of all the frontier AI labs. I met with technologists who are AI-skeptical. I met with investors. I met with academics.

Of course, then I convened and engaged with the entire national security enterprise of the U.S. government and said, “What are we doing to integrate and adopt AI capabilities to advance America’s national security?”

In fact, toward the end of my time as National Security Advisor, we put out a national security memorandum, which sounds kind of boring but was a lengthy document. In fact, the Trump administration has not discarded it. It sent clear directives to all of the national security agencies of the U.S. government, including the Pentagon and the intelligence community, but also Treasury and Commerce, the State Department, and so on.

What it said was, “Right now, America’s in the lead at the frontier. We could debate by how much, but we are behind the curve when it comes to integrating and adopting these capabilities and applying them for national security purposes. We’ve got to move faster, with greater urgency and with greater dexterity.” I think that remains the case today, more than 6 months after I left office.

Are there people inside the Pentagon who see the promise and the opportunity here and also see the threats and risks that come with advances in AI? Of course there are—brilliant people. But culturally, bureaucratically, and institutionally, are there obstacles to taking advantage of the full suite of capabilities that are there today and increasingly will be there in the future? Hell yeah, there are.

And that's something that I think takes leadership from the White House to try to break through. We did that particularly in the later years of the Biden administration. I see evidence that there is some push along those lines from the Trump White House as well.

Oh, and you asked about Congress. I should say a word about Congress. I think, at the moment, it's hard to talk about it monolithically because there are members who get both the opportunity and the challenge, who recognize how transformational this is going to be and are thinking hard about what role they should be playing. But I would say, in the main, when you look at the pieces of draft legislation, the hearings, or the public statements of most members, I think there is still a gap between where Congress is and where it needs to be in terms of a sense of urgency and priority for this issue.

Yeah, that sounds safe to say. For people who want to change that, there's this notion in the AI safety community that if we can just get some time with members of Congress and show them our scary demos—you know, show a chatbot that will help you make a bioweapon, or at least attempt to—and that will snap people to attention and get them to engage more seriously. I wonder how much those sorts of awareness-building efforts have changed the overall landscape in your mind. Has any of that reached you, or have you seen an impact from that?

Yeah, look, demos, going back years now to early in my time as national security advisor, definitely made an impression on me. You can describe what an LLM can do. I can coherently understand that, but when you actually see the capability live and in color, it does have a different kind of impact. And that probably is the case on the Hill.

But I would submit to you that, generally speaking, with complex areas of regulation and legislation—and this is among the most complex there is—Congress really only tends to react when there are real-world impacts. That could be a catastrophe like 9/11, which gave us the Patriot Act, in part for good and in part for ill. But that also could be a more cascading set of developments in the world, like job displacement in key industries or announcements by other countries of military capabilities that give them advantages over the United States. And it will be a collection of those moments, in my view, that is really what's going to get Congress into action, more than just demos of the tech.

And that's not great because it means that we're going to be playing catch-up as opposed to being ahead of the curve, but I think that is the world we're living in right now. I would also say that the Trump administration has sent a very clear signal: They don't want to see legislation. They don't want to see an effort to impose something from the Hill, or even from the executive branch.

Even setting aside that question, it seems to me this is a very difficult area to be confident about what is right from the point of view of rules, standards, trust, and safety. And coming back to your first question, we don't know where the tech is going and when. So, it's very hard, sitting here today, to dictate things too confidently in light of the profound uncertainty that we're contending with. That's why I think preparation and flexibility—being ready, developing a set of rules and standards, and developing a set of tools that Congress or the executive branch could use as necessary as we go forward—is the right thing. Choices about when to deploy those, based on how we see the tech evolve, should be made in careful dialogue between Washington, Silicon Valley, and the rest of the country.

I think you're absolutely right that sitting down today and trying to write a long list of rules or standards is an ill-fated undertaking. We've done a couple of episodes recently on more creative attempts to try to bring dynamism into the regulatory environment, with things like the SB 813 proposal in California, where there would be some sort of marketplace for regulation. Just going a little bit deeper into your AI worldview, though, I know it's been your job for the last 4 years to anticipate and, ideally, prevent a lot of bad things, and then respond if necessary. Have you had time to develop any sort of positive vision for the AI future? Are you imagining a new social contract? Are you signing on to Bernie Sanders's call for maybe a 4-day workweek? What, in your imagination, does the upside of all this look like?

One of the things about the way that I think is that I tend to think in quite linear and concrete terms, and so I have a hard time with these big abstractions about new social compacts or about recursive self-improvement leading to a nation of geniuses in a data center. These kinds of concepts, I just have a hard time wrapping my head around. So I tend to think of things in a pretty straightforward, linear way, and I see 4 big buckets: security, economics, society, and then these existential questions, where there are both risks and opportunities.

Our job should be to identify the risks very concretely. So, in the security space, there are non-state-actor risks like cyber and bio. In the state-actor domain, there are risks around wonder weapons and military advantage by one of our adversaries or competitors. But there are also opportunities in the security realm: AI could actually enhance America's national security if we remain in the lead on it and deploy and adapt it effectively.

Similarly, in the economic domain, there are genuine risks of job displacement. There are genuine risks of concentration of wealth and power that come from essentially the benefits flowing to a very few if we're not careful. That is a big risk. And then there are massive opportunities for productivity and problem-solving, and on issues around climate and energy or public health, or what have you.

So what I try to do is think, all right, let's get very concrete about the risks, identify them, and make a typology of them. Let's think very concretely about the opportunities. And then, are there steps the government can take to increase or expand the opportunities and manage and minimize the risks? If the answer to that is yes, act. If it's not, then don't act. And that's really how I think of it.

Then there's this fourth bucket, the existential. And here I think it's just too imponderable at the moment. These kinds of things—the robot apocalypse or superintelligence leading to nirvana and to space—you can tell the very positive story and you can tell the very negative story. For me, I'm really focused on the first 3 buckets more than the existential questions right now.

What can we do in the security domain, the economic domain, and the society domain? And on society, the risks of misinformation and alienation are real. But the opportunities of giving every person access to teachers, doctors, nurses, and personalized medical care—these are exciting opportunities.

So that's how I think of it, basically, in a kind of boring way, to be honest, but in a way that I think the debate about AI is not entirely contending with. I've asked research assistants of mine here at Harvard, "Get me a set of articles that actually walks through what the real risks are and what the real opportunities are, beyond the headline or the whiz-bang or the doomerism." And honestly, there's not that much out there that really walks through this in a deep and coherent way. I think more of that kind of analysis could help us really bring a greater degree of rigor to the conversation about where AI is going and where government action should go to try to make sure that it works more for us than against us.

Yeah, that's interesting. It leaves me thinking that I take the existential concerns more seriously. Certainly, I would bet my P(doom) is higher than your P(doom).

Just to be clear, I take them seriously. It's just that, right now, I feel like I am personally less equipped to give definitive answers to how to deal with the existential questions than I am to deal with immediate issues, like the cyber risk that AI could pose in 2025, or how to think about dealing with the fact that, whatever happens overall with growth, productivity, job gain versus job loss, a lot of people are going to be disrupted and will need a better form of social insurance than we've currently got on offer.

To me, I want to put more energy and emphasis into working through those questions because I actually believe that if we answered those questions effectively and designed a set of approaches that allow us to manage risks and opportunities in those buckets, it will set us up better to deal with those bigger questions of misalignment, AI risk, P(doom), what have you. I'm not trying to shrug those off or say they aren't significant.

I'm just suggesting that racing right to that question, which I think so many people do, takes you past things that are very live challenges right now, that are far less imponderable, far less contestable, and have far less of a spectrum of where people fall on them. And we've got to do the hard work of building consensus around those issues. I think, in doing so, we'll build muscles that will put us in a better position to handle the existential.

Yeah, that does seem to depend on a certain timeline assumption. I guess part of how I've operationalized my thinking is that I do take pretty seriously the idea that Daniel Kokotajlo and the AI 2027 folks could be right, and one of these frontier developers just might set off an intelligence explosion in the next 2 to 3 years. But then I'm like, there's probably not much I can do about it if that is going to happen.

And so, maybe I just have to try to stay sane and play for longer timelines and more mundane scenarios. It sounds like you're saying the same thing but with a little bit less weight on the short timeline.

I guess what I'm saying is we have to take seriously the possibility of AI 2027. I not only read the piece, I shared it extremely broadly. I called people and said, “How right is this?” I talked to a lot of technologists who said, “This is a totally credible scenario.” I've talked to others who said, “It's less credible.” So, I of course take that very seriously.

I guess my point is, sitting here in 2025, let's assume that is the direction we're heading. I still believe that a more concrete and application-specific conversation about what we're going to do to manage particular risks and enhance particular opportunities will actually set us up better, both in the development of policy tools and in the depth and specificity of the conversation that can be had between the technology community and the policy community across this country.

I actually think that will help put us in better shape to handle that kind of extreme scenario if and when it comes. And sitting here on your podcast today, I'm not denying that could come in 2027. I don't throw my hands up and say, “Well, there's nothing we can do about it.”

What I'm saying is, let's work through the things we know are coming and are already here. Let's get good answers to those questions, and in so doing, I think it tees us up better for these kinds of deeper existential questions about intelligence explosion and takeoff and the like, all of which I spend part of my time both worried about and studying and talking to people about, because I do take it very seriously.

That's a perfect transition to getting a lot more concrete. I think we're talking about a week after the AI Action Plan was released, and a ton of stuff in there was broadly pretty well received, including by people who I think expected not to like it. What was your take? What would you say are the highlights and lowlights, from your point of view, on the AI Action Plan?

My overall reaction is that I thought it reflected a bipartisan consensus on a number of critical points, including the kinds of investment and support the U.S. needs to be providing to ensure we stay ahead of the frontier, including even in areas related to trying to build a global AI alliance and working with other like-minded countries as well.

I thought the way that they handled some of the risk and security issues, including on bio, was laudable. They re-upped a proposal that the Biden administration had put forward in the executive order on pre-synthesis DNA screening, for example. So, I thought all of that was to the good.

I thought there were some politicized elements to it, but that kind of goes with the territory of the way this current administration operates. My main concern with it actually was the gap between what it says about needing to stay ahead of China in this race and deny them access to high-end compute, and then the reality of what the Trump administration is actually doing with this H20 decision and other things. So, my biggest concern is that I think there's a dissonance between the AI Action Plan and what the Trump administration is actually doing, and that's a concern to me.

Yeah, I definitely want to unpack the whole export controls and China thing in some depth in just a second. Another big question I have around all these action plans in general, with a sort of “Are we actually going to achieve abundance or not?” background context in mind, is how likely we should think it is that the government is actually going to adopt these technologies.

Last I heard—and you can tell me differently—from folks in the Biden White House and some folks in the Trump administration, there's still no ChatGPT or anything similar at the White House, right? I think people are going home and running Deep Research queries on their own computers and then trying to memorize those results and bring them back into the office, and that just seems like not a great baseline from which to lead the national charge toward global AI leadership. So, yeah, how do you think about that disconnect?

It's a disconnect. I acknowledge it. I am now an avid—“avid” would be an understatement—user of these tools. I used them much less frequently when I was actually national security adviser, even though I was working on the issue every day.

There are a few reasons for this. The culture of government in terms of new technology adoption is historically slow, and that's true in this case as well. There are legal issues, particularly in the national security enterprise, especially if you're talking about things on what we call the high side and classified compute and the like.

But in general, I think it's imperative that, whether it's the White House or the agencies or beyond, the adoption and incorporation of these tools get accelerated, because you're right. I think it's axiomatic that you cannot effectively lead and govern on this set of issues if you aren't an effective user of them for purposes of advancing America's interests. So, I think we should see more of it.

It's not clear to me. I think this is going to be piecemeal to a certain extent, and there'll be some early adopters than others. But I would be an advocate for seeing this get incorporated as rapidly as possible.

How about the recent deals that President Trump announced when he went to the Middle East with the UAE and Saudi Arabia? My sense is that you would not have supported those deals, or that we wouldn't have seen similar deals under a Harris administration. It seems like a complicated business. I think there are a lot of arguments going different ways. What do you think are the primary considerations that people should have in mind, and where does that lead you to come down on those deals?

Look, for me this actually isn't too complicated. Number one, I would like to see the world running on American AI, not Chinese AI, and American chips, not Chinese chips. And that means diffusion. It means that we should be selling our chips and seeing data centers built around the world, including in the Gulf, where countries like the UAE and Saudi Arabia have a deep, vested interest in having AI be a central part of their national development enterprise. So, one, I believe in diffusion.

Two, we need security and high standards to ensure that there's no diversion of chips to China that would vitiate the export controls, and that you deal with things like insider-threat challenges, cyber challenges, physical-security challenges, et cetera. We should hold everyone to a high security standard.

And then, third, I believe the United States should not trade one form of dependence on any part of the world, including the Middle East, for another.

We were long dependent on the Middle East for oil. We should not become dependent on the Middle East or anywhere else for compute. So, I believe that the lion’s share of the compute build-out by American AI should be in the United States of America.

For me, the question in assessing these announcements coming out of the Middle East—and, by the way, that’s all they really were: announcements—is that the devil is in the details. It’s hard to know exactly what’s going to happen. I support selling chips to these countries as long as they meet the necessary security standards. But I support a limit on how many chips we sell because I think we should not end up in a circumstance where all of the future build-outs and all of the future training runs are being offshored and outsourced.

I think that would be repeating a mistake we’ve made too many times with past industries, and we shouldn’t repeat it here. I’ll be the first to say I engage deeply with the leadership of the UAE on their AI ambitions. I very much supported deepening partnerships between U.S. technology companies and UAE technology companies, and I continue to support that. But I think there again, the critical point for me is: let’s not end up dependent on others for access to compute in the future. Let’s make sure that we are building the lion’s share of the compute here in the United States.

I think that dependence thing seems very common sense to me. The other dimensions of the deals that I would highlight would be values and control. There’s all this talk—and we’re going to get to China next—but there’s all this talk about making sure that AI reflects American values. I think it takes some squinting, to put it mildly, to say that the governments of these countries really uphold or adhere to American values.

On the control side, this seems like it might actually be one of those things that you might do a little differently depending on how seriously you took the AI 2027, 2028 superintelligence scenario. You might say, “Geez, a 5-gigawatt data center in Saudi Arabia might be just the kind of place where somebody might set off a recursive self-improvement loop that would be outside your U.S. jurisdiction and control. Maybe we just shouldn’t set up the preconditions for something like that to even be possible, and yet it seems like here we go.”

I agree very strongly that we do not want leading-edge frontier training runs done in the Gulf or in a lot of other countries, too. I want them done here in the United States of America. I think it’s incumbent on us to ensure that we break the bottleneck on power so that it’s not a hard constraint on being able to conduct those training runs, and we should do them here in the United States. I very much agree with that.

When it comes to values, my view on this is that there are a number of countries around the world, including in the Gulf, that have different values when it comes to democracy and human rights than the United States does. We deal with them economically. We deal with them in technology. I don’t think that that is wrong. I just think we have to be clear-eyed about what we stand for and try not to directly aid and abet repression and things along those lines.

But you could take that argument to its logical extreme and end up really severing relations with a lot of countries around the world, and I wouldn’t be prepared to do that. So, I think it is a factor, and it’s something that should be part of the discussion and part of the conversation. It regularly is between our country and countries in the Gulf, but I have a hard time saying that I would put—I certainly wouldn’t put those countries or others similarly situated on a blacklist.

Why doesn’t the same analysis apply to China? It seems like a lot of what you said there in terms of different attitudes on democracy and human rights, but we still have to deal with them. We can still trade, et cetera. That could be applied to China.

But it seems like there’s a sense that there is some unique threat that comes from China today, and I always ask this question. I feel like I never get a great answer. What is that threat? How should I understand the threat as an American who grew up in the Midwest and has certainly lived through some of the deindustrialization problems that we’ve had, but I consider those to be in the past? Sitting here in 2025 and looking ahead, why should I be concerned about what China is going to do such that we can’t deal with them in the same way we can deal with these Gulf countries?

First, let me say that I believe we should not decouple from China. I believe we should have an economic relationship with them. I’ve been very much on record saying that we’re competing intensively with this country, but we’re also going to have to learn how to live alongside one another as major powers for the indefinite future.

So, you’re not going to hear from me, “You can’t deal with China at all.” I am focused on what I have called the small yard and high fence, where these high-end capabilities with national security applications can be controlled and restricted.

To directly answer your question—what should we be concerned about with respect to the PRC? I told you earlier I like to be concrete and not just speak in abstractions, so let me give you a few examples. First, you said the China shock is in the rearview mirror. The author of The China Shock, David Autor, and one of his co-researchers just wrote a quite compelling piece in The New York Times saying, “Get ready for China shock 2.0,” and it’s going to be worse.

That is because China is pursuing a strategy of massive subsidization of strategic industries and then trying to flood the market with cheap goods that undercut workers and businesses who play according to a different set of rules in the United States and elsewhere. If left unchecked, that could easily have the effect of further hollowing out our basic industrial base, including in the industries of the future. Economists call this overcapacity. I call it basically flooding the world with cheap manufactured goods based on a state-directed capitalism, where China’s not playing by the same rules as everybody else. I think that’s a challenge.

We don’t have to go to war to deal with that challenge, but I do think we need a set of measures coordinated with other countries in the G7 and beyond to push back against that. So, that’s one.

The second type of challenge is best exemplified in an episode that happened back in 2019, when the general manager of the Houston Rockets made a comment about freedom fighters in Hong Kong. The Chinese government went nuts and basically told the NBA, “You’ve got to shut this guy up.” The NBA essentially had to bend to the will of China on that particular point.

Is that a threat to every American? No. But if the entire world is run on Chinese AI, and China is effectively saying, “There’s going to be a price for your speech,” then that is a threat to the American way of life. China has shown a willingness and a proclivity to do that. So, that’s a second challenge.

Third, while I was in the job, I had to deal with something called Volt Typhoon, which has been publicly reported upon, and I can’t say too much about it. But I can say this: We have seen the pre-positioning of malware in critical infrastructure in the United States by the PRC. That’s a direct threat to water systems, electricity, pipelines, whatever it may be, and something for us to be very concerned about.

Why are they doing that? Fourth, China has engaged in the world’s largest peacetime military buildup, probably in all of human history. Why? Part of it is about Taiwan, which is its own risk and threat, and maintaining peace and stability across the Taiwan Strait, I think, has to be a paramount priority for U.S. policy because it would be totally catastrophic if that happened. But they’re also looking to project power globally in a different way.

So, what makes China different from other countries? One of the things is that they are the only country that actually has the attributes and capacities to compete with the United States, matched with an ambition to surpass the United States as the world's leading power economically, technologically, militarily, and diplomatically. And I think a world in which China is writing the rules, China is running the tech stack, and everyone is dependent on China is a world of greater coercion, less freedom, less good economic opportunity for Americans, and a greater possibility of China flexing its military muscles to subjugate others. So, that is my set of concerns.

Now, can this be managed in a way where we have a stable relationship with China? I engaged in intensive and deep diplomacy with my counterpart, Wang Yi. And I believe that we paired intense competition with intense diplomacy to good effect. I am also determined that competition not turn into conflict, and that we maintain space for cooperation on key issues, including, by the way, on issues associated with AI risk, where we inaugurated an AI risk dialogue with China to talk to them directly about issues that challenge both of our countries.

I'm not someone who sits here and just says, “It's Cold War time, baby.” Not at all. I do not seek a Cold War. What I seek is for the United States to maintain the capacity to ensure its way of life and not have that way of life undermined, pressured, or put at risk because of too much capacity and power accruing to a country that does not share our values and has shown a willingness to use its coercive power, frankly, recently against the United States with the rare-earth magnets thing, but also against other countries. So, that's a long answer to your question, but that's how I see the set of challenges that we're confronting when it comes to the PRC.

Yeah, I think most of that sounds very reasonable to me. You look at a giant military buildup, and one does have to wonder what exactly they're planning to do with it. I think you could offer benign answers, but in your job, it's not right for you to just accept the first explanation. I totally get that.

How do you think they understand what we are trying to do? I'm not a big analogy guy, but a short story of the run-up to World War II, at least in the Pacific, was like, “We cut Japan off from oil. They were like, ‘Well, we have a limited window here where we've got to go for everything or we're screwed. So here we go. We'll roll the dice and do this Pearl Harbor thing and try to push out and establish our perimeter, and hopefully we can hold it.’” Obviously, that was terrible for them and terrible for a lot of American guys who had to go fight that war.

I think it seems like it would be hard for the Chinese government, the leadership class, or whatever, to look at being cut off from chips at the same time that everybody's saying this is the new oil. It's the new big thing, the thing that's going to drive the next transformation, and not feel like they're in a similar spot where they say, “We're trying to hold them down, and this does seem like it is what we would be doing if we were trying to hold them down.” How do you think they understand it? Do you think they're getting it wrong, or are they, in that fundamental sense, getting it right?

It's interesting because the people who make that argument also tend to make the argument that the export controls don't work. It's an interesting overlap, and I see a pretty deep tension between those 2 propositions. Now, I think China, to a considerable extent, buys its own hype: that the export controls are futile because it's going to make more and better chips and it will be just fine. That, to me, disrupts the Japan analogy.

Do they complain at great length about the export controls? Absolutely. Do they say that this is containment and suppression and so forth? They do. Do I believe that they have decided they must go to war to deal with the export controls? No. I have not seen that.

I think the analogy has a certain elegance to it, but it requires someone to produce evidence that this is actually a targeted policy of saying, “The highest-end chips and semiconductor manufacturing equipment that the United States designs, we're not going to put in the hands of our competitor.” That, to me, is not the same thing as a full-on oil embargo and is a sensible national security measure.

Frankly, if you look at China, they're engaged in a whole series of their own national security measures, including various forms of export control on things where they think they have some advantage. And I don't really see anybody saying, for that list of things, which are pretty elemental, “Hey, that's a reverse Japan-U.S. thing that's going to lead the United States to feel it has no choice but to attack China.”

So, this is what competition looks like when power is increasingly measured and exercised in economic and technological terms. China has cards to play, and the United States has cards to play. We have to look at them and make determinations about how to play those cards, but I am a skeptic of the argument that we are inexorably pushing China to a position where it feels it has to lash out at us militarily because we have high-end chip controls.

Yeah. Again, it seems like a timeline thing, maybe. I would put myself in the camp of people who do believe that the export controls will be effective, at least in the sense of limiting what China can do in AI on some important dimensions, at least in the short term, as long as the AI 2027 or Dario Amodei timelines go.

If we find that there's no AGI in 2030 and we're headed into a longer-term scenario, then I'm not so sure, because I always say I've seen them put up a hospital in 8 days or whatever, so I don't doubt that they could build a chip industry in 5 years. Maybe not, but who can say.

It seems to me that there has been this evolution in terms of whether we're trying to keep them down and how they should understand it. You mentioned this term “small yard, high fence.” But it seems like this has really evolved, right? The story that people are telling—maybe not you, but you could tell me—is that the story I'm hearing in defense of the export controls has moved from, “We want to deny them these super-cutting-edge military applications,” to, “Well, maybe we can't do that, but we can at least deny them the ability to build frontier AI models.”

And now it seems like we're at, “Well, we can't really do that, either.” This is not an exaggeration: the latest stuff that I've seen coming from Leonard Heim, who's been almost synonymous with this export-control analysis at the think-tank level, is that hopefully we can field 10 times as many AI workers as they can, at least. We can't deny them these other capabilities, but we'll just have way more AI agents running in our economy than they will.

This does seem to me like it has shifted from a small yard, high fence to a general denial of the broad-based economic benefits of AI. Again, to me, it feels like if I'm them, even at the rank-and-file consumer level, I'm asking, “Is the United States trying to deny me my AI doctor?” That doesn't feel good. At what point do we actually become the bad guy in some sense? At what point are they right to understand us as trying to hold them down?

Look, the way I approach this may be overly simplistic, reflecting the fact that I'm not a technical expert on chips or AI. To me, it's common sense. You have a high-end technology that has significant national security applications. Some of those applications are exquisite, like in a weapon system. Some of those applications are more general, like building super-powerful AI that gives you a massive military and intelligence advantage over your adversaries.

But either way, it is an input for building national security advantage. We have in China a serious competitor. Why are we giving them that input to be used against us or our allies? We shouldn't do that. That's how I look at it.

Now, what impact does that end up having in the AI race? The way that I would put it is, I think it gives America a distinct advantage at the frontier. We should try to sustain that advantage because being ahead at the frontier, from my perspective, has a significant number of national security and strategic benefits.

So, that was, in my view, the original impetus behind the export controls: don't provide your deep competitor an input for its national security advantage to be used against you, but be discriminating in what you're denying so that it's not things for gaming applications, broad-based kind of stuff. And it is definitely the case that generative AI and large language models had an impact on how to think about that basic calculus of national security advantage, no doubt about it. It made us think differently about what exactly is inside the yard.

But the goal here, fundamentally, is still very much focused on trying to sustain and secure a national security advantage for the United States and not cede it over to China. That leaves this kind of dual-use issue that you flagged, which is a real one, right? These chips can be used for national security purposes, and they can be used for scientific, benign economic, and other purposes, too. The challenge we have here is that in China's system, they have a doctrine of civil-military fusion. Their technology companies have deep ties to their military. And so it's not easy to say, “Well, we'll give the chips for this purpose but not for that purpose.”

There's no way to actually enforce that in a meaningful way. And that is how we end up at the policy that we end up at. But are we trying to deny access to AI doctors in China? Absolutely not. And do I think the functional outcome of an export-control policy will be the denial of AI doctors to Chinese people? I do not believe that it will.

So do you think Leonard Heim is wrong when he says we can run 10 times as many agents? That seems to be the upshot of that, right?

Well, let's think about it this way. What's absent from that statement? I don't have the context of that exact statement, so it's hard for me to directly comment on it. But what's absent from that statement is what we started this conversation with: AGI, ASI, powerful AI—whatever you want to call it. Is that coming? Who's going to develop that capability first? To what end?

I would like to see the United States win that race. Do I think denial of access to high-end compute to China is an advantage to the United States in that race? I do. Where do AI workers in data centers come into that? I don't know. What I will tell you is that my logic on this is that the profound national security implications of being ahead of the frontier for the development of military capabilities, intelligence capabilities, and economic statecraft tools—not just broad-based economic growth, but coercive economic statecraft tools, among many other things—mean that I want to see the U.S. there.

If I've got an input that we have some control over, I'd prefer not to hand it over to my competitor so that they can get there first and use it against us. That's kind of the baseline logic behind the controls. In this respect, I think the Japan analogy is really not that strong, because what we're talking about ultimately with super-powerful artificial intelligence is a kind of transformative national security technology.

Let's say we had it. Leave aside the chips. Let's just say we had that technology to be used for particular national security purposes. Would we hand it over to a competitor of ours? No. We'd keep it for ourselves, the way we do with every other type of technological military advantage. Now, China has caught up to us in a number of areas in the military, but we didn't go hand them the jet engine and hypersonic missiles and so forth. We said, "That's on you. You have to figure that out for yourself. We're not going to help you in that regard."

I see a similar dynamic at play for the very high-end compute, given how central to the future of national security the question of frontier AI would appear to be.

One statement that has really been a burr in my saddle for a while is Dario's proposal from his "Machines of Loving Grace" essay. He basically says, "We should use these export controls to keep our lead, build an international alliance," which you alluded to, "but then eventually we'll get around to basically making China an offer they can't refuse." Or, in his words, quote, "Convince them to give up competing with democracies in order to receive all the benefits and not fight a superior foe."

To me, that basically reads like regime change. Would you say that's going too far? We shouldn't adopt that as our policy. How do you think about the fact that one of the 3 or 4 top AI company leaders is pushing for that?

So, I've been explicit that I don't believe regime change in China should be an object of U.S. policy. In general, either way, we should shape our own policy. Our job should not be to shape China's government. We have to deal with the government as it is and try to design a strategy to ensure that America's security, economic, and other interests are protected and our values are protected.

I am not a regime-change guy when it comes to China. There are others who are. If we went down the list of prominent technology leaders and their views on various geopolitical issues, we'd find a lot of different, quite interesting things, many of which I agree with and many of which I disagree with.

So, I don't take a lot from what Dario says. I have huge respect and admiration for him as a person. I think he's an ethical person, and I think he's passionate about democratic values. There's a way in which that, to me, is admirable. I have a different point of view than he does when it comes, it sounds like, to the policy question of regime change.

I guess one other very big-picture question on China is how much hope or room you have in your scenario or world-building for the possibility that the regime there could change. That could be through an actual regime change, or it could be through evolution, perhaps fairly suddenly in terms of the way that their government is acting. We've seen these dramatic reversals on COVID lockdowns, for example.

I was really struck recently by the release of the Kimi K2 model, which is leading some creative-writing benchmarks in English and also comes from this company that has this very Western-loving aesthetic. The company's called Moonshot. It's named after "Dark Side of the Moon." And it just has me thinking, "Geez, maybe all this engagement that we've been doing for the last few decades actually has worked on a much deeper level, where there's potential for real friendship, in a serious way, between the 2 countries."

Maybe that just hasn't quite hit a certain level of leadership yet, but maybe it could. Am I too fanciful for hoping that something like that could surprise us on the upside?

That is not my base case. In fact, what you've seen is a deeper consolidation of power under Xi Jinping, the centrality of national security in all decision-making, with their definition of national security being social control at home, and the deepening and strengthening of the Chinese Communist Party under the grip of Xi Jinping. That is what I think you have actually seen over the last few years.

My base case is that that's likely to endure. But, of course, I'm absolutely not ruling out that your optimistic scenario could come to pass. It could. My main point is that we should make policy that is not directed explicitly at a particular institutional or systemic outcome inside China. We should say we'll watch and see what happens with the Chinese government and its relationship with the Chinese people.

What we can do is set policy that is not about changing China from within, but rather about protecting and defending our interests and creating an environment that is more conducive to and leans in the direction of American interests and values, and does not end up allowing Chinese interests and values to dominate. So, could what you're saying be right? I can't rule it out. I don't think that the evidence really points in that direction.

I'll say one last thing, because you said, "Could we be friends?" Look, I believe that at the people-to-people level, we should have a lot more exchange than we have today: American students in China, cultural exchanges going back and forth, all of it. I believe there's no reason why, even if we're competing intensively, we cannot deepen the connections at every level of our societies between our 2 countries.

Look, we're 2 big, ambitious, dynamic nations with very proud peoples who share a lot in common and can benefit from engaging with one another. And I am all for that while remaining clear-eyed about the nature of the government in Beijing.

How do you think the rest of the world is looking at this competition? We are trying to export chips, obviously, and have some newly announced deals to do that with at least some countries with deep balance sheets to pay for them. China's open-sourcing its models at the AI event in Shanghai last week. The premier—no less than the premier of the country—showed up and said, "We're committed to making AI an open-source public good for the entire world to benefit from."

It seems to me like, if I were in countries 3 through 90, or 3 through 193, I would find that to be a pretty compelling vision. How do you think the soft-power competition is going for us right now?

Well, I think of it in 2 respects. One, I do think that this is a soft-power move by the Chinese that is meaningful. We want to give you really cheap or free, high-quality AI. People are going to like that.

On the other hand, I think most countries are pretty unsentimental. Another element of soft power is innovation and technological prowess, and they're looking at who's actually got better stuff. If it turns out it's basically equal and so forth, then, yes, they're going to want the easier, cheaper, open one.

But if they're looking and saying, "Hey, the U.S. is pretty impressive. A lot of countries are betting on the U.S. to stay in the lead and even extend the lead in the AI race over time," that's its own form of soft power. I saw that when I was National Security Advisor, and I continue to see that: countries basically saying, "We like the American tech stack because it's damn good, and we think the United States is going to stay at the cutting edge."

I think it's our job to make sure that that remains the case. If it does, I think the soft-power piece of this will take care of itself.

As you think into the future broadly, can you envision any sort of stable equilibrium? I struggle to get to a point in my imagination where we have, obviously, the sort of MAD equilibrium for the nuclear threat. I feel like we're on this track to create a new AI Sword of Damocles that will hang over all of the future as we race to defend ourselves against one another.

And yet, I still can't really imagine a version of that that's super stable. You probably, I'm sure, have seen the Superintelligence Strategy document from Dan Hendrycks and coauthors that talks about MAD. Yeah, I did find that. I applaud them for looking for some sort of stable equilibrium. It didn't ring true to me as something that seemed like it would be super stable.

Do you have any sketch of a stable equilibrium in the presence of powerful AI between these great powers?

I do. I think it is a version of managed competition where we presume there's not going to be an end state where one side just wins and the other side loses, where we have to come to terms with the fact that we're going to live alongside one another as major powers. There is going to be intensive competition, jostling, and a little bit of elbow throwing, but with sufficient guardrails so that intensive competition doesn't spill over into conflict. I think that is manageable.

That is basically the blueprint that we pursued over the back half of the Biden administration: compete like hell, but intensively manage that competition so it didn't tip over into conflict. I think that can be sustained even in the face of increasingly powerful AI. I also believe that powerful AI presents opportunities for direct engagement between the US and China to manage risks that afflict both of us. That is kind of a modern version of arms control.

On the one hand, we're building our capabilities; on the other hand, we're talking to each other about both arms control and nonproliferation. We did that quite effectively with the Soviet Union over decades. So, would I use the phrase “stable equilibrium”? Maybe not. Do I believe that we can avoid great-power war without giving up on competing vigorously? Yeah, I think we can do that.

I think we can keep competing vigorously and avoid great-power war. It requires stewardship and deep, intensive diplomatic engagement between the two sides. I think we've shown in this iteration that this is possible. If we continue to build the muscles for it, I think it can withstand even the advent of very powerful artificial intelligence. How's that for optimistic?

I would love some optimism. Does it require us to mutually swear off the pursuit of strategic dominance, though? Because it seems like if there's an expectation that's mutually held that each side wants to get into a position of strategic dominance, then we're in an arms race scenario and we can't really trust each other. This stuff is much harder to verify, it would seem, than anything in the nuclear realm. You can see the silos from space, but who knows what's going on in any given data center, right?

It seems like as long as there's this notion that we're going to try to achieve strategic dominance—and that we might achieve it, and they might too if we don't—then I have a hard time imagining how that doesn't create some really serious risks. Maybe we can manage them, but we're creating a lot of problems to manage if that's the shape of the competition we're going to be trying to manage.

My answer to you is going to sound a little bit evasive because, frankly, all good diplomacy is, at the end of the day, a little bit based on evasion and ambiguity. But I think some degree of uncertainty in this phase is manageable and is better than explicit declarations one way or the other on concepts like strategic dominance. I think it's just too soon to tell exactly what this world is going to look like, and the band of possibilities is too wide.

There's an expression about getting across the river by feeling your way across the stones. That's basically what I would say we have to do here, rather than try to impose a framework today borrowed from deterrence theories that relate to nuclear weapons or from something else. We should have a few tentpole concepts: We would like to stay in the lead when it comes to the frontier of AI, and we would like to ensure that no other country has capabilities so advanced beyond what we have that they can just put us at risk and there's nothing we can do about it.

We need to keep the lines of communication open and make sure that there is no mistake, miscalculation, or inadvertent escalation. Let's do all of that, and then I think these questions around enduring advantage and strategic dominance will sort themselves out as we come to understand exactly what it is we're dealing with here, which I don't think we understand right now. I think we should front-run all of that with some decisive declarations about our policy and strategy.

Maybe the same answer, but is there any version of a grand bargain that you could imagine between the 2 countries now or in the near future that would take us off the trajectory we're on of decoupling and intensifying competition and release some of that pressure?

I think President Trump is giving some thought to his version of what he will call a grand bargain. I'm not sure exactly what it will contain. I'm pretty skeptical of the grand-bargain concept because I think, at a fundamental level, both countries are going to keep competing. As long as that competition is a feature of the relationship, it will need to be managed, and it can't just be solved and put away.

It is a chronic condition of the US-China relationship, one that I don't think is susceptible to the cure of some strategic condominium, G2, or grand bargain. So I think we're better off in a world of managed competition than either going for defeating them or doing a grand bargain where everything will be good. That makes me a little boring, I confess, but I think it's the right recipe.

All else equal, I think I prefer boring people in power. What have we learned about AI and the future of conflict from recent conflicts? My sense has been maybe not that much yet, because it still seems like we've got wires to the drones and people are controlling them with PlayStation controllers, rather than handing that sort of stuff over to truly autonomous systems. What would you say are the big lessons?

You're catching glimpses of the future without being able to fully see it in the present. One element of it is just scale and attritability, just a different form of fighting wars than we've seen in the past, and we're beginning to see how that plays out on the battlefield in Ukraine. This is where abundance—the notion of abundance as a core construct of modern war—comes in.

Another is that you are beginning to see the incorporation of various AI capabilities into some of these drone platforms. There are still humans in the loop, but you could now glimpse a world in which you could have fully autonomous weapons and see how that could actually play out. We're going to have to grapple with the lessons of that in American military doctrine.

Then you have to think about this well beyond the war-fighting phase. It's intelligence, logistics, command and control. Here, I think we're beginning to see the incorporation of AI capabilities in that conflict, and that's something that we can draw some lessons from as well.

My biggest concern, fundamentally, is that the United States military—the Pentagon—because of the bureaucracy, the Congress, and the defense primes, is not necessarily poised to adopt AI capabilities at scale and with speed. I think the PLA is better situated to do that. That makes me nervous, and it should be a kick in the pants for all of us on a bipartisan basis.

Cool. This has been great. I really appreciate the time and thoughtful discussion. I do genuinely mean it when I say I think, all else equal, I'd rather have boring people in power. Anything else you want to leave people with? It could be the biggest questions you're watching, anything else we didn't touch on that you want people to understand, or we could just leave it there.

No. One thing we didn't cover as much—we touched on the economic and job-displacement stuff—is the question of how this impacts the lives of everyday people: how they make a living, how they provide for their families, and how they find meaning and purpose professionally. These are really profound questions that I'm wrestling with and trying to get a better handle on, in addition to everything else that we've talked about today.

As usual, I find there are a lot of different answers from a lot of different people about where this is all headed in terms of the impact on our economy. This is the first technology I can think of with such profound national-security applications that the government really had very little to do with. It's essentially a private-sector-led and -driven technology.

That's uncomfortable for a guy who thinks about government policy, and it's made more uncomfortable by this thing we've been talking about: how wide the band is of predictions about the timing and scope of AI capabilities. That's just the world we're going to have to inhabit and do our best in, and have a can-do spirit about this rather than just stick our heads in the sand.

Yeah, totally. For what it's worth, I think if the political class can get the international relations right, I'm personally very optimistic about the working class's ability to figure out what to do with the peace dividend and all the extra free time that they might have. All right, fair enough. Fair enough.

Thank you again. This has been great. I really enjoyed the conversation. Jake Sullivan, thank you for being part of The Cognitive Revolution. Take care.

Thanks a lot.