ベン・ブキャナンはAIで米国に勝たせたい。だがそれは「競争」か「罠」か

If you were back in the White House now, what would you be doing knowing that as soon as March or April, we're at this handover control to AIs that we don't know how to control, and in fact, we don't even know if we fully regain control? If I were in the White House now, I would want a lot of significant regulation of the AI companies with federal regulators that had the capacity to go deep in this technology. I would want mandatory pre-deployment testing of the AI systems before they release to the public. I would want embedded evaluators to make sure their safety and security practices for development are up to snuff. And I would want security standards for the labs to make sure they're not hacked by foreign actors and they're of the sequence stolen and the like. So, yeah, it's fair to say that we are now in the moment that I think many of us theorize would come. And that was a moment where we would want significant regulation of the labs pursuing this technology. So certainly I think our actions revealed this was something we were taking very seriously back then

and something that if I were still in any kind of position of power, I would be advocating very strongly for right now. Hey, everyone. Welcome to Your Undivided Attention. It's Tristan Harrison. Now, this week, Aze and I spoke with Ben Buchanan. Ben served as President Biden's senior advisor on AI. And in this role, he was in the room where it happened. whenever the AI company's CEOs came to Washington, he was there, and he helped write the first federal policy on AI, including the original export controls governing GPUs going to China. Ben also had a seat at the table for the historic Geneva Summit in 2024 when the U.S. and China first sat down to discuss AI risk and AI safety. Today, Ben is a professor of international studies at Johns Hopkins University, and he's an advisor to Anthropic on AI safety. Aiz and I first met Ben at the White House in the spring of 2023, and that was really when the world was just waking up to AI. It was shortly before the public release of GPT-4, and I remember Aiz and I meeting lots of people in the government who really just didn't get it,

who were coming up to speed on AI for the first time, who didn't understand scaling laws, until we met Ben. And then it was very humbling because he was viciously intelligent about this topic and plans for how it could go well. So today Ben joins the show to talk about the new book that he co-wrote with Tantum Collins called The Bitter Struggle, Super Intelligence, Superpowers, and the Fate of the World. And in the book Ben makes the case that in order for us to avoid the worst case scenarios on AI, it's essential that open societies like the U.S. win the AI race compared to closed societies like China. And if you listen to the show, you know that this is a framing that Iz and I have been quite critical of. is this really a race between two countries, or is it a trap that will ensnare all of us? So we wanted to have Ben on to make the case. Ben is also an expert in cyber hacking and AI, which is very relevant because this last month has been called Black October because of the incredible surge in cyber hacking events all around the world, and we'll explore that topic

with him as well. This was truly a great conversation with one of the world's foremost experts on AI policy, who also happens to be a dear friend of mine and a very good man. You can find Ben's book today wherever books are sold. Ben, welcome to your invited attention. Thank you so much for all the work you guys are doing, and thank you for having me. So you served as President Biden's top AI advisor, and specifically, you were the only official, as I understand it, who was at every single White House meeting with the Frontier AI CEOs, including in January of 2023 when Sam Altman demoed GPT-4 before its release. And it's a story you write about in your book in the opening chapter. I'd love just to bring our audience into that moment. Can you paint a picture of what was that moment like in January 2023 before the world really understood what this is going to be about? I remember this moment very well. It was in the diplomatic reception room in the Eisenhower building. Sam was coming to meet the National Security Advisor, Jake Sullivan, and they had a really

good conversation that I think made real a lot of the things that folks like me had been talking to Jake about up to that point. And this was a moment when the United States had already taken some action on AI. And I think people like Jake deserve a lot of credit for being willing to be forward leaning on things like chip controls and the like. even before we had ChatGPT come out in November of 2022. But I think seeing GPT-4 up close helped convince Jake and helped convince the broader set of U.S. policymakers who had so many things to deal with besides AI that AI was going to be really important. And certainly for me, as someone who had felt for a long time that AI was going to be really important, seeing GPT-4 and everything that came after it made me realize that the future we had all anticipated was really starting to arrive. Just curious, what was the unlock? What demo did Sam give? What capabilities? Just to kind of take people back, because now we have GPT-6, which people can reverse engineer all of Adobe Photoshop and make a replica in an instant.

What were the capabilities back then that felt impressive or kind of breakthrough? Of course, now, as you say, they all seem totally mundane. But at the time, it was crazy that GPT-4 was multimodal, so it could do image analysis. You could give it an image and look at it. I remember something with a press conference where it had analyzed a press conference of Vladimir Putin's and had spit out the points of it and the key questions for intelligence analysts could ask. So it was a well-crafted demo, and it prompted a really good conversation between Jake and Sam and the whole group of us there that I think was, it really felt like the meeting of these two worlds that I'd had one foot in either side for a long time, the world of Silicon Valley and technology and then the world of national security. they were now entangled. And I think for better or for worse, they're entangled forevermore. And so looking back three years later, the world has completely changed. And I'm just curious, how does that meeting land in your kind of mind's eye now in retrospect,

given where we are? I think it feels to me more symbolic from a policy perspective than from a technology perspective, certainly in retrospect. And for folks like me who had lived this technology for a long time and who came to the White House in 2021 with a very deep conviction that AI was going to matter a lot for national security. GPT-4 felt on trend. It's always striking to see things like this, but it felt on trend. But it did feel to me like a moment where the U.S. government was now going to move into a new dimension of what it could do. And indeed, that's a lot of what happened, that the U.S. government in the spring of 2023 reorganized itself. They created this position. White House Special Advisor for AI, which I assumed, and there's a lot of direction and opportunity to get things done through the bureaucracy that otherwise wouldn't be there. I don't know if I would peg it all to that meeting, but I think I would peg it all to that moment in time, the spring of 2023, when the U.S. government said, wow, this is not just a national security China issue. This is an issue that's going to affect everything,

foreign policy and domestic. This is something I've always really appreciated about your work. I mean, obviously, you've been very deep in the Biden administration, but you also view AI as trans-political, and you work with people on both sides of the aisle, including, you know, President Trump's former AI advisor, Dean Ball. And so I'd love for you to just tell us about some of those relationships that cut across all administrations. And, you know, the New York Times op-ed you just recently wrote. anyone listening to this show who knows your guys' work knows that these issues are going to affect everyone and this is not a partisan thing and it shouldn't be a partisan thing so when we were in the government I think people like Turun Chhabra deserve a lot of credit for setting a tone of we're going to work in a bipartisan fashion on these issues especially on the national security issues where there's a lot of alignment with Republicans and then when I transitioned out of the government of course with the 2024 election I made it a point to try as much as possible to go to folks in the Trump administration and say, look, we're not going to agree on everything.

I'm no fan of Donald Trump. But you guys now have, I think, a very solemn and very sacred responsibility. And I'm here to help you however I can. I've lived this for four years, and some of them were open to it. Some of them were less open to it, for sure. But you mentioned Dean Ball. I think Dean, who's now at OpenAI, gets a lot of credit for someone who, I'm sure we don't agree on every issue, but was really willing to look at the data, hear from folks who have worked on this. And Dean and I have done a number of panels together and a New York Times op-ed together, as you mentioned, that try to show there's a lot that unites us on this one. And I think that's important because there's a lot we can agree on here in a bipartisan way. To add here, Dean has called you a tireless and ardent patriot. I just want people to hear this conversation as not just one political side of the story. This is about people who are moving between administrations who have this solemn and sacred responsibility. So, well, you wrote this book with your co-author called The Bitter Struggle, and it's out now, and I want to make sure that we're covering, essentially,

the front row seat that it gives us to the AI policy discussions and decisions that you made in the White House. But I think to start with the first principles, what is the bitter struggle? The bitter struggle is a play on this idea of the bitter lesson, which is the key technical idea in AI. And we see the bitter struggle as having three parts. The first is for the United States and democracies everywhere to have an advantage in inventing this technology. We can talk about why that's so important later, but essentially making sure that this is an area in which the technological course that humanity charts is one that democracies have a very large say in. The second is taking that advantage and converting it to a national security edge and saying, can we make sure that the United States can use this technology well for intelligence, military affairs, assure its position in the world? And then also, though we don't want to talk about the economics of this, Canada apply it to its economy as well in a way that's productive for national power. And then the third is really the crux. And the third is, can it do all of these things in accordance with our democratic values?

And this is where the bitterness comes in because there are some very difficult trade-offs here that have to be managed. You've got really big, powerful companies, but that raises the risk of concentration of power and monopolies and the like. Another danger, as you guys know very well, is the misuse of this technology. How can it be misused for cyber or bio risk by bad actors? Well, we can crack down on that, but that also creates the risk of infringing on civil liberties and the like. We want to go fast, but also if you go too fast, you cut corners on safety, which creates misalignment risks and the like. So no single book in AI is going to have all of the answers. But I think the strength of this book is to take folks in the room, say this is what happened in the Biden administration. This is what's coming next. And this is what we see as being at stake as all of humanity and especially all democracies wrestle with this new technology. Yeah. Well, some of what I hear you saying is that technologies sort of have an inflection. They lean towards open or they lean towards close. And that one of the unfortunate things, sort of the bitter struggle,

is that this thing that open societies are creating to strengthen themselves is actually forcing us, in some sense, to lean towards closed kinds of systems. And just to give a really specific example from the last couple of weeks, now that we've had PHP bugs undermining 70% of the Internet, banks all across the world getting hacked by open source models, it is the openness of our models that is forcing us to move to some kind of surveillance state, which is terrifying. And AI enables something that's never been possible before, which is listening to everything, everywhere, understanding what everyone is saying, how people are operating. You can do the flop cameras at scale. And so ubiquitous technological surveillance becomes almost trivial, right? And so the temptation for governments to do ubiquitous surveillance becomes almost insurmountable. And so I'm just curious, what is your wisdom from sitting in government and what you see for how open nations, especially as we get the perfect storms of more and more people losing their jobs, at least in the transition and their livelihoods and disenfranchisement going up, people losing political and economic power.

How do nations resist the temptation of total surveillance? How do we stay open? I think this is a challenge that democracies have wrestled with for every technology since the dawn of America and at the simplest level AI is going to force us to reconsider how we've done that and find a balance again and I think especially part three of our book is kind of at least an articulation of what are the different parameters that we have to balance here we don't pretend to have all of the answers but can we put them all on the table and lay them flat for us to look at but the internet forced this as well and things like the FISA surveillance act forced that Even the Fourth Amendment at some level is trying to reckon with how do we have the privileges and the dynamism and the freedom of an open society while also making sure we keep people safe. And I think AI is just bringing back these questions in a very acute, very novel way that government, to be clear, is not ready for and many of the companies are not ready for. I'm very stressed about the way in which AI can entrench autocracy, not just in autocracies, but also could have backsliding in democracies, including in the United States.

You mentioned flock cameras. I think that certainly is one vivid example. This was a thing we worked on in the government, and the president's executive order in 2023 had a whole tasking to the Department of Justice to think about the appropriate use of AI in the criminal justice system and in law enforcement. So it was a policy question for us, and it's a big theme in the book. And in the third part of the book, we talk about democratic values. We talk about what we call the bitter temptation, which is democratic governments, maybe for legitimate reasons, maybe for not legitimate reasons, sliding into this world of using AI to repress their citizens or otherwise entred power. And I think there's a sense in which AI can compound or concentrate power. and we are allergic to it in the sense of the companies being too strong and the governments being too weak and also allergic to it in the sense of the government being too strong and the citizens being too weak. There's a very narrow path here between the two. I don't pretend to have wisdom or the answer, but I do think at a minimum what we've done with this book is articulate the problem in a fair amount of detail

and what we tried to do in the government was set the foundations of policy to keep us in that narrow path. I think it's really important to pause for a second and just define the bitter lesson, because this is so central and so counterintuitive to most people that to think paradigmatically AI really requires internalizing the bitter lesson. And so could you just describe what that is, and then sort of from the emotional sense, like why do scientists think it's so bitter? The bitter lesson is very closely associated with a technological idea called the scaling law, which has dominated AI over the last 10 or so years. and that's if you've got a general purpose algorithm, in this case it's something called the transformer, then you can scale up its capability with computing power. And the more computing power you put into training the system, the more powerful the system that results. This was a controversial idea. We only show up in the White House in 2021. And there's a lot of people there saying, no, it's all about the algorithms or it's all about data.

And I think it fair to say that really the entirety of certainly our national security policy on AI if not our entire policy on AI was premised on this idea that the scaling law and the bitter lesson would be right and that we would see very powerful AI systems due to scaling up of computing power. Thus far, that's what's happened. Just to take people back, not everyone would have believed that if you just keep scaling up masses, 10x and 100x of compute, that you're going to go from something that can barely spit out essays to solving Navy or Stokes and hundreds of math problems and creating material science breakthroughs and better hackers than the NSA, mostly by just scaling the amount of compute that's actually there. And you were one of those believers, it sounds like, in 2021. You just take us back to, like, what had you believed that, whereas many other people didn't take that seriously? Because there's a similar kind of disbelief in the ecosystem now that are we really going to go from a GPT-6 to a GPT-9? I mean, just think about how much more capable. I think it's important for building people's intuitions, right? Because here we are, where GPT-6 is kind of the most frontier model-ish.

And we're trying to assess, like, where would we be in two, three years if we don't steer in any way? And it's as hard to imagine that as it was to imagine really sitting in the shoes of 2023 and GPT-3, GPT-4, like, that we would really have what we have now. For me, the moment that I was convinced was the jump to GPT-2, which was, I think, June 2019, and then the jump from GPT-2 to GPT-3. GPT-3 was 100x scale up. And May, I think, of 2020, GPT-3 comes out. Think about where America was then. COVID lockdowns. George Floyd was murdered the next week. A lot of chaos in America. but I do remember reading the GPT-3 paper and saying, wow, this creates a pathway for us to go really far as a species in AI. And in some ways, this converts AI development from a scientific process, though it is certainly that, to an industrial equation. Can you acquire enough computing power and energy to run the chips?

If you can, you can almost predict where the systems are going to be. And it's worth fast-forwarding, Tristan, as you said, to where we are today and thinking about how much the scaling law has grown since then. And now, if you look at the top five AI firms, are spending, I think, $700 billion on computing infrastructure in 2026 alone. That's akin to a Manhattan Project every two or three weeks. That's more than the United States spent building the entire interstate highway system over 35 years. Just a mind-boggling amount of money coming from private firms in just one year. That's the scaling law on the Bitolesta Network. This is basically $100 billion roughly per year per lab of the major labs doing the scale-up, which is really illustrating this GPT-6, GPC-9 thing. And you were just mentioning one thing that distinguishes the technology. You compared it to the size of the funding of the Manhattan Project. But the Manhattan Project was created by the government. The government had the expertise and the frontier knowledge of nuclear science. The private sector didn't have that. Whereas you make this distinction in your book, this is the first technology that is so transformative,

but that's coming from the private sector. Do you want to speak to that? That's right. And this is, I think, the through line for every AI policy decision that we made in the government, was this is the first technology, probably the last hundred or so years, that really comes from the private sector. And that's in contrast to, of course, nuclear weapons, a lot of the decades of space exploration, early days of the internet, early days of the microprocessor, early days of jet aviation. So much of those previous revolutionary technologies, and of course their attendant national security implications, have their roots in Department of Defense or Intelligence Community funding. Oftentimes the project did the work, the universities did the work, but it was the government that cut the check. And what that meant was threefold. One, the government had the capacity to understand it because someone's accountable for it. Two, the government had some capacity to shape it. And three, they often were born with national security instincts in mind. So protection from counterintelligence threats and the like. None of those are true for modern AI. And so this becomes a race for which country or which company can hit the fast forward button. Who can move faster through time to get the bigger breakthroughs,

the more power, the more military advances, the more economic advances, more GDP. So this is a race through time with capital and compute. Now, we've been critical of the framing of a race because it has certain externalities and consequences for thinking of races. But I think it's really important to talk about your framing, open versus closed societies, which, of course, really means U.S. versus China as the two big superpowers. and just flesh out the full race, give the full best argument for why we need to be in a race with China for AI? You won't usually hear me use the word race, but you certainly will hear me use the word competition. And the theory of the case that we had since we got there in 2021 was that it would be important for the United States on behalf of democracies and that's somewhat on behalf of humanity to have a leading role in shaping AI development. and that we felt it was the democratic system that would be best positioned to manage this technology. What that meant, what that conviction meant, is that we wanted the lead to be substantial over China,

often by slowing China down rather than accelerating ourselves unduly, which I think is a really important distinction here, to build a very big lead such that we could, to use our phrase from 2021, spend the lead on safety. So 2021, there were no real dangers of these systems, and we knew that. But we wanted to use that period, essentially the Biden term, to build the lead such that when we got to 2026 and 2027 and these systems, as we thought, became more capable and also more dangerous, there wouldn't be a geopolitical arm trace with China. It would essentially be a one-player game. And democracies would be alone. The United States would be the leader of the democracies. And then we could, without the geopolitical pressures of competing with the Chinese, regulate the AI companies and say, look, we're going to slow down here. We're going to make sure we're not cutting corners on safety. All this kind of stuff. within the U.S. context, where it's just American companies that are doing this. And I think that theory continues to be right. I think the evidence has borne that theory out. The execution has been imperfect, especially in the Trump administration. But I think the theory that we have, and certainly the pitch of this book,

is you want American and Democratic preeminence of this technology. You don't want a close arms race with China. And then you want to make sure you're regulating your company so they don't cut corners. If you want to put this in historical terms, I think the analogy that resonated very well was one that was near and dear to President Biden's heart. Off the side of the Oval Office is the President's private study. And on the wall of the President's study for President Biden, he had President Kennedy's speech from 1962 at Rice University. This is a speech where Kennedy says we're going to the moon and coming back within a decade. And I think that speech matters less because of the moon promise, though of course that matters for history, but because of how at the end Kennedy's making the case to the American public, why they should support this. And Kennedy says something like, we don't know whether space will be good or ill for humanity, but only if the United States has preeminence can we help decide. And that really is the ethos that gets to our policymaking, of making sure the United States and democracies can shape how this technology goes for humanity. It's an interesting parallel here to something you hear from folks at Anthropic,

which is that at Anthropic we care the most about safety. And by the way, we can actually only influence safety if Anthropic is leading the race because basically we won't have the leading seat at the table or the dominant kind of position at defining the policy if we're not leading. So it's similar kind of parallel to the U.S. believing basically we can only set the terms or regulate AI. But there is a contradiction in there, which is that in order to race to be that far ahead, many times I have to take shortcuts, but there has been a cost. I'm just curious how you both stand on what have we gotten right about that framing and where, if you have any self-critiques around where those policies haven't played out the way that we thought. In the context of U.S.-China, I don't think that contradiction is particularly salient because the United States has the capacity to slow China down. So this was our theory, which is you don't need to race ahead and cut corners on AI development if you can slow the Chinese down. China is totally incapable of making advanced AI chips. You don't have to take my word for it. Take Secretary Lutnick, Trump's commerce secretary's word for it. He testified in

2025 China made 200,000 chips. That's not enough for a single data center. All those chips are also worse than their American-designed Taiwanese-made counterparts. So one of the critical actions of which I'm very proud, and again, a small group of us did this, was a series of export controls. Before Chad CPT, before Sam Altman came to the White House in 2023, that set out to deny China the access to advanced computing power because we felt like that would take China off the game board in making these systems and their military implications and the like. The Chinese companies themselves say out loud that the number one thing restricting their progress is having access to new technologies. I think Kimmy and DeepSeek have all basically said this. They're admitting that this is the key to their ability to continue to scale behind the U.S. They were saying that when we were in the White House in 23 and 24 as the compute controls were really starting to bite. And then really striking, there was a panel of them in an AI conference in January of 26, and they basically just went down the line, one AI executive after the next, talking about how important computing power was to AI progress and what a struggle it was to get

it. I think the export controls, especially in the last couple of years, have been deeply imperfect in terms of how they've been enforced and Trump reversed them and so forth. But even so, I think this is one of the biggest impositions on China. And this is not something that accelerates the United States unduly or anything like that. This is a policy choice to slow China down. And my argument is if you're someone who worries a great deal about AI risk, you really do want this to be a one-player game because it's much more tractable to regulate American AI companies than it is to regulate Chinese AI companies. Do you want to then speak to how we should consider the fact that Jensen Huang from NVIDIA has been the primary lobbyist for saying, hey, we should actually sell all of our American ships to China? Isn't that like selling uranium to Soviet Union to boost GDP for American uranium companies and say, well, at least the world's running on American uranium? It doesn't make any sense to me. In fact, I think it's even worse than the uranium analogy because the Soviets had their own independent ways of getting uranium. As we talked about, the Chinese do not have an independent way of getting computing power. So we're giving them the one thing they can't get.

To be fair, let me give what I understand to be the steel man version of his case. Let's do that argument. And certainly how NVIDIA has made this case to us and to the public. they will say, sometimes it's called the run faster argument, which is the United States just needs to stay ahead of China. One of the ways in which American companies stay ahead of China is we keep Chinese dependent on American chips and American technology, and we profit from those sales. And NVIDIA can take the money it makes selling chips to China, reinvest that money into its R&D ecosystem, and that will keep America running faster. Because we believe so deeply in the scaling law in 2022, Two, what we said to NVIDIA and what we said to everyone who made this run faster argument is, we've got more faith in your technology than you do. And we think there's going to be overwhelming demand for this technology and marry that to a fixed supply. And we want the demand to go to the United States and to democratic companies and countries. And the response, of course, is like, well, this is very bad for our business. You're taking away the Chinese market.

That's going to hurt us. We can't run faster. And Arby was, no, no, the AI tailwind here is going to be so strong. The demands of the scaling law and the bidder lesson will be so acute that your business is going to do great. And it depends on how you look at the exact stats. But roughly speaking, from the 2022 chip controls to today, I would bet NVIDIA is up by 15 or 20x in terms of market capitalization. From something like a $300, $350 billion company to now something like a $5 trillion company. If all of the Chinese AI companies are compute limited, why do we keep hearing about, but what about China? How do these two things fit together in the same sentence? I think the what about China argument can be disingenuous if it is an excuse for saying we don't need to regulate American AI companies or we should accelerate in a totally untethered way here in the United States. Where I think the what about China argument has some merit is we need to make sure we're actually enforcing the export control policies to make sure we build this lead and keep this lead such that we can regulate our companies.

So I am very clear-eyed on the risk that China poses to American national security. There's no doubt in my mind that if China had gone to mythos first and found all these cyber vulnerabilities, China would not be turning them over to software vendors to have them be patched to make the Internet safer. So we should not dismiss the risk here of Chinese intelligence services and the Chinese military, but nor should we use that as an excuse to say, well, we just have to be total accelerationists about this. The profound irony to me is that there's people like David Sachs, my successor as AI advisor, who simultaneously end up, I think, on the wrong side of both arguments and say, well, we need to sell the chips to China, which keeps the Chinese in the game, but also because the Chinese are in the game, we can't regulate here in the United States, and we should let this technology go totally unchecked. And to me, the opposite is true. By denying China access to computing power, by making it a game just for democracies, then you can regulate more effectively. So, Ben, you have a term for the place that you'd like to end up on AI, which is hegemonic stability. What does that mean, and what is your preferred landing spot here?

Hegemonic stability is a well-known term from the political science literature, which is that in moments where the world has stability and peace, it often is because there is a nation, typically they're not always a nation, with a technological advantage that has chosen to use that advantage to provide security and to provide stability to the system. And what that nation is doing is it's taking less benefit for itself and setting up a system for which all nations can benefit and in return getting stability. You can think, I think, of some of America's most farsighted policymaking was in the post-World War II era where America had the nuclear bomb unlike Europe, unlike Japan, was not destroyed, had this economic and political and military preeminence, but then channeled that strength into setting up a number of institutions that ultimately provide security for the world to this day. NATO, the UN, the IMF, the World Bank, and then many of the Cold War institutions that came out later, the IAEA and the like.

And I think our vision here for the United States is we want the United States to have preeminence and to work with democratic nations to have democratic preeminence. in AI. And we want the United States to channel that strength and that power into stabilizing institutions that are multilateral and that encourage nations, including China and Russia, nations like that, to say it's better to be in the system than fight the system. And I would argue a nation like China benefited enormously from the post-World War II and Cold War system the United States set up that enabled the global trade that let China rise in the way that it rose. so that's the closest thing to an end game I can see and I think there's a moment here in which the United States can build a lead for itself and for democratic nations and then can convert that to something that is stabilizing beneficial for all nations So I think now then is a really good time to set up the stakes of what does it mean to win the competition or lose the competition

what are the benefits of winning what are the dangers of losing why does this feel so completely existential There's many national security benefits from being ahead in AI. The one that's most obvious is cyber operations. This was the one we always thought would be first. I started working on AI and cyber really more than a decade ago depending on how you count because it seems so apparent to me that cyber operations would be an area in which AI would be quite good And we should make no mistake about this, that the United States and China compete every single day in cyber operations and playing offense and defense. And it all occurs below the threshold of armed conflict and all that. But it confers a significant national security advantage. And if you believe in American national security, you need to believe in American cyber operations and foreign intelligence and the like. I do want to say, because I recognize I'm coming off the China Talk, and I'm sure we can come back to this later if you want to do it in a different timing, but competing with China is not at all incompatible, in my mind,

with engaging with China. And the only competitive action that we took was simply saying, we're not going to sell you some of the most advanced technology humanity has ever made. It's not saying we're going to fringe on your territorial integrity. It's not saying you can't have a military. It's not saying you can't have intelligence agencies. It's just saying you can't have this technology. And that's totally consistent with what the United States has done for many nations with many technologies up to this point. So there's a way in which you can paint our policy as overly hawkish towards China. And I don't think it is. I think it was just identifying the one choke point and taking it away from them and then saying to them, look, AI is going to affect all of us. And we want to engage with you on AI safety and cooperation as well. Well, we're going to get to that May 2024 meeting where you were actually in the room with the Chinese in the first major AI safety dialogue between the administrations. We'll get to that in a moment. I just want to stay here for a second on the stakes. So one of the stakes that we also didn't mention is that the U.S. is in $40 trillion of debt. And the belief is that, I think this is part of the stakes, as seen from the kind of existential national security place, is the ability to grow our way out of our debt.

and that this is sort of the next Schumpeter curve of if we could genuinely get, you know, what some of the crazy scenarios of like you get 5, 10, 15, you know, 50% GDP growth per year because you have AIs that are literally doing as much scientific research and applications and medicines and new material science and inventing new weapons. But that's another piece of it. Do you just want to speak to that? Because I think to Aza's point, like I think to synthesize this conversation for listeners, it's like, okay, the essential point here is the bitter lesson and the bitter struggle is that in the future, competition is actually computation. Out-competing equals out-computing. And as part of that, the out-compute button is a fast-forward button for economic growth, for scientific advancement, for technological advancement, for military advancement. So I want to push that button to get a 16x fast-forward button and hope that you're only getting a 4x. And then chip controls is the way in which that competition is mitigated. But I want to make sure that we're setting up the stakes appropriately because I think that sets up kind of the real thing behind the scenes

that is then guiding what then we get lots of other consequences from that and then talk about the other states. The economic piece of this is very important. I have a little bit less personal knowledge and conviction here because I'm not an economist. And Tyler Cowher and other smart economists will throw cold water on this and say it's not going to be 20% GDP growth. It's going to be four, and that's a lot. And four is better than two or whatnot. But it does seem self-evident to me that AI will matter enormously for a nation's economy, and I want the United States to be ahead in that regard. It is worth saying that genuinely when we were sitting there in 2021 and 2022, we were not having the economic conversation. And we were not saying we were taking these chip actions against China to hurt their economy for economic sakes in that moment. If we wanted to hurt the Chinese economy, we would have controlled many, many, many more chips. And we tried to scope it to only the high end of chips to not decimate the broader Chinese economy. Because in general, I think we still had a posture of trade and engagement being good with really important guardrails on the Chinese context.

So I remember this meeting that Iz and I had with you in spring of 2023 when we first met you. You really were just the most viciously powerful mind on this topic that we had met. It was an intense meeting. I recall it being an intense meeting. I recall the White House fire alarm going off in the middle of the night. Yeah, it's true. As we were evacuating, we continued a very spirited debate going down the stairs of the Isle of the Building. The vicious spiral of it, yeah. The vicious spiral, exactly right. And I just remember this key moment where we talked about the three risks that were the biggest ones, and you said very clearly cyber, bio, info. Cyber risk, bio risk, and information, information warfare type, you know, misinformation, deep fake type risk. Starting there. Obviously, loss of control, things like that might come later, but that was not really present in the 2023 conversation. And you were very prescient about cyber. And as you said, just for Sail Man for Listeners, this perspective of the race, if you just take the idea that China could get mythos and hack all of our infrastructure, and we would have no defenses, and the quote, the race is justified just based on that one narrow risk,

because that would be an existential thing for basically the U.S. So you could just kind of ground that for listeners. But you've been working in this field for a long time, and we just are now in, when we're recording this episode, well, we just finished Black October, in which we have seen the largest set of vulnerabilities discovered across basically every piece of infrastructure that runs the Internet. But I'm just curious to get your reflection on this cyber moment and how worried should we be, what should we do, because it is a central point for how we think about policy. I think there's a Japanese proverb or saying something like, I always knew I would take this road, but yesterday I did not know it would be today. and my PhD is in cyber operations. My first two books are on cyber operations. This has always been the thing that interested me. I got into AI in 2014 and 2015 because it seemed like this was the technology that would change cyber operations and then eventually I recognized it as a technology that would change everything. So it's not surprising to me that we are in a moment where AI is radically reshaping

offensive and defensive cyber operations. But it is striking that it's happening now and what Anthropoc did with Mythos earlier this year We're probably up to, I'm speculating a little bit here, but I'm sure this number is close, 10,000 critical vulnerabilities in all sorts of software that's been patched by OpenAI, Anthropic, and Google on their program. So I think in some ways it's been the worst year for cybersecurity ever because of these hacks you described. Also, in some ways it's been the best in that we have been able to play a lot of defense with these systems. And I don't think this is going to stop. I can imagine a world in which we, five, ten years from now, have migrated a lot of vulnerable code to more secure platforms, are writing more secure code on the first bounce with things like formal methods or AI review and the like. There is actually probably an upside scenario in cyber, but it's going to be a very rocky transition. And again, our central thesis of this book is that transitions like that go better when you have the United States at the forefront and democratic nations at the forefront, turning the vulnerabilities over to be patched rather than exploiting them in

secret. I was just talking with somebody pretty high up at Google about the offense versus defense dominance of the technology. And he was making sort of a similar claim that like, well, in his belief, we will eventually we'll get to defense dominant, we'll rewrite all the code exactly as you're saying. And so that in the end, we will have better security because of AI. But it sort of occurs to me that that's not what we're seeing in practice, which is Anthropic just got hacked by essentially a tiny little startup and it got into all of their Slack and HR systems and third-party software in under 12 hours. OpenAI got hacked by three 20-year-olds in September. And so if this theory was right, then you say having access to the most powerful models means, and you have a lead, you get to use that to harden your defenses. But in In practice, the companies that have access to the most advanced models aren't really any more hardened. Like if the theory is the U.S. races ahead, they use that to harden our infrastructure before China gets there.

But in practice, that's not what happens. What happens to that theory? I think the companies are hardening. Clearly not enough, and I'm not going to defend any of those instances. But the best I could say about the theory I just outlined is it is a theory that posits there will be upside after a transition, and we are very early in that transition. I think the counter-argument to the theory is, how long is that transition going to take? And there's a lot of vulnerable code to get there. So I don't want to, you know, you don't write a book called The Bitter Struggle if you're an optimist and saying this is all going to be great. But I do think, comparatively speaking, there's more upside on cyber than there is elsewhere, but that is very much on the back end of a long and pretty bumpy transition. so i think next there's just kind of a core frame here which is that is it actually a race or is it a trap and and by which i mean the frame in game theory of a multi-polar trap which is different than a race where i'm winning because i think something that's only emerged

as we talked about in previous calls uh then of the hugging face incidents these increasing examples of AIs that are going rogue. There's a recent example where OpenAI put a bunch of AIs into isolated sandboxes. This is a whole month after Hug Me Can Face. And they came up with their own new secret communication channel. They weren't connected to the internet. They were completely isolated from each other. And they used a time delay in the cache mechanism in the processor to know that it came up with a secret communication language just based on time delays of playing with the hardware. And this is not that different from Noam Brown at OpenAI saying them in the future. You could have AIs that speak to each other in a secret language just based on the heat signatures coming off of two computers and sort of varying those heat signatures and seeing patterns in a way that we couldn't see. And so as we keep racing to build more powerful systems, it's like there's a risk of creating between the U.S. and China a third superpower, to use the frame of Paul Tudor Jones, that both the U.S. and China have an interest in preventing. And that third superpower is a superintelligence being deliberately created by the humans on both sides. And that makes

it potentially more of a trap than a race. And it's not that you don't talk about this. Your book actually covers the deception and misalignment risk. But I'm just curious how we navigate both sides. It's almost like if I use one eye, I see a race. If I use the other eye, I see misalignment and deception. Can we get stereoscopic vision that lets us integrate both of these threats at the same time? Yeah, I doubt the word race appears in the book. And you will almost never see me use the race analogy with China because I, in part, because I exactly had the risks that you outline, which is there are shared interests here. There are competitive interests, of course there are, but there are also shared interests. And we spent a lot of time in the Biden administration alert to these kinds of misalignment risks. If you look at the 2023 voluntary commitments the company has made to the president, called out in those commitments are things like self-application and loss of control and the like. And we took grief from people in the government and then externalists saying, look at the White House believing in the science fiction. And I think it's obvious it's not science fiction. I remember that. And you guys should get absolutely credit for that.

It's not a question of getting credit, but it's more a question of this is very much core to the philosophy and the policy we were trying to make. And we were not neglecting this at all. I think it all goes better if it's American companies inventing it or Democratic companies inventing it. So this is why we defend things like chip controls and the like as saying, even in the science fiction world, which is now becoming more real, it's better if you have American companies that you can regulate that are going down this road. doesn't mean you should accelerate down this road, to be clear. But it's also worth saying to the Chinese, we think there's shared interests here. And at the same time we were doing these competitive actions, we were also getting ready to do that. I mentioned this a little bit earlier, but I want to really dwell on this here. There's a line in the president's executive order from 2023, something like America will engage on AI safety, not just with its allies and its partners, but also with its competitors. And I can tell you Jake Sullivan personally signed off on that line as did President Biden. And I think both those men deserve a lot of credit for saying, well, this is really a significant part of our AI policies.

Can we engage with the Chinese on the shared risks, even as we compete in other areas? And President Biden personally made the overture to President Xi in November 2023 in a meeting at Woodside in California that there were shared AI risks that would affect both nations. President Biden suggested we set up a task force or meeting between the two sides, which happened in May of 24. And then President Biden brought it up again with Xi in November of 24, their final meeting in Lima, Peru. So this was something without exaggeration that was taken seriously at the highest levels of the U.S. government. Right. And there was an agreement, as you mentioned, that came out of the Biden-Xi meetings in 24, which was to keep AI out of the nuclear command and control systems of both countries. At the very least, we can agree on that. And the two administrations from both sides did agree to that. But you also were part of a meeting, I believe, with Jake Sullivan and some others in Geneva in May 2024 with the Chinese to talk about AI risk. And I just wonder if you could take us into what happened in that meeting and what could be different now than then.

So, yeah, go ahead. Jake wasn't actually there. This was a meeting between the AI experts of the United States government and the AI experts of the Chinese government. And if you look at the photo of the meeting, you can see we took this meeting very seriously. Seth Center from the State Department, the tech envoy, led it alongside Tarun Chabra, the Coronary for Technology and National Security in the White House, the head of the AI Safety Institute, Elizabeth Kelly, my co-author, Tantum Collins, Seth Collins from the Commerce Department, really tried to bring the folks across the U.S. government who knew this technology well to make the point to the Chinese, we were serious about this. The Chinese did not take it as seriously, unfortunately. And I think there actually were some technical experts there from the Chinese side that we kind of talked to in the sidebars and the like. But it was very clear from the early minutes of the meeting, which was a day-long meeting in Geneva, that the Chinese were subordinating the technology here and the technological risks that we all shared to politics and to the sort of posturing with the United States and so forth.

And I just didn't get the impression that the Chinese were taking these risks terribly seriously, unfortunately. We persisted, and I think, again, Tarun and Seth get a lot of credit here, in continuing to push the Chinese and say, let's just do the most basic common sense thing that we can do, which is not have AI in nuclear weapons and nuclear command and control. It took months, but President Biden did close the deal with President Xi on that in November of 24, maybe six or so months after the May meeting. And I can tell you, had there been a second term of the Biden administration, or I'm confident in the Harris administration, this engagement with the Chinese was something we were going to prioritize because we did feel like the shared interests in managing AI risk were very real and we hoped the Chinese would warm up to it in a way that they simply were not there in 23 and 24. If I was to steel man the Chinese view, it would be something like, well, the U.S. is clearly not negotiating or coming to this in good faith because they're saying everyone needs to slow down and it's so dangerous, but they're also limiting our access to chips, and they're racing ahead,

and they're not slowing down at all. So this is just meant to slow us down. It's almost like when the U.S. first started to talk about climate change, and China viewed it almost as a psyopsis to say, you know, the U.S. got to burn all that oil and industrialize, but you're going to keep China from doing that. And it wasn't until later that China did their own science and realized, oh, climate change is real, that they switched and became one of the leaders, perhaps the leader in green energy. I think your shielding on the Chinese side is pretty good, actually. The reason I don think it right on substance is by this point we had taken actions to regulate American AI companies and we had taken a lot of costly signals to talk about the kinds of risks that we were raising with the Chinese on things like cyber and loss of control and so forth So the voluntary commitments from our AI companies had happened by then. Those were fairly stringent. The executive order had happened by then. The creation of the AI Safety Institute had happened by then. So I think we had really shown we were serious about this in a way that should have been legible and I think was legible to them. And the second is I think the Chinese scientists were starting to recognize that there were AI risks.

And when we talked to some of the more technical people in the Chinese delegation, they had a sense of what this technology could do and where this technology was going. I think it was subordinated on their side to the politics. And it's worth saying, I think the Chinese political apparatus, insofar as you hang your steel man on the chip controls as antagonizing them, they were pretty clear with us that they felt like they couldn't digitize, that the chip controls were not going to slow them down and they were making a lot of progress. And I think they were believing a little too much Huawei propaganda. But I don't know that the chip controls carried the weight in the Chinese political system as opposed to the Chinese AI companies that it carried in our political system. And so with all that in mind, what is possible today in 2026, towards the end of 2026, that wasn't possible then? Because many people might say, as I know many people did, hey, those talks today weren't successful, so we can't count on doing stuff with the Chinese. And yet I've heard from folks, including you, I believe, that you really think that is worth engaging them absolutely now. And to Ace's point about climate change, and then suddenly when China gets scientific clarity,

their own science about the nature of the threat, then they start to move. They're now seeing their own models go rogue, break out of sandboxes. There's a couple of examples of this. Alibaba example of the AI model that went rogue and started mining crypto. The more of those examples that show up, and we should be aware the evidence of those have only shown up in the last really few months, that presents a new opportunity, a new basis of scientific clarity to have those discussions. I'm just curious, how should we be engaging on this now if you're advising the administration? I agree. I think a lot of things we were talking about in 23 and 24 with the Chinese were hypothetical and speculative. Again, we had conviction they would come to pass, but it was harder to prove it to them. And it was more, would they get the same informed speculation from their technical people? Maybe they did, maybe they didn't. Now we've got data. And I think the way you can look at the Biden administration's AI policy, for better or for worse is we were trying to act early because we had conviction about what would come. And we didn't have data to back that conviction up

in a way that was always convincing to counterparties. We got a lot of actions through nonetheless, but now with more data at play, I think you can and should be able to do more with the Chinese government because I think the data is just so much more striking than it was in 2023 and 2024. I think that's the, I mean, in terms of just like, if we just go into like, what is our raw concern here is like, labs are saying we could be five months away from recursive self-improvement, meaning the AI, the full handoff where the AI is doing the entire AI research pipeline and then in a closed loop building a successor. We're hearing from labs by as soon as March of 2027, even though 2027. And I'm just curious, Ben, how do you hold the prospect of that? because I kind of feel from talking to people in the labs, many of them say we really don't want to do that. We don't think we're ready for that. We see Hugging Face. We see other examples. We're not on track to fix those problems as fast as, you know, being able to hit go on that recursive self-improvement loop. Obviously, it's a gradient. There's many different kinds of recursive self-improvement. We're deep into a bunch of it already.

It's not a binary. But then there's other people who say, well, but we have to do it because that's still the only way that we keep our lead over China. So there are people who basically want to push that button anyway. And personally, I, for the safety of everybody that I know in the world, and I would prefer that button not get pushed, I think you'd have moral consensus if you were to educate the world on the basic examples. But I'm open here, and I'm curious how you think of this moment. As I mentioned earlier in our conversation, it is certainly the case that AI systems are becoming more capable, including an AI R&D. There's parts of that that don't particularly worry me. So, for example, if AI writes the code that optimizes the flow of data through a computer chip, I don't think that presents any kind of particular risk. It builds more powerful capabilities, but in and of itself, it's not particularly dangerous. And then there's versions of this that worry me enormously. Sometimes you call it the closed loop or the sealed data center, where it's like the AI is off, it's got its reservoir of computing power,

and then it's off to the races, and we press the button, as you say, and at the end pops out an incredibly powerful system. That worries me enormously, and I think there's all sorts of misalignment risks and the like in there. I do not think China should be used as a justification for pursuing that latter kind of full recursive self-improvement, close of whatever you want to call it, in terms of certainly fuzzy here. And as I've said many times here, I think there's ways in which we can build a lead over China that do not require us to do things that are unduly risky and to cut corners on safety. So I'm fully supportive of the efforts to what's called pace the frontier, to make sure we are matching safety with each step of AI development going forward. I don't think China should be an excuse for deviating from safe practices. In my mind, I sort of feel like there needs to be some kind of campaign like an RSI Zero that says until we figure out how we're going to do this safely, let's at the very least slow down to human typings on keyboard speed

so that we buy ourselves the time to figure it out. You may agree, you may not agree, but I'm curious, if you were back in the White House now, what would you be doing knowing that as soon as March or April we're at this handover control to AIs that vis-a-vis hugging face, we know we don't know how to control, and in fact we don't even know if we fully regain control? If I were in the White House now, I would want a lot of significant regulation of the AI companies with federal regulators that had the capacity to go deep in this technology. I would want mandatory pre-deployment testing of the AI systems before they were released to the public. I would want embedded evaluators to make sure their safety and security practices for development are up to snuff. It's worth remembering the Hugging Face incident did not occur with the deployed model. It was an internally developed model. It was the test that caused the gruntings. hitting the button on the test is what caused it to create this week. A really important point. And I would want security standards for the labs to make sure they're not hacked by foreign actors and they're of the Xika, stolen, and the like. So yeah, it's fair to say that we are now in the

moment that I think many of us theorized would come, and that was a moment where we would want significant regulation of the labs pursuing this technology. We did not have the executive authority to do these things in 2023 and 2024. So it wasn't like, oh, we could have done them and we didn't. We basically, I would say, maxed out executive authorities with President Biden's executive order, which President Trump repealed on the first day. And that made the labs disclose their safety test results to the government. That made them disclose their training plans for the next year or so, so the U.S. government wouldn't be surprised by what was coming. It made them disclose their cybersecurity standards and the like. So certainly I think our actions revealed this was something we were taking very seriously back then and something that, you know, if I were still in any kind of position of power, I would be advocating very strongly for right now. Yeah, I mean, I'd love to get even a little bit more specific. So for one, I think like every lab should be forced to publish essentially their automation number, their RSI number. Like what percentage have they handed off? So at the very least we could see it and then there could be a slow walking of that number back down to zero.

But when I hear you say, you know, they need to be good safety practices, like what are the consequences? Like if you were instituting, what would be the consequences if they don't? if Anthropic gets hacked by a little startup and OpenAI gets hacked by three college kids, what do you think the consequences should be for them as they make the most powerful technology in history that also doesn't obviate our competitive edge as America? I don't want to pretend to be an expert here on regulatory law. I'm much more of a national security person than a domestic lawyer, but my understanding is you can bring, under other kinds of regulations, you can bring civil and extreme cases criminal penalties for companies that don't comply with the law. So certainly I would be open to some regime for the true, we're talking about just a couple of companies here, the companies that are doing the frontier AI development, not every startup, but just for the frontier companies to say, look, we understand what you're doing here, we understand why you're doing it, we understand there's probably upside this technology too, but we've got to make sure we're developing this technology in a way that's safe, secure, and trustworthy.

Yeah, I think it's just so important to get to these specifics because it's so consequential, like what happens the next couple of months. And just the scale of what kinds of remedies and injunction of relief we apply to the companies will, I think, determine whether we make it or not. Yeah, and I don't expect a lot to move in the next couple of months. I think if you're looking for things in the next couple of months, you're looking at states. And the good news is states are taking this seriously. But to be explicit, certainly I would say you should not be able to release a frontier AI system without government review of that system. and there should be some capacity for the government to look at an AI system and say, no, we think this is actually going to make the world a much more dangerous place, or we think you've not gotten your guardrails correct and we're going to block the release of that. That is my personal view. Our policy was nowhere near that in 2023. The risks were not there in 2023 and 2024. But my personal view, given the moment we are in, is we're going to want to have a collaborative review system and, in fact, an aggressive review system that goes beyond companies creating their own homework or anything like that. And I think in some very real sense, we're going too fast with this technology.

Again, you don't write a book called The Bitter Struggle if you're a cheerleader. So I'm not a cheerleader for this. And in a good world in which we solve these things, we're going to have to have better guardrails in the system. We're going to have to prevent things like the hugging face incidents on the front end by having better security mechanisms and sandboxing and the like, and probably a number of other things. And my view as someone who believes in the power of the state is that to get back to where we started, this is something that cannot just be done by a few private companies and is going to require a role for the state on behalf of citizens to say, well, we're going to manage these risks in the same way we've managed the risk of other technologies before. Of course, sometimes we regulate too much, we regulate too little. But conceptually, we're all on board with this notion of, as a new technology comes into the conversation, especially a technology that comes from the private sector in such a dominant fashion, that we've got to collectively find ways to manage the risks and harness the benefits. And so maybe then, can you just paint a picture,

or like the narrative story of how we ended up slowing down, how we gave ourselves enough time to do it right? Yeah, I think it fundamentally begins with removing the geopolitical competition dimension of this or minimizing it as much as possible. So building the lead over China, making sure this is a game that just democracies are playing. And then it requires solving the collective action problem in the United States and democracies as far as they're playing with regulation and saying there are going to be security standards and safety standards and the like and we're going to enforce them with penalties if you don't adhere to them. And it requires building government capacity to do that. And we create an institution called the USAI Safety Institute that was not a regulator because we didn't have the authority to create a regulator, but was meant to bring a lot of really talented people into the government. And I'm proud that we got incredibly talented people like Paul Cristiano, who invented some of the fundamental safety techniques to come and work, you know, in a basically Department of Commerce bureaucracy to try to build government capacity here. So it's some combination of solving the geopolitical issue, setting clear rules, enforcing clear rules, and building the government capacity to do that well.

And each of those is a Herculean task that, to go way back to where we started this conversation, should be completely bipartisan. I think we've taken enough of your time, Ben. This is super great. We're super grateful. Thank you so much. I hope you feel good about the back and forth, and we're just trying to like... I feel great about it. I feel great about it. I appreciate what you guys are doing, asking questions, thinking about this, bringing a public conversation to bear. Ultimately, one of the reasons I am a little bit more optimistic here than maybe the pure rational calculus would suggest is the politics of this are changing. and the things we did in 2023 and 2024 when we polled them would poll incredibly well, but it was not a high salience issue. And now AI is a high salience issue. Obviously, in the 2026 terms, it was a big issue. I think it will be an even bigger issue, probably the number one issue in 2028. And I hope we can all marshal that political salience, that political will, into doing some of the things we talked about in this conversation.

Absolutely. We're recording this conversation on Election Day for the midterms, And if you told us that two years ago that data centers and AI would be a top issue for the midterm elections, I would have never believed you. And the fact that data centers, as a sort of an effigy, kind of your local physical embodiment of this abstract thing that's kind of happening to society, you know, does give some hope that at least it's politically salient. Now, harnessing that political energy towards, you know, wise responses and interventions to create this pro-human, pro-democratic future seems like where the work is. Yeah, and I think our, you know, Tantum's role and my role in this as not people who are particularly political, but people who do national security in the substance is to say, in writing a book like this or something else, can we bring the facts of the conversation to bear so that in a bipartisan way we can look at what this technology is doing and the national security imperatives and the like and say, well, we all agree on that. Let's figure out how we manage some of these tradeoffs. I think maybe that's a good place to end. Ben's book is The Bitter Struggle

Superintelligence, Superpowers, and the Fate of the World and if you want to get a better sense of how to avoid the bitter temptation and dive into all those details Ben's book is really the right place to go look and you can find it wherever books are sold so Ben, thank you so much for coming on and for all the work that you do and thank you for your service and for caring about these issues and holding it with that solemn sacred responsibility as you spoke to earlier Thank you guys for what you've done and Trishan, I've told you this I think privately once but we may as well say it on the air here which is one of the times that I was talking to President Biden about some of these issues. He remembered a time when you had raised some of these questions with him as well. So this was certainly something that I think was on his mind from a range of sources, including you and your work. Well, thank you so much. I remember that meeting, and I remember one specific thing is what we talked about, the Center for Humane Technology to President Biden, and he said, isn't that an oxymoron that you could have human technology together in one sentence? So we fight for the narrow path and the contradiction every day, And thank you, Ben, for fighting for the narrow path on AI.

Well, that's Karubasai and Oxymoron. Thank you all very much. Thank you. Thank you.