AI policy and the battle for computing power

20:48Ben Buchanan:What is you as you're looking at how this relationship is, what would you like to see in the developing relationship between government and the private sector as it's meandered across these administrations? And to your point in the last in that last answer, what what would be an ideal to go for between public and private sector? there's always that tension about, you know, we've seen in the news as we're recording this, some of the concerns between the Department of Defense or Department of War, depending on how you're choosing to label it, and, you know, a particular organization that doesn't want to put certain models into combat scenarios.

21:37Ben Buchanan:And so, which is anthropic, I guess there's no reason not to name them outright. How do you see that relationship developing over time? What would be a healthy way for it to develop that brings the larger good into picture? And are there anything that jumps out as we really don't want to do that? I know you just named selling maybe to potentially hostile countries and such, but I'm just curious how you see that relationship developing.

22:06Chris Benson:Well, I think it depends on if we're talking about the relationship between the government and AI companies or the government and chip companies. So when it comes to the chip companies, companies like NVIDIA and the like, our posture was we thought their technology was so important, so fundamental, and also so scarce that we did not want it going to countries like China because of the ways in which it would modernize the Chinese military and the like. And because China really struggles in part because of the controls on chip manufacturing equipment, China really struggles to produce similarly powerful technology.

22:41Chris Benson:China will not make a chip as powerful as the one Trump has agreed to sell them until 2028, according to China's own roadmaps. If you believe the Chinese propaganda, they won't get there until 2028. So there just is, as I said before, an extraordinary advantage. And we put a policy that I would defend of export controls and say we want this technology, especially given how scarce it is, every chip that gets made will get sold. We want this technology to go to democracies and ideally American companies. So that I think is the first aspect of the relationship. Then the question is, well, what's the relationship between the government and the AI companies themselves, the ones that are developing the systems?

23:22Chris Benson:I'll leave the news reports aside. Obviously, it's a timely subject, but I'll tell you, we thought about this a lot in our administration. And the president signed a document called the National Security Memorandum. Basically, it's an executive order for the Department of Defense and the intelligence community. It's got an unclassified and classified component to it. And then classified component is pretty straightforward, which is it directs these government departments and agencies to work with the private sector and to say, let's bring this technology in to the private sector. A lot of the ways in which we did things in the past is outmoded or broken.

23:54And we need something that is newer and more capable

23:58Chris Benson:and able to keep up with changing times and changing threats. and I'm very proud, the president signed that in 2024 and I'm very proud of what we did in the time developing that and then before we left office to start that ball rolling to say there's a way we can work very collaboratively. That does not mean it's no holds barred and President Biden said this is a technology that poses significant risks and we need to have guardrails in place to make sure this technology is not misused and the like and we were very alert to try to craft those guardrails in a way that gave the department the flexibility they needed to fight and win wars, but also made us worthy of the values we are defending.

24:40Chris Benson:And that is something that the Department of Defense and the intelligence community worked with us and was totally on board with, of saying we're going to use this technology in a way that's consistent with our values.

24:51Ben Buchanan:As we talk about guardrails and values, and that's a huge topic right there, especially when you talk about specific applications of AI. The reason I brought up the anthropic thing that was the current affair thing before is it is typical of a concern that people across the political spectrum have in terms of the appropriateness of putting AI into specific cases such as combat and stuff like that. What is your guidance? Like what is your own personal feelings around the right place to draw a balance in that? You know, what's the responsible place to land in terms of how you match AI up with security critical or safety critical, you know, applications of it?

25:43Ben Buchanan:Do you have any guidance on where you think things should land?

25:47Chris Benson:I don't claim a lot of expertise on the particular service. I think I – and again, not commenting on the Anthropoc case in particular. I do think in our administration, we recognize that it varied by use case. So there are some use cases, I don't know if you call cyber operations combat, where autonomy is really important. Yeah, I do. I mean, I would too, but I can introduce you to some infantrymen who wouldn't, and men and women who wouldn't. Missile defense is another area where autonomy has historically been very important. And there probably are some great Lockheed systems that have an autonomous mode from the 1990s on this kind of stuff.

26:24Chris Benson:So I think it depends on the area. The DOD in our time, and I had nothing to do with this, but they revised a policy called 3000.09, which talked about appropriate levels of human judgment in this. And I think that policy is -

26:40Ben Buchanan:Very familiar with that personally.

26:41Chris Benson:You probably know more about it than I do, but that was something that was important to us. One thing that I was a little closer to that I think was important is to say, this is not just the United States deciding this on its own. We need to get a group of nations, ideally a group of democracies, but also broader than that, to work on this problem together. And the answer probably is not, we will not have any autonomy in military systems. No one is saying that. But we developed a document called the Political Declaration on the Use of Autonomy in Military Systems, I think. And we got something like 58 countries to agree to that and the set of principles that would guide that work.

27:15Chris Benson:So I think that is vitally important as well. Wherever we decide as a nation to draw the lines, it's vitally important that we go and try to set that as the, in a collaborative way, the norms and standards with the rest of the world and admin on the rest of the democratic world. And that's something I think we're very proud of, uh, having gotten the ball rolling on.

27:35Ben Buchanan:I think is kind of one more question along the line of, of guardrails in general and kind of where to land. Um, there's a lot of debate, uh, in the, in, you know, out there in general with people that care about policy and how AI relates in terms of what I'll call kind of speed versus caution in terms of how you set guardrails that are appropriate and stuff. And interestingly, I'll make the comment so that I don't put it in your mouth. I've seen the current administration go both directions on that at different times, which is somewhat confusing in terms of how to interpret. But in general, as you're looking at the private sector racing forward in terms of development of models and capabilities, as you would expect from the technology sector.

28:24Ben Buchanan:And then you're dealing with certain areas that are safety critical and you're worried about, and it's not necessarily just military topics, but other areas, there's obviously transportation and there's even psychological impact. A hot topic these days. I have a daughter who's about to go into high school and the impact of AI playing into systems that are social media and things like that. So there's a wide array of concerns to address this. How can people think about that kind of speed versus caution paradigm in all of these different aspects that impact their lives and or which they are watching politicians in Washington, D.C.

29:12Ben Buchanan:engaged in. Do you have any thoughts around how do you, as you're looking at such this massive array of possibilities, how do you assess those? Do you have a framework in your own mind about how you would say, this is a good place to go, this is not such a good place to go?

29:30Chris Benson:I mentioned at the top of our conversation that in some respects, this is the first private sector invented revolutionary technology in the last 100 years. I think the last one was the railroad, if you go back to the late 1800s. And the railroad had many of the characteristics that AI has. Private sector invented, huge capital expenditures. If you go back and read some of the literature, this kind of promise that it would transform the economy. They even said it would transform the climate. It would make everything wonderful. It was just similar kind of utopian vibes that you get from sometimes these Silicon Valley companies.

30:05Chris Benson:And the early days of the railroad were incredibly bloody. And there was train derailments, train deaths. There was no standardization of anything. So there was no time zones. There was no brakes, no air brakes on a lot of the trains. There was not even standardized track gauge width. So you got all sorts of derailments. There was poor coupling between cars, so the trains were coming apart. Thousands of deaths in the early years of the railroad. And it eventually was the case in a very halting, imperfect process. some combination of the government and private sector worked this out. We got time zones from the railroad companies.

30:42Chris Benson:We got standardized track with gauges from the government. Air brakes, coupling between cars, Railway Safety Act, all of this stuff kind of emerges over several decades. The net result of this is the trains are safer, but also the trains go faster. And the reason I kind of give this historical analogy is to suggest that too often, I think, speed and safety are put into tension. And again, look at JD Vance's comments. We're like, I'm not here to talk about AI safety. I'm here to talk about AI opportunity. Well, my view is that we get AI opportunity through AI safety. And through not incredibly cumbersome regulations and the like, but through developing technology that is safe, secure, and trustworthy and people can trust.

31:25Chris Benson:And that is still my general principle. And it shows up in all kinds of different ways. We've worked on domestic things and kids' online safety. And the principle, I think, applies across the board. But at a big picture, that is the most important principle, I think. The second important principle that I thought about a lot is competition is good between companies. Competition is sometimes even good between nations. But you can have a competition that becomes a race to the bottom. And I was worried, and I think a lot of people were worried, that if you had the United States and China in essentially an arms race, or perceived to be an an arms race for this technology, that would create incentives for both sides to cut corners on safety and to race ahead in ways that might be foolhardy.

32:09Chris Benson:So part of our thinking was, we want to build as large of a democratic lead as possible for American companies to essentially make this a democratic problem. And to then say, well, we'll coordinate, we'll figure out whatever coordination and regulation is necessary within our own borders. But this is not a thing where you have democracies and autocracies racing to integrate into the military. So developing a big lead so that you can spend it on safety, I think, and more generally spend it on safety and trust, I should say. That was a big part of our philosophy. And again, I think that shows up in the policies.

32:42Ben Buchanan:So, Ben, as you were just – as we went into break kind of talking about how different geopolitical concerns affect how people are perceiving potential like arms race to AI or whatever. And that begs a larger question, especially in today's climate, is kind of the international order has been definitely being, I'll gently say, transformed in recent times. And so many of the relationships that for many, many decades, 80-ish years that we've been relying on to coordinate are shifting. as we are, like, even if that hadn't happened, you're having AI technologies and chip concerns that aren't just affecting, you know, we've had a somewhat of a US centric conversation so far, but it has, you have to get allies working on different policies and try to align those to where we have that kind of guideline safety net that you were talking about just before the break.

33:51Ben Buchanan:And so is it getting harder to do that in today's geopolitical? If you were able to influence current policymakers in this administration, how would you guide them into a safer world with AI where we're able to get many parties internationally to kind of agree on what those safety and guidelines should be? What is the best way to do it? It seems like it's harder with some of those relationships faltering at this point.

34:21Chris Benson:I think that's absolutely correct. And we knew the international dimension in this was really important. And this was something that our companies said to us all along. So this is not a bleeding heart liberal thing. This is American companies saying, this goes better for us if there's one clear set of standards and if there's interoperability between regulatory structures and the like. So this was something that was a priority for us in every aspect. You can see it in the president's executive order on AI. You can see it in the National Screen Memorandum I mentioned. We had something called the Hiroshima process that the G7 Group of Nations did.

34:56Chris Benson:We had the UN resolution that was unanimously passed, including the Chinese co-sponsored the resolution. So obviously each of these documents was different. But our view was we had to show up as a country in every international setting to make the case for what our vision of AI looked like and then to hear from others and to work it out. And that was international diplomacy. And there's a great team at the State Department that deserves a lot of credit for the work that they did. All of that is harder now because the relationships have frayed for other reasons, tariffs, politics, and like. I think ultimately that will hurt us and that will hurt us and will hurt our businesses.

35:31Chris Benson:And the proof is going to be in the pudding on that. So yeah, I think the harm is real. One thing I also want to say is this is not just going to be a question of democracies. And I believe very firmly that democracies need to have preeminence in AI. I recall a quote from Kennedy. He's given a speech at Rice University in 1962. Everyone remembers his speech because he says that's when we're going to go to the moon and come back. But he talks later on in the speech about space. And he says, we don't know if space is going to be good or ill for humanity, but only if we are first can we help decide.

36:11Chris Benson:And President Biden had a copy of the speech in his private study outside the Oval Office and I always thought it was just a great metaphor for this. So I'm all for democratic preeminence in AI but I also think we need to talk to our autocracies. And we had quiet conversations with the Chinese in Geneva. I mentioned the UN resolution and I do think there's an aspect of this that is being able to engage with all nations about a technology that affects all of us. And I'm worried we're also losing the ability to do that just as we're losing the ability to talk to our friends about it.

36:44Ben Buchanan:As we've kind of talked about where the West is and our capabilities a few minutes ago, talking about chips specifically, versus Chinese capabilities and where they're anticipating chips coming up, one of the things that we've seen, particularly over the past year or so, is kind of a surge of scientific papers released by Chinese researchers. If you go to Hugging Face, there are a lot of Chinese open models that are there at this point. They are, if you're just doing a numbers count, they are definitely catching up and kind of passing it that way of measurement. Like as we're at this point of at least inflection at some level, and I'm not sure that I know what the inflection is, that it's a fairly complicated thing to discuss and its intricacies, but what might this mean for the relationships between ourselves and China in this case in terms of maybe steering away from an antagonistic relationship in the future, finding our way back to that kind of Nixon opening up China moment potentially in the future, whether it could be this administration or whether it needs to be a future administration that does that.

37:59Ben Buchanan:How does this surge of open models matter in the sense of should people be using them? Should they stay away from them? Like, how should they think about them? We get a surprise, I get a surprising number of questions about that topic is like, do I want to use a Chinese model? Am I putting myself at risk? But so there's the, there's kind of the several questions you might say packed into one, there's kind of the, how do individuals assess those? And how does it affect the relationships of these national bodies in terms of how they are developing the relationship in the future? I definitely am hoping that we don't get into a worse situation in the Taiwanese Strait.

38:41Ben Buchanan:It'd be nice to find an opportunity to back away from some of the risks developing there on an ongoing basis?

38:51Chris Benson:Sure, well, we can start there. I obviously would like to, no one wants a warrant in Taiwan. I mean, I think it's a very important piece of the global puzzle, but I'm with you on that. Now that I've left the government, I advise American AI and cybersecurity companies. So obviously I have a preference for American systems and I think some part of me still has the old national security policymaker in me too that of course has a preference for American systems. But I think the Chinese developers are very talented. And if you look at the team at DeepSeek and look at what they have done. They are very talented algorithmic engineers.

39:24Chris Benson:I was a big fan of, we worked very hard in the AI executive order on high-skill immigration and bringing AI scientists to the United States. I'd love to have all of them move to the United States and start their companies here. So I have no ill will towards them or the Chinese people generally. That said, I do think sometimes DeepSeek or companies like that can be presented as rebutting the thesis I outlined earlier of the centrality of computing power or the dominance of the United States and democracies in computing power. I don't think it does that. And in fact, if you look at the DeepSeq systems, they're all trained on either smuggled or stockpiled American chips.

39:58Chris Benson:And they're constrained in their performance. They, in many respects, lag U.S. companies because of their inability to get U.S. chips. And if you look at the DeepSeq v3.2 paper that came out in December of 25, for example, they acknowledge this in the paper. They say, essentially, we're constrained by a lack of computing power. And the DeepSeek CEO, when I was in office in 2024, said, my issue is not talent and it's not money, it's computing power. And he's right. He's absolutely right. So I think several things are true here where DeepSeek people can be very talented, which they are, and also it can be the case that computing power remains incredibly important.

40:34Chris Benson:In fact, probably the most important US advantage. So none of this changes my mind on the policy side of it and what we did or didn't do. Again, I think we should go further than we did, strongly further than President Trump has gone. But it is a reminder, I think, of the Chinese ability in this space. And it makes me wonder what they would do if they could get even more of our chips. And the answer is they would be really good.

40:59Ben Buchanan:I'm curious, as they continue to work toward developing this, and you've talked about your expertise in cyber, and maybe you're kind of starting your career on the cyber side, moving into AI. Can you talk a little bit about how you expect AI to create asymmetries in cyber and the various aspects of cyber? Talk a little bit about what the junction of AI and cyber looks like and how you expect that to evolve. How is it influenced? We hear about cyber attacks pretty much every day. Everyone does in the news, and I don't think anyone expects that to stop. So going back to our agreement that it is certainly part of warfighting, it's also part of daily life that every listener here has to consider.

41:46Ben Buchanan:So could you give us a little bit of a level set on how cyber and AI interact in that way?

41:51Chris Benson:I think cyber operations is one of the most immediate ways in which AI will impact national security. And again, it's part of how I got into AI because of this connection with cyber. And I agree, cyber is part of warfighting for sure. But as you said in your question, it is also part of day-to-day life, not just for ordinary people, but for nations. The famed U.S. strategist George Kennan has this quote. I think he talks about the perpetual rhythm of struggle in and out of war between nations. No place is that more real than cyber operations. The United States conducts offensive cyber operations, defensive cyber operations, not just in wartime.

42:26Chris Benson:It's a key part of intelligence collection. Every nation does this. So an advantage in cyber operations and offense and defense translates in a very immediate way to an advantage in national security, which is why it's so important. Now, I think AI will have a significant effect both on offense and defense and cyber operations. There's a number of different angles to this. Let's take the most obvious one, which is vulnerability discovery. Vulnerabilities are weaknesses in computer code. If you can find them on the defensive side of the ball, you can patch them and reduce them. If you find them on the offensive side of the ball, you can exploit them and use them as a tool for intelligence advantage and the like.

43:01Chris Benson:The question for decades has been, can AI find software vulnerabilities? DARPA ran a thing called the Cyber Grand Challenge in 2016, tried to begin answering this question. We did another version of it called the AI Cyber Challenge in 2024 and 2025 that went much deeper. And now I think the world is providing the answer. Anthropik, which is a company I advise, but I do not speak for,

43:27Chris Benson:published with their recent model release, published that I had found something like 500 high severity vulnerabilities in open source software. So some of these have been in the code, I think, for years or decades. So this is a tangible proof, not in a theoretical kind of way, but in a very practical real world way, that AI is changing cyber operations. And I do think there's a lot of opportunities for nations to use it to their advantage. And I hope the United States is continuing the effort we put in place to try to get there first.

43:58Ben Buchanan:I appreciate that. A great way of framing that. As we wind up and kind of as a last question, we often ask kind of what we call the future question to finish up and like where you're thinking is that. And I think I want to kind of craft it a little bit on this in terms of let's take the current politics and the challenges of that out of it a little bit and just talk about democracies in general within AI. If we're looking at what the next few years might look like, and let's say that some of the angels on our shoulders prevail, and maybe some of the scary things that we worry about don't occur, so that we kind of get a hold of.

44:48Ben Buchanan:maybe rational minds are take hold for the future. How do you measure democracies? Um, kind of like, are we winning as an, as a simple way of putting it to democracy and AI? We've talked a little bit about, you know, I talked quite a lot about that in the conversation. How do you look forward and say, yes, democracy is holding true. We have a competitive advantage in this particular way or that particular way. And we can measure it that, um, AI development is better for that. I know that's a little bit of an oblique way of putting it, but I'm trying not to put you into a Republican versus Democrat future kind of thing and more just democracy going forward.

45:29Ben Buchanan:How does that look to you when you're thinking about healthy democracies moving the ball forward?

45:34Chris Benson:I appreciate it. Again, I don't think it's a Republican or Democrat thing. I think this question of what does success for democracy in the age of AI look like comes down to three things. The first is, are we inventing the technology? And are we the ones for the species pushing this technology forward, bending it as best we can towards safety and justice and the like, but making this technology ourselves? And that's part of the reason why things like chip controls are so important. The second is, are we adopting this technology? Are we adopting it to a national security apparatus? Are we adopting it to things like cyber operations?

46:06Chris Benson:Are we adopting it in our economies, in our businesses, in a way that propels our economy and our prosperity? and of course propels our security, and advances our security as well. So that's the second question. Those two questions I actually feel pretty good about. I feel best about the invention one. I think we have all the winning cards on invention. The only way America loses is if it folds and does things like sell chips to China and the like. But I think we have the winning cards on invention. Adoption is doable. It won't be easy, but it is doable. Savvy policymaking can do that. And then the third set of questions is the harvest, which is, are we using this technology in accordance with our values?

46:42Chris Benson:We talked a little bit about lethal autonomous weapons. That's one dimension to it. Another dimension to it is, are we using this technology domestically in a way that guards against some of the risks that it could pose? What are we going to do about this technology and its impact on jobs? I'm a national security person, not a labor economist, but that's really important, I think, in a democracy. What about the ways in which it can centralize power in the hands of a couple companies or in the hands of a government, a surveillance state? What about the degree to which it might undermine the social contract and the ability to participate in democracy?

47:20Chris Benson:You mentioned disinformation. You mentioned kids' online safety. there's many many many different aspects of the how do we make sure ai is advancing rather than undermining democratic values question i don't have great answers to all of them i think i've got answers to some but those are the three things i would use if we're sitting here in five years or ten years and we're evaluating whoever's been in office it's is america inventing ai or our democracy is inventing ai systems are they applying them better and are we doing so it's consistent

47:46Ben Buchanan:with our values all right that was fantastic thank you very much my pleasure canon really Really appreciate having you on the show. A lot of great insights there. Thanks for sharing them with our audience. And look forward to having you back on the show at some point in the future to talk about some of the extension of what you've been working on at that time. You're always invited back. Looking forward to talking to you again.

48:08Chris Benson:Thank you so much.

48:16Ben Buchanan:All right. That's our show for this week. If you haven't checked out our website, head to practicalai.fm and be sure to connect with us on LinkedIn, X, or Blue Sky. You'll see us posting insights related to the latest AI developments, and we would love for you to join the conversation. Thanks to our partner, Prediction Guard, for providing operational support for the show. Check them out at predictionguard.com. Also, thanks to Breakmaster Cylinder for the beats and to you for listening. That's all for now. But you'll hear from us again next week.

48:51Ad,'AAAAA