B (2:45)

Oh man, this is great. Promo for wartok. So. So starting at the beginning, at least for me, you know, people have started to talk in the past decade about kind of where artificial intelligence and nuclear weapons mix. It's by no means a new issue, right? I mean, we've, we can talk about Soviet dead hand system or perimeter as it's referred to more currently there. We can talk about different Hollywood takes on AI using nuclear weapons, you know, Terminator 2 Skynet, Linda Hamilton grabbing the fence while Los Angeles detonates around her, and War Game with Matthew Broderick. So there's, there's actually quite a bit of literature out there as well as some policy relevant occurrences throughout history. And for us and for Chris and I really, we were thinking about this in our former roles in the last administration. I think in general, people who work on nuclear weapons issues are kind of like, hey, we have a lot of other problems like why, why do we need to talk about artificial intelligence within our nuclear policy for the first time? Some of those problems are practical, right? The US is thinking about how many more nuclear weapons it may need. It's dealing with these big ticket nuclear weapons modernization programs that are getting delayed or costing more money. They're worried about geopolitical factors that are kind of perceived to be tightly related to the number or types of nuclear weapons adversaries have. China wants to acquire more territory, Russia wants to coerce a NATO state or a partner in Europe, etc. And those are putting stresses on US security guarantees that date back decades. And we're always tied to nuclear weapons issues. So when you throw AI into the mix, which is kind of unclear to most nuclear policy people, why it's a game changer, how is it going to be applied? What does it really change? Does it make nuclear weapons thinkers think more about offensive advantages or defensive advantages, et cetera? I just think it adds too much. It's like, adds another dimension to the nuclear policy debate, which is why you haven't seen it represented in official documents that much. Now fast forward to the Biden administration, the Nuclear 2022 Nuclear Posture Review, which is probably the first official government strategy document that really goes into some detail. And maybe Chris, I think was more involved at the time in that and he can expand on it. The people who are drafting the review and of course the leadership that approved it wanted to make sure there was some language associated with artificial intelligence in as it relates to nuclear policy. And at this point, you know, the sort of think tank and academic debate circles have really started to talk about AI for the past few years. So in 2022, there was a sentence that was put into the Nuclear Posture Review, which was very much in a paragraph that was focused on the risks of unintended nuclear escalation. So what if a nuclear weapon gets used by accident? What kind of controls are in place? This is where artificial intelligence enters the scene as a matter of government nuclear policymaking. The sentence reads, in all cases, the United States will maintain a human in the loop for all actions critical to informing and executing decisions by the President to initiate and terminate nuclear weapons employment. So there you kind of have a staple for US policy, official government policy, which at least among the five formal nuclear weapons states was a first. I think we saw later that year the United Kingdom and France adopt versions of this commitment as well. And then of course, as you referenced, the United States worked for a couple of years to have a similar statement made by the People's Republic of China, culminating in 2024, where there was this Biden and Xi joint statement related to keeping a human in the loop for nuclear weapons use. A much simpler statement, less expansive, but you know, in the annals of US and China arms control diplomacy, I think you kind of call it a win when you can get the same sentence on two readouts of a meeting. I wouldn't go so far as to call it an agreement, of course, but at least we see that both countries share the same intent here and now. I think a lot of the conversation that I've witnessed outside of government is how do you make that statement or that those shared statements into something real? What do you need to do to actually have some understanding that that commitment is going to be lived up to by, you know, either country that's made it. And of course then you really get into hard stuff, which is like understanding how AI is being integrated in each country's militaries, which is obviously a well kept secret. I don't know, Chris, where did I go wrong? What did I leave out?

C (7:14)

No, that's all right. I think just to give a little backstory on kind of how this filled in. So the National Security Commission on AI, which was led by Eric Schmidt, published its final report in 2021. One of its recommendations was to have kind of restrictions on AI for use in nuclear employment decision making. And it's actually very, those series of words are quite important because I think it kind of gets garbled sometimes people say no AI in NC3, and that's profoundly wrong. Right? Because actually AI has to be throughout our NC3 complex. So all sorts of, it's going to be hugely beneficial to our early warning systems, detection capabilities, et cetera, et cetera. So I don't, I mean, there Might be some people who are arguing against that, but not, I don't think, think that would be wise. And I think most people aren't. It's really in the employment decision making and it's this fundamental idea that, that the, you know, the, the person who presses the button is the, or makes the decision for the purchase and to press the button is the President. And it has to, to maintain that it has to stay that way. And that I think is a little bit of a unique thing where it's something that we really, you know, there's, there's pretty broad agreement that's something we don't want to outsource. There's not a lot of other things I think in the tech stack like that. So the National Security Commission on AI recommended this in 2021. It was, it was pushed into the Nuclear Pass review. I'll say, you know, some inside, the inside baseball backstory there. I remember when I was in the White House in kind of mid-2021, when I, when I got first got there, I talked to the people who were spinning up the Nuclear Posture Review and I said, hey, we should do this thing that says that we're not going to use AI for nuclear weapons decision making. And I remember a bunch of people at DOD were just like, okay, that's weird. Sure, why not? Like, doesn't seem like why would anyone do that? That's a strange request from the White House.

B (9:10)

We've got word count anyways, Chris. We've got all the we want.

C (9:13)

So it kind of just slipped in because it was like, sure, people are requesting this. Obviously it's not something we're going to do fine. We have like bigger fish to fry. And it's just interesting because that shows how quickly this debate has moved. Right? Like now this is something where if you said that people would be like, oh yes, this is something that actually we think about every day and maybe you would be on one side or the other, but it would never be seen as a weird statement or request. And I think that it wasn't that long ago. But, you know, I'm thankful that, that we have that statement. Obviously we built into it to also get commitments from the Chinese there, which is, which is rare that they are willing to say anything on nuclear policy. But yeah, it just shows like, how quickly this has really changed over the last five years, that this kind of like, you know, very high level risk was not something that was like seriously thought about in a lot of policy circles. But obviously now, now is, is pretty significantly

A (10:08)

like I don't know how much better this makes me feel that like, there is just a human being who has, like, you know, white blood cells as opposed to a computer who's going to be making this final decision. I mean, so Chris assigned me Command and Control, Nuclear Weapons, the Damascus Accident and the Illusion of Safety by Eric Schlossler, the same guy who wrote the Big Mac book Fast Food Nation, which I had low expectations and it dramatically exceeded them. Like, really well reported kind of Cold War, Cold War stories. And one of the things that really struck me was this command and control problem, right? Where like, okay, say the Soviet union nukes Washington, D.C. and all of a sudden the Pentagon doesn't exist. President's dead, vice president's dead. You know, you go down the list of successions, we're down to, like, person, you know, 25, who, by the way, is like, probably, you know, doesn't have a cell phone on them because it's like 1954. And then you have this question of, like, which human being, like, how far do you delegate it down to, like, Is it a 50 year old in Nebraska? Is it a 35 year old in, you know, West Germany or Italy or Turkey? And my sort of takeaway from that book was like, once you get to that point where either you, like, have the nukes flying and you have really stressed presidents who have like five minutes to make the call of, like, which, you know, psyop A or Psyop B to do, or, you know, we're down to like, the kernel somewhere. We're already in like, a pretty terrible position. So if we're in that moment and it's a, I like, I mean, we kind of seem pretty fucked anyways. I mean, I guess maybe like, the best case for this is that, you know, if AI has the sort of the risk of like, something going awry, like maybe is a little lower in the peacetime or like in a heightened, you know, warning phase as opposed to, like fully midway through nuclear holocaust. Anyways, I'll stop here. Thoughts for now?

B (12:24)

Yeah, look, I agree nuclear holocaust is bad. So whatever we can do to stop that from happening is great. I do think that, you know, Schloss book is great. He highlights a lot of historical examples that continue to animate these discussions even today about the risk of inadvertent nuclear war. Now you throw AI into the mix as well, and it makes it all that much more scary, in part because I, as somebody who works in nuclear policy, like, can't hang with Chris, as somebody who works more on the emerging technology and artificial intelligence side of things in a conversation about what I can and can't do for my area of work. And that's largely true of many of the people who are now focused on AI and in the nuclear policy and non proliferation community. But one thing we do know is that since some of those events that were highlighted in the in command and control, the US has actually changed the way it tries to mitigate those types of accidents. For example, you know, using different types of warhead designs or explosive designs to make sure that warheads don't accidentally explode. I think he cites example about the Titan 2 ICBM exploding in a silo and throwing a warhead. We try to make sure that kind of thing can happen anymore. We don't have liquid fueled ICBM, we don't have warheads with sensitive high explosives, etc. As much as we can. There's also, there's been much more of an emphasis in the past few decades on positive controls and negative controls. Right. You want, you never want a nuclear weapon to go off when it's not authorized to go off. You always want it to work when you do want it to work. And that has led to a lot of technological and design elements in nuclear weapons that I think all nuclear weapons states try to employ now. And I do think that that sort of safety culture has really increased the reliability that our system works. And we're not going to have the types of false alarms and accidents that Eric worried about. I'm not saying it's impossible, but it does make it a lot harder than

C (14:18)

it used to be.

B (14:20)

AI introduces potentially some new failure modes.

A (14:22)

Right.

B (14:22)

I think getting. And some of the recommendations I know like the Future of Life Institute and other organizations that have really been pushing sort of how do we manage AI risks that are going to be introduced in nuclear policy decision making. Have Talked about how AI is integrated into NC3. Is there going to be transparency? Is there going to be reliability to find NC3, sorry, nuclear command and control. So these are the, this is the sort of suite of systems that forms an architecture to enable nuclear decision making. So it's your communications, your ability to communicate with your nuclear forces, if you're the President, your ability to kind of command and direct the nuclear forces, have secure communications and be able to have authorized orders and then control those forces as well, including their deployment need to bring them back home. Right. If you don't want to use them, that sort of thing. And so that's an entire sort of infrastructure that includes not just the People in the chain of command, the people who are advising the President and supporting any decision making, but also all the technical means by which you can manage the nuclear forces. Some of the utility for artificial intelligence in the nuclear policy world comes from can you use artificial intelligence to better support nuclear use decision making? So can you more rapidly detect, as you outlined, an incoming nuclear attack? So maybe a President has more time to make a more prudent decision with more information available about whether he should attack now, he should ride out that enemy attack that's incoming or do something else. So there's a sort of rapid intelligence and battle domain awareness and force analysis fusion that can happen that probably takes people, even if it's a few minutes longer. Those few minutes matter a lot.

C (16:11)

But.

B (16:11)

But you might have be able to have some sort of frontier model that's integrated into the NC3 system that does that a lot more quickly and frankly, maybe more accurately. You could also have AI recommend options. Right? We think these targets aren't as important for the political objective you have. We think these targets have already been destroyed by other means. I mean, the type of conflict you were talking about, which is a general nuclear war, it's going to be a pretty fuzzy picture, right? You're talking about needing to worry about warhead fracture side. You're talking about targets that may not have been hit, but may have been destroyed because some other target next to them got hit. And how are human beings supposed to keep track of all of that in real time while the President is being forced to make decisions on a minute by minute or hour by hour basis? I mean, we're talking about some pretty hairy stuff. The other side of that is, of course, hey, if all these nukes are flying around like, does it really matter? Does this level of specificity matter? And Jordan and I, before we started recording, we're talking about these sort of, you know, we stipulated the insanity of a general nuclear war. But at least in the United States, we've always thought about like, yeah, but how can you make it slightly less insane? Or how can you actually achieve some advantage so that you're not a completely destroyed society at the end of that, but you're a mostly destroyed society. I mean, these are the types of debates that are very strange, loving, but you can imagine that little bit of accuracy advantage or decision making advantage that I can provide really could be incentivized in a US NC3 system, maybe less so in other nuclear weapons states.

A (17:46)

And before we move on, just for all the kids out there, the reason you have The Internet is because of this very question where the whole problem of command and control being such that you couldn't have the bases communicate with each other, like led various scientists in places like Sandia to try to come up with sort of distributed ways to communicate where you could have some part of the network fail and it be okay.

B (18:16)

Some of us Remember pilots in 2022, right? I mean, getting it in our house for the first time.

C (18:22)

And look, one thing I'd say on what Pranay said, first of all, I'm imagining know chatgpt telling policymakers, like, I'm not saying we won't get our hair must this hilarious Image, but anyway, ChatGPT

B (18:34)

looks a lot like Austin Long in that scenario. But.

C (18:38)

But it's really the question of like, look, this is obviously like nuclear use is obviously kind of a pretty fundamental, you know, barrier that we have. We as a species haven't crossed since 1945. And it's something that, that's a, you know, once you initiate that decision, you're potentially opening a Pandora's box to a whole other host of policy outcomes that we may or may not want. So like, that is the decision that has to be, you know, with a human, obviously. I think the scenario where like one's nuclear weapons start flying either way, like, yeah, at that point, like all bets are off. I'm sure decisions are delegated. I'm sure, you know, AI is probably making a ton of decisions, potentially even to include employment decisions, but not the initial one like that, I think is how good is.

A (19:24)

How good are tactical nukes at clearing mines in the water?

B (19:30)

I don't know. I don't think the US has any, but maybe we could ask the Russians to help. They have a lot, much more diverse array of tactical nuclear weapons. And you know, there's been people like Sergey Karaganov in the Russian academic space who've been like, you know, what we really need to do is light a nuclear weapon off so everyone remembers how terrible nuclear weapons are and then everyone will listen to us. So I, I don't know. Jordan, do you want to write a letter? I could help you draft a letter if you'd like to say, well, it

A (19:52)

could go the other way, or you just do like a Little Davy Crockett 1 in the Strait of Hormuz and everyone's like, oh, this is not that bad. What are you guys worried about?

B (20:02)

Some of your new listeners to Wartok maybe will really like the excursion that we're on now.

A (20:10)

Let's come back to. Sorry, I Interrupted you, Chris?

C (20:13)

No, that.

A (20:14)

Okay, okay.

B (20:15)

There's nothing as good as. I could never remember what I was saying.

C (20:17)

Yeah, I'll leave the Davy Crockett's and the straight upper moves be

A (20:24)

so pranay. You had this, you had this kind of long list of potential AI use cases when it comes to sort of like targeting and force planning and up planning and what have you. I'm curious because we have this big debate now about we've got China rearming, there's this question of nuclear modernization. How many more weapons does the US need, what type of them? Like where do you spend the money? Is there a world where these tools get you to a confidence level where you can feel like you have to spend less money to achieve the same amount of deterrence?

B (21:11)

Yeah, that's a really good question. I've only thought a little bit about this, but sort of thinking about the strain that the US is under where it needs to have a nuclear force that feels like is sufficient to do the many things that the US wants to be able to do with its nuclear weapons. In this case maybe deter two adversaries at once, support multiple adversaries and far off places in the world at once, etc. So there's going to be like a primacy, right, like on cost efficiency here because the US is not going to be able to just double its arsenal. And I don't think that would be a prudent choice, prudent expenditure of resources anyways and it takes a long time to do that. Making nuclear weapons, as I think Chris and I have both experienced, whether it's in the nuclear space, the delivery system space or any other major acquisition is extremely expensive, time consuming and then you find out five years later it's even more expensive and more time consuming. So finding any sort of efficiencies where let's say you have to use, threaten to use or use fewer nuclear weapons in to achieve a certain objective than you may have before you brought AI into your NC3 system, could be worth it, right? Because you could imagine a scenario in which if the United States has not, and we see this play out in conflicts in the conventional realm as well, if the United States is not very sure that they have destroyed, you know, achieve the weapons effects they needed to against a certain type of target, they may need to use additional weapons, right? So let's say the United States is trying to destroy mobile missile launcher that's in the forest somewhere. These things can move around and the intelligence information you may have may be slightly dated and so if the United States is trying to destroy that using a nuclear weapon and misses or isn't sure, it might need to use two or three or however many, because part of what the United States likes to do in its nuclear strategy is threaten an adversary's nuclear forces. So if you can do that more efficiently, let's say you can use loitering conventional capability or something that you know will be able to action very quickly upon an execute order being given and is already in theater and can do it more quickly than any of the US Nuclear forces, like, hey, guess what? That's a target that you don't need to have a nuclear weapon reserved for anymore. And so that could lead to not just less strain on this sort of nuclear force as it stands today, but in the future, if the US Is thinking about having a more efficient way in which it can execute these or pursue these political objectives related to nuclear use, maybe there's even a future where you can have fewer nuclear forces. That might not be right now, but it does lead to some potential benefits in arms control down the road. It does bring up an interesting question, which is, as we've been talking about, having a human in the loop for nuclear employment or a decision to terminate nuclear employment. And maybe this is something Chris will have more familiarity with if it, if the President says, okay, let's go on this particular option because I want to be able to destroy China's nuclear forces in this hypothetical conflict. Now, if you have a bunch of systems that are essentially autonomous and already in the region and that sort of employment order has been given, you can imagine a scenario in which these systems are then going to go autonomously, go and hit the targets they're supposed to if they're already in theater. And you may not have the President sort of approving the strike of each of those types of systems on a target hit. He's just kind of given this overall blanket, like I approve, you know, kind of option 1A, and that's what we're going to try to do. So there's an interesting question, I think, for nuclear policymakers, which is, yes, you want a human in the loop. Yes, you want the President or his successor making the original decision to begin nuclear employment. But do you need that decision applied to sort of every system that maybe has some sort of autonomous capability, of which of course, the US does not have in any of its nuclear weapons delivery systems now. But if you're thinking 30 years down the road, maybe people will see the benefits of that in the future, just to bring this back to the Skynet conversation we really want to have.

C (25:17)

I think like we. I think like we said, it gets pretty murky. I think, you know, it's very clear that the initial decision you want to have human control over, but then beyond that, I think it gets into the details of the conflict and it gets into. It gets hard. And I think there's going to be delegated decisions, and I think it's ambiguous. I think that even taking aside nuclear use, but just the fully autonomous weapons to be, let's assume for now without nuclear weapons on them, that is even we're seeing this play out in real time right now in the news. But that is admittedly a very murky and complicated area. I mean, obviously there's been a bunch of news stories about Anthropic's position on this and negotiations with the DoD, et cetera. But even on this, it's notable that Anthropic's position is not no fully autonomous weapons. We don't think that our technology is ready for it right now. And the reason for that is a recognition that actually probably we are going to have and probably need fully autonomous weapons at some point. Obviously, you want them to be secure and reliable, but just saying no fully autonomous weapons is probably not a militarily viable posture. And that's exactly why the US has opposed the bans on killer robots, is why it's proposed alternative frameworks for allies, and it's why DOD has 3009 and why anthropic is taking this position. So given that, then the question is, is there kind of a fundamental difference when it comes to nuclear use of autonomous systems? Like, is that a red line? And maybe it might be right. Maybe, maybe this is something where just like the added value of having, you know, a fully autonomous system in theater, as opposed to, you know, ICBMs and as opposed to, you know, manned systems that are out there is just strategically marginal enough, particularly given that, like, you know, once we're starting to get into, like, the nuclear use world, like, kind of all bets are off anyway. Like, you might be able to make an argument that actually the normative value of prohibiting fully autonomous nuclear delivery systems is greater than any strategic benefit that they could confer. I could see that, but I can also see how it's hard just because the fully autonomous weapons debate inherently is. Is murky and something where, like, red lines are difficult, I think I would probably come down on the side of I would be certainly comfortable right now and for the pretty long foreseeable future having A bright line saying we don't want fully autonomous weapon systems. And this is, you know, there's a reason why the US has expressed concern about some of our competitor adversaries, weapon systems that are unmanned and concerning. Right. I mean, the US has talked about the Russians for a long time about the Poseidon system. You know, that's, you know, something that, you know, there's not only, you know, strategic and arms control compliance concerns, but also technical concerns and concerns about kind of accidental use and risk and what that could mean for escalation. So, yeah, I guess my broader take is like, everything here, it's murky, but I think it's a fair thing to say that for the foreseeable future that might be another bright line. But like I said, I think there's actually pretty few bright lines in this. And I think that we spent a bunch of time when I was at the AI Commission and also at the White House thinking about, well, what are, okay, we have this nuclear employment decision red line. Okay, that is something we want to make sure is kept in human hands. But what's after that, we want to build on that. What are other things that we want to say, like really have to be. And there's not anything that's really clear. Like because of where the technology is going to go and the kind of inevitability of it and the dominance that you're going to get from having increased automation and weapon systems, there's reasonable discomfort from DoD from drawing red lines anywhere else. And I think the answer is that we need to make sure that the systems we have are really secure, really safe and really reliable and really meet our intent. And we also need to, we need to develop some kind of architecture globally that promotes other countries using those systems. Because if other countries are using systems that are prone to accidents, that's also very bad for us. That's a very hard thing to do. And we can talk more about that, but there aren't any very clear solutions there. But it's obviously in our interest. But yeah, in terms of any kind of hard red lines, they're relatively few and far between. Yeah.

B (29:34)

And I think where Chris is, has articulated a sort of a position in the sort of subsidiary questions on exactly how we specify the role of AI. Further specify the role of AI or absence of AI as it relates to nuclear weapons seems to really track with where the current administration is.

C (29:52)

I think.

B (29:53)

In one of the articles related to the anthropic issue in the past couple weeks, Pentagon spokesperson essentially said there's been no change to how the Department of War considers a human needing to be in the loop and any decision to employ nuclear weapons. And he said that there's been no policy considerations underway to put that decision in the hands of AI. Congress, I think in last fiscal year, maybe two fiscal years ago in their National Defense Authorization act had a sort of sense of Congress that stipulated a few sort of opinions about artificial intelligence and nuclear weapons. It was promoting AIML and decision support role. Something that we talked about a little bit, right, like sensor and intel fusion and, and sort of leading to a more informed decision making process with regard to nuclear employment. Made sure directed the Department to take care that integrating AI in this way didn't actually bring more risk to strategic capabilities. And also restated a version of the policy we've been talking about which is that there needs to be human safeguards and a human in the loop. And they even reference requiring positive human actions in the execution of decisions related to nuclear employment. Which to me points to not just the President giving his order and saying go forth and employ our nuclear force, but that anytime there's a decision, maybe even a delegated decision to a theater commander, let's say, let's say we live in a world where the US has lots of theater nuclear forces and needs to make these more sort of battlefield oriented decisions, even that commander needs to potentially be in the loop then for any of those execution decisions. Which to me sort of goes beyond the, the language we talked about in the Nuclear posture review, this P3 statement, this US China joint statement, and points a little bit more to where kind of Chris is leading, which is like what is the right amount of kind of automation as it relates to nuclear decision making. We don't have a real sort of, we don't have the Davy Crocketts we can use in the Strait of Hormuz anymore. Maybe there's a world in a decade from now where the US has more theater nuclear options like that because that's been identified by multiple congressional commissions and multiple administrations as a, as a gap that the US faces against Russia and to some extent China. But that to me is where these sort of like tactical execution decisions and I collide, right? Like how much of that do you want to be delegated to only a human. How much do you want that to be Based on an AI's quick analysis of the, how the battle space is developing? To me that's like where the really interesting conversation is moving forward.

A (32:30)

My sense is like the reason we're still having on these, these Human in the loop versus human on the loop discussions is because the technology isn't there yet to just press a button and have a thousand drones like do the thing. And once that does exist, there is as Chris said, a like very strong competitive logic to just having your drone fleet go over a country and figure out where all the, you know, ballistic missile launchers are and shoot them or just, you know, figure out like the proper way to do it. So I'm with you there on it being hard to imagine a world where there are like really strong kind of legal restrictions or ones that like stick around a week into a conflict or something. But on this sort of continuing to have the humans be part of, not just the president deciding it, but also the theater commander and then the two guys in the silo. I wonder to what extent of this pranay is just the sort of like hope and the sort of reasoning from some of these Cold War case studies where you had human beings who like could have chosen to interpret something more dangerously or less dangerously. And there's something kind of like nice about us all having a soul and not wanting to, you know, kill millions of people that, you know, we're, we're a little more comfortable and you know, we have a number of various like American and Soviet military personnel sort of deciding to like, you know, chill out for an hour and that be in that, you know, continuing to be something that we're all going to want to preserve in the future is just like the people who are in these jobs not really being like, you know, super excited to do the thing.

B (34:30)

Yeah, no, I think that's right. This is the, maybe this is the dovish and inspiring portion of war talk, but there's like a couple of fundamentals I think that I haven't seen evidence that AI is going to change. Right. One is it seems like people in positions of power, whether it's it was in the Soviet system, in the US President or you know, Mao in China when the Chinese first tested nuclear weapons and thought about use cases, frankly in the Sino Soviet split and the war that was, was going on at the time, people really don't want to use nuclear weapons. There are very strong incentives to avoid using nuclear weapons. Right. In a conflict. I think you're seeing a lot of the development of sort of drone technology, one way attack drones and automation or automation light that's used there. Whether it's in the Ukraine, Russia conflict and we've seen the rapid evolution of military technology used there or in the current conflict in the Middle east people want. It seems in general that countries would rather trend towards these sort of conventional non nuclear attrition based warfare models if it's possible because the consequences of going in the other direction are so terrible. Now I think you're right to point out that the, we've seen these sort of like heroic figures throughout the Cold War storytelling about these near accidents. As I mentioned before, I think all countries that have nuclear weapons have really worked hard to mitigate the types of risks that have been presented by those events. So you're not just reliant on somebody saying not today. I'm not turning my key because I think this is a fake or whatever. You're, you have an entire system and architecture that makes sure that no one person is really put in that position. And that's when we talk about AI for decision support purposes. Like you don't want the information that gets to the President to be bad information. Like you want him to have the best possible information or she to have the best possible information available before making such a consequential decision. I think our system has always been looking to optimize that sort of maximizing that decision time and maximizing the integrity of the information a president has. There's of course been. Yeah, go ahead.

A (36:35)

But, but here's my question for not like waymos are better at driving than humans and maybe they'll kind of make some mistakes that humans wouldn't make. But at this point I would take awaymo driver like 10 times out of 10 versus my replacement level human driver. Now the sort of human being making the targeting decisions or the human being making the kind of like intelligence judgment about you know, what's happening in the Politburo or the Kremlin or whatever, like clearly we're not there in 2026, but the amount of things that the AI can do better than the humans in five or 10 years just seems to be such that like we're, we're again like depending on legislation, we don't quite have the competitive pressures that you would, you know, as directly that you would have in a, in a, you know, corporate marketplace. But I'm just, it's hard for me to like imagine that a lot of this intelligence kind of like gathering, collection, synthesis, whatever, targeting work will just have agents do a better, more thoughtful, more thorough job than your sleep deprived 25 year old.

B (37:54)

Well, I think that's probably right. But there's also, I mean nuclear weapons use is inherently a political decision. Right. And so until we see these sort of agents be able to deal with that. And I think in large part that is, that takes away the sort of cold, strange, Lovian analysis of like, well, Mr. President, if we are able to execute our plan and take out these targets, we think the enemy will have no choice but surrendering. And then he's not thinking about the sort of political fallout, the willingness of the people in the other country to fight on. And like, I mean, these are all sort of behavioral and psychological calculations that could be analyzed and maybe AI can get pretty good at doing that. But when it comes to the sort of decision making that will take place, it's going to be a President's assessment of how is this all going to come together from a political standpoint, both geopolitically and in domestic politics. And so it's almost like our system was always designed for the President to have to make that fateful decision and for it to be essentially a human decision that's made, one that incorporates the President's own experiences, thoughts, feelings, you name it. And it's not just a, the product of cold analysis. Otherwise we just, you know, feed a nuclear war plan into, scan it into a computer and let the computer do all the stuff. Yeah, and we could have done that a while ago really without AI.

A (39:13)

And I think that's kind of the, like the, you know, the Iran is strike is a great case study for this. It's like, okay, great. You know, a computer can tell you with like 97% certainty that if you bomb this thing at this time, you'll kill, you know, the supreme Leader and all of his friends. But like, then what, like, is an AI really going to be able to sort of predict with a high degree of certainty like, you know, who's going to be the next leader? Like, is there going to be a, you know, civil unrest? Is that going to be quelled or not? So, all right, I think if you

B (39:46)

ask it, it'll probably give you semi intelligent ideas. But you know, I think we spent, Chris and I both spent a ton of time doing tiger teams and playbooks and things like that to do this sort of scenario based planning. And that was a very human intensive effort. And you can imagine your starting point with AI might not be so bad, but you ultimately, like, you bring people in because these are people making decisions not just in our country, but in adversarial countries where you might be engaging in this conflict.

C (40:15)

Yeah, and it's interesting, there is the studies recently that showed that in War games and Pranay can probably speak to this, but in War Games AI is actually substantially more prone to resorting to nuclear weapons use than humans. And obviously this is right now. Right. So this could change in the future, particularly as models get better and better reflect human behavior and intent given presumably human intent is not to always result in nuclear weapons use. But that said, that just shows to me you're close.

A (40:42)

Like, when people play war games, don't they always want to use the nukes? Isn't that like, like the last day? It's like, okay, I guess we'll just use the nuke.

B (40:51)

I think it's like, but it's. These are sometimes contrived, right? Like, sometimes it's what you want your war game to test. If you want your war game to test the likelihood that an agent will use nuclear weapons, as Chris is outlining, that's very different than testing, like, how easy is it to get to restore deterrence and get to peace after nuclear use. In the latter scenario, you actually need the countries, the games, the game countries involved to use nuclear weapons. So then you can kind of test like, okay, how do you reduce, how do you limit escalation from there? So yes and no. Right. And also it depends on who's playing. Some people just like to pretend to use nukes.

C (41:25)

Yeah. And it's not to say that that's like the system is inherently prone to that, but it just shows that, like, given the gravity of the risk and also given really it's like the minimum, the relatively minimal cost. Right. Like the cost of having the President make that initial decision is really not that high because like we, yes, it is going to be a very stressful few minutes, but the system is very well set up to do that. Like, there is redundancy if even in the event of a decapitation strike, like, we've planned extensively for that. So to, to remove that entirely adds a substantial amount of risk for not that much benefit. And if you actually think about why, why do other countries have automated decision use? I mean, really only one. Like, it's not because there's some kind of like massive strategic advantage. Like, the Russians don't think like, oh, there's a dead hand gap and that's why we need our own dead hand. Right. No, it's because they don't trust their people to use the we and because they don't have as professional of a military as we do. Whereas I think we generally have a high degree of confidence if the President issues a nuclear use order, our people will follow that. And that is why they train extensively for that. So therefore the utility of having that be kind of automated down the chain and even from up top is a lot less. Whereas in their system I think that there's some questions and particularly in the event of decapitation strike, I think that all bets are off. So therefore having it just be automated might actually be preferable, but like very different circumstances.

B (42:55)

I think you highlighted one risk Jordan about in the decision support space which we haven't spent a ton of time talking about. And I would recommend people read. There's a new Texas National Security Review roundtable. Our former colleague Mike Horowitz and a bunch of other scholars contributed to that. People should take a look at. It's right at this sort of AI and strategic stability or nuclear deterrence set of issues. But one of the concerns I think that's been expressed outside of government is, you know, if you bring more AI agents into the decision making and decision analysis and support process for NC3, don't you create new areas of potential cyber vulnerability where adversaries can sort of plant deepfakes or fake information into this decision making process in ways that they haven't before? I'm not sure how you guys think about that. To me that's sort of like a different flavor of an existing problem which is sort of cyber vulnerabilities and NC3. But that's something that's been highlighted in the scholarly community and perhaps maybe focused on a little too much given kind of the, the limited way in which we're talking about artificial intelligence sort of like slowly crawling into the sort of nuclear decision support space. But I'm not sure if you guys have opinions on that in or outside of the nuclear space. Not to take over Wartok Jordan, but that was something that crossed my mind.

C (44:16)

Chris, Then the misinformation problem that we face with I just crossed the board that I think that everyone kind of wants to apply it to their pet issue. But I think the fundamentals of it are actually pretty similar and I think it's actually pretty unclear how that's going to play out because you know, first of all you can use AI to check whether or not something is made by AI and whether it's misinformation. I mean even just like go on on Twitter right now. It's interesting, right? Like yeah, there's a bunch of information misinformation but like even GROK will generally, even GROK will generally like at least can identify at least a big chunk of things that are clearly false very quickly. So, so there's, I don't know, it could cut a bunch of, a bunch of different ways, I don't think. It's like, I don't see a lot of applications in the nuclear space that are like fundamentally unique and indifferent in my mind.

B (45:07)

The other issue I think that's been highlighted, sort of how AI interacts with nuclear deterrence is whether it, you know, to use a turn of phrase, it sort of turns the oceans transparent. Right? If, if your nuclear platforms, your safe second strike is based on ballistic missile submarines and adversary countries are able to crunch data in a way with coming from satellites, undersea sensors, you know, you name it, that makes, that increases risk for ballistic missile submarines that could be game changing over time, right? I, I think that that's not that close. And then I, the sort of question is sort of how can you use artificial intelligence in a defensive mode to prevent that type of early detection from happening. But like to me there's probably going to be a significant undersea competition related to AI integration that impacts nuclear deterrence. And I again, I don't know that it's that close, but it's something that, I really think that if you're the US and you put the large substantial portion of your nuclear forces on submarines because you sort of are the best at undersea quieting right now, you could envision that Even like a 10% increase in risk there might change how the US thinks about deploying its nuclear forces in the future.

C (46:22)

I am profoundly worried about this. I just, it seems infeasible to me that we are going to be able to hide a ship that is hundreds of feet long, weighs millions of tons anywhere in the world with the technical detection capabilities that are going to become online. Just because the whole advantage of AI is being able to parse out the signal of the noise through the noise. And you're going to need so much less signal in order to figure that out. So whether that's undersea detection or from oversea space based stuff, it seems like the idea that we can hide these massive things just because they're in the ocean with the extremely advanced technical detection capabilities that are coming online is just something we can't bet on in the next five to 10 years, let alone the next 50. So does that mean we should scrap the Columbia CAST class summary? No, I don't think so because it's just too important. But I think we have to plan for the eventuality that it might not be the invulnerable second strike capability that we think. And that's really scary when you're planning 3050 year procurement decisions that are hundreds of billions of trillions of dollars. I don't know, I'm not sure that we're. If there really is a sea change here, pun not intended, then I think that we need to posture ourselves accordingly.

B (47:48)

Yeah, and that might be an area for, I don't know, in calmer times, you could imagine countries coming together to say like, hey, we think we should try to avoid risks to our stable second strike. Like we can do the advantage seeking and competition elsewhere, but for ssbn, we don't want to do that. Now. The problem is, I think for the US is given our nuclear strategy, like we, we want countries to have stable second strike, but if push came to shove and we went into the, the type of nuclear war that Jordan outlined earlier in the podcast and the US is trying to attack adversary nuclear forces, then you actually want to have those advantages in detection. I think the US is probably pretty good at that. It's probably leaps and bounds ahead of other countries in being able to do that anyways. But if you think about the benefits that Chris just outlined of integrating AI into creating those risks for undersea platforms, then I think the US would not want to foreswear that.

A (48:43)

Right.

B (48:44)

They'd want to keep pace or be better at it than other countries. So that, that to me could fundamentally change how we've thought about sort of stable nuclear deterrence, mad, whatever you want to call it since the end of the Cold War. And again, maybe it's not here now, but I don't see why it wouldn't show up on our doorstep as we think about these issues. And then in the coming several years,

A (49:05)

can we do a little proliferation detour?

B (49:09)

Yeah, I thought this was all a detour.

A (49:15)

I mean, we did a show with a very nice Japanese scholar. I think this was like March of last year. And I remember the big takeaway that I had from hit with this guy being like, look, at the end of the day, if America's nuclear umbrella goes away, we'll just get nukes because we hate China that much. And we've had a number of pretty dramatic developments over the past 12 months. Probably capped off with Iran getting, you know, Iran's leadership getting bombed, you know, presumably because they weren't able to put the same cards on the table that North Korea has been able to do over the past five years. So I mean, are we just going to be living in a world five years from now where there's 10 more countries with nuclear weapons? I mean, the odds seem to be increasing where what's the, what's the right way to think about all this stuff?

B (50:09)

Yeah, I'm happy to start. I do think there's an AI angle to this as well. So I think a lot of countries are under stress for the reasons you outlined. Right. Is the US commitment credible still? Both from what is the political messaging coming out of the executive branch? Is the US going to spend enough money on the types of nuclear forces that are most relevant to the things that allies worry about, et cetera? The politics part of it is something that, you know, is, is always going to be kind of a toing and froing depending on who's in the White House, I think. And then the capability side of things is, is, as Chris outlined, it takes a long time to do stuff like make a new nuclear weapon. So I think there's a lot of, a lot of allies and I was in Japan recently for a dialogue. There are a lot of allies who are much more open to the idea of needing to hedge against a sort of US retrenchment or failure of the US to kind of grow its nuclear force or adapt its nuclear force for the new era. I think that's very real. I don't think that any country is like on the cusp of doing it. I do. You know, if you think about President Macron's speech last week, clearly France and the United Kingdom are. The US Nuclear allies are thinking about this risk as well. They don't want to see proliferation in Europe, so they're trying to figure out ways to develop more cooperative mechanisms to not necessarily extend deterrence to replace what the US does for its European allies, but complement what is the kind of existing nuclear deterrence paradigm. So France has talked about, or Macron talked about maybe deploying French nuclear capable aircraft forward in other countries, growing the French nuclear arsenal from what I think is currently around 290 or 300 warheads to some indeterminate number higher than that, and forming these more kind of cooperative dialogues where you might do joint conversations with other European allies on the threat environment, where nuclear policies should go, where the French nuclear arsenal should go, how they can more readily adapt to sort of defending an ally as opposed to just themselves. So it's a very lively debate in Europe and I think one of those might. That's probably going to play out in some way in Asia too, I guess where artificial intelligence comes in to me. Is proliferation easier now because of artificial intelligence? I mean, you could imagine artificial intelligence being used for overcoming sort of the sort of prying eyes of foreign Intelligence services that would want to stop them from pursuing the bomb, finding alternative materials, doing like modeling and simulation of different types of materials to achieve more rapid centrifuge design and efficiencies and enriching nuclear material. You know, coming up with explosives designs and lens designs and all these types of things that go into making a nuclear weapon. You can probably find ways to use artificial intelligence for supply chain optimization, avoiding export controls. There's an entire architecture of non proliferation related export controls that much of the world has agreed upon at the end of the Cold War to try to prevent these types of materials and technologies from getting into the hands of states, non nuclear weapon states that could proliferate. We have failures, right? With North Korea we have some successes. Iran has had faced a really steep hill to try to get there. Then eventually their program was discovered in the early 2000s and we've been in this sort of like on again, off again conflict or, and diplomacy with Iran since. And then, you know, are there ways in which artificial intelligence can help automate your analysis of the sort of tomes of scientific literature that are now out there since 1945 about nuclear weapons design and assistant engineering? Right. How can you set up a civil nuclear industry that's optimized for a clandestine weapons program? I think all of those things for a country that doesn't have a lot of money but is really worried about its security environment, could really be useful in ways that have not existed before. So I do think that that creates a new set of proliferation risks that we haven't really had to deal with in the past.

A (54:10)

But Pranay, what's the like threshold there? I mean presumably if you are a big boy country like Korea or Japan or Poland, like you figure this out. So is what you're saying more like if you're, I don't know, one or tier two tiers down, if you're like, I don't know, a Vietnam or something, then it then like what would seem to be kind of a bridge too far now become something you really want to be able to consider.

B (54:40)

I think that's probably where the benefits are maximized. Right. If you're one of, if you're sort of a smaller country, you're not going to be able to expend the resources. You don't have an existing civil nuclear program. Right. Japan has tons of stockpiles of reprocessed nuclear material. South Korea has a really mature civil nuclear industry. They sort of are, they're sort of 10 yards down the, the straightaway anyways. And there are other examples of that too. Other. A lot of countries are interested in nuclear energy now because of the AI data center boom. But for countries that like haven't given it much thought and don't have a civil nuclear industry, don't have the technical expertise, that's probably where you have the like maximum benefit of.

A (55:24)

So if you're just like a Kazakhstan and you get some like, you know, cracked version of ChatGPT or some like Chinese open model circa 2032.

B (55:34)

Jordan Steiner, dictator of a tropical island in the Western hemisphere. You know like that's.

A (55:41)

And that sort of also played Venezuelan leader. Let's not, let's not.

B (55:45)

I don't want to cut you. And then non state actors as well.

C (55:49)

Right.

B (55:49)

We're talking about sort of the, the sort of like the, the that end of the spectrum. But yeah, for mature countries, I mean some of this optimization stuff may help. But if they really are committed to do it and are willing to face the kind of consequences of doing it. Right. International opprobrium, end of an alliance with the US High cost. The potential that an adversary wants to kill their program before it really gives birth to a weapon. I mean those are all the other risks that matter a ton.

C (56:17)

Yeah, I think that fundamentally, I would say there's not that many countries. It is such fundamentally a political decision to pursue nuclear weapons. And for every country that is either aligned with the U.S. or a partner with the U.S. whether or not they are going to be able to do it more easily from using AI or not. The threshold for most of those is whether they want to cross that political Rubicon. There are some countries that might be in the bucket of would like to proliferate but can't. Myanmar or something. Kazakhstan maybe. Kazakhstan actually had them gave him up. So even then I think it's relatively few. The non state actor point is is fair. I think the interesting question here is more what are the strategic capabilities that AI is going to confer that are not nuclear but are, are still very strategic in nature and have the potential to kind of massively change, you know, force posture, architecture, how we do, you know, big grand strategy, planning. Right. Like that. That is, that is the sea change. I think that can happen with AI and you know, you can. Whether it's like, you know, mass precision strikes or something like using you know, tiny drones that you know, hit every single individual target you want or something like that. Like you don't even need a nuclear weapon because you can actually create the precise effects you want on a much more micro Scale. So like that is, you know, the proliferation of those capabilities I think is something that we should be thinking a lot more about. Less how AI is going to contribute to the architecture of the previous century and more how it's going to create a new architecture in the previous, in the current century that is going to pose similar strategic risks for the US and we do have this sophisticated non proliferation regime for nuclear weapons that is obviously there's many kind of US and allied led efforts. There's also a global regime set up in a very different era. But look, we're facing all sorts of debates on export controls with AI right now. And I think we're debating the, the kind of, you know, the onesie twosies at this point, right. We're debating whether or not we should be enabling China's AI capabilities or not, which is like the fundamental here, right. And, and I think there's actually much more sophisticated and complicated debates that we should be having about like does there need to be a kind of non proliferation regime as it pertains to like strategic, militarily, strategic AI capabilities. And I think what we're seeing in the Iran conflict is like obviously AI is starting to confer very significant military capabilities for advanced militaries, right? And like that is only going to go up and that's building off of, of the Ukraine, Russia conflict, the Armenia, Azerbaijan conflict where we've seen this kind of percolating and now we're seeing the US military starting to apply it at scale. So like this is happening, it is happening fast and we need to think about like look, the US military is going to want to preserve this edge. It needs to preserve this edge. And a world where every other country can kind of have mass precision strike capabilities at scale is of probably a pretty scary one for the US because it means that our conventional advantages are going to be a lot less than they used to be. And then that gets you into, into a lot more discussions about nuclear power. You don't want to have so or kind of threats. So what do we do to prevent other countries from getting those capabilities in the first place? That's the conversation we should be having right now, not whether or not China should have them, which is unfortunately where we are.

B (59:54)

Yeah, I think Chris points out a really damning feature of this debate which for people who are kind of nuclear policy experts who are focused on nuclear deterrence in particular scenarios. I think one thing, one thing I would say up front and I think most people would agree is that we can have all these conversations about why the US Nuclear force needs to grow or change or become more diverse or not, how AI should fit into it or not. But in a scenario like, let's say that there's a conflict over Taiwan, where China is able to utilize its local military advantage, conventional forces utilize artificial intelligence to sort of retain that advantage, and pushes the United States into a really difficult decision, which is whether it should use nuclear weapons or not first in that conflict. I don't think it's a good thing for any American president to be put in that position. Right. I mean, part of why the United States wants to maintain the strongest possible conventional force, which is becoming harder and harder to do, given that our interests are widespread and adversaries are building up, is because we want the onus for nuclear use decisions to be on the adversary. We want them to have to face the, the sort of tax of breaking the taboo that's lasted 80 plus years in any kind of conflict. Now, if you're the Russians, maybe you feel a little bit more comfortable doing it than other countries, but I tend to believe that most countries don't want to be the first to use nuclear weapons after 1945, and I think that's true for the US as well. So you can really gain an advantage if you can bring AI into your military capabilities in a way that preserves some sort of local military advantage. Because the likeliest cause of nuclear use is going to be one of these sort of extended deterrence crises that goes into a conflict in either Europe or Asia, far away from the US and the US Is faced with a decision, if it's losing a conventional war, to use a nuclear weapon or vice versa. The adversary is faced with that decision. You want that pressure to be on the adversary, I think. And so for all the reasons that Chris stated, it's really important for the US to not face that disadvantage and be in that place where the onus for nuclear first use is on the US And I'm hoping that, you know, as the US kind of navigates this space and how to. How to think about artificial intelligence and its military capabilities, that's one of the driving factors here.

A (62:01)

Well, so Pranay, you did a show with Mike Kaufman on the excellent Russia Contingency podcast back in October, kind of reflecting on the experience of that fall where Ukraine was kind of moving and grooving a little bit and you had Russia and Putin in particular start to make more noises than have ever really been made, at least definitely since the fall of the Cold War, about the use of nuclear weapons. I'm curious, five months past that podcast, do you have any other reflections you'd want to share about how that experience has continued to inform your thoughts on this work?

B (62:46)

I mean, not really. I still think that at least for people who work in this field, in this kind of post Cold War era, that's still probably the closest we've gotten. And we have to take moments like that seriously. It was sort of a very real example of where an adversary may consider using nuclear weapons. And again, it's comes from the dynamic that I highlighted that Chris started off, which is the potential failure of a conventional war campaign.

A (63:10)

Right.

B (63:10)

The potential eradication of conventional forces giving rise to consideration of nuclear use in a country. I still think that that's probably where we could potentially see nuclear risks arise once again. Maybe it's in a different part of the world, maybe it's a different adversary, different belligerence in the conflict, but that's what I see as sort of the greatest risk for nuclear use. Now I'm hopeful that there's not going to be a conflict like that ever again where, where nuclear use is seriously contemplated by any country involved. But that still kind of motivates me to work in this space and think about nuclear risk, including kind of new nuclear risks that might be presented by emerging technologies. Because I don't think we, and I think people will disagree over how, how close was it and were the conditions met. And that sort of debate has been playing out ad nauseum for the past four years or so. From my perspective, like, any closer than, like absolute peacetime is close enough. Right. So we should, we should sort of focus our efforts on making sure even that risk, however high or low it was, is mitigated to the greatest extent possible.

A (64:17)

Chris, like, like, you know, Pranay had a pretty busy fall. When you're sitting in the White House, like, not doing that, but like your boss is worried about nuclear war starting, like, how do you, how do you feel? What was your, what was your experience of that moment, the, the kind of

C (64:37)

Russia moment in 2023? Yeah, well, look, I will say, you know, Pranay was in the thick of that and, and for good reason. You can't have everybody focused on that issue because the rest of the gears of the government do have to keep operating and running. And if, if, you know there are, there is something that is the most important thing at any given time, but if everyone is focused on the most important thing at any given time, then you're never focused on anything but one thing at a time. So you know, I was largely, largely separate from that. But look, it does induce like enormous amounts of stress. It's interesting. I mean I think you do kind of when there's things like that that happen, whether it's, you know, this high level concern or any, any other, you know, initiation of a conflict or something, you do feel it throughout the whole system in just because senior leader decision making and focus and attention and time, which is your most valuable resource always in government is all focused on that issue. So it means that you can continue to kind of move some of the smaller rocks on other issues, but you can't move any of the bigger rocks just because no one's going to be there because everyone's just focused on this one thing. Yeah, I mean I would defer to Pranay and others sense of exactly, you know, where we were in the kind of, you know, exact, you know, how likely things were. But obviously look, we're, there was a, it was a pretty unique moment. I think we're, we're, we're likely to see other, other moments like that where I think senior leaders like especially in other countries like very seriously ponder nuclear use, especially tactical nuclear use in order to achieve battlefield ends. And I think, you know, as, you know, particularly in a circumstance where they're at a conventional disadvantage in a potentially existential conflict and we can think of other conflicts that could kind of rise to that level. Yeah, this is like a super possible reality and it's really scary. And I think that the there, there is real. I think some people think like oh, the nuclear priesthood like is kind of always focused on the issues of the past. But I think that when you read, when you hit these moments when it becomes really real, we might actually see a battlefield nuclear weapon explosion having people who have actually done a lot of thinking about like how could we react? What are the options here? What does this mean if it does happen is just supremely valuable because if that spirals out of control, the world can end and you're kind of preparing for a relatively low likelihood eventuality. But these are exactly the very high ROI moments where we want people who have like thought a lot about this and are you know, able to react calmly and rationally or you know, ask ChatGPT or Pranay.

B (67:31)

Yeah, or whoever's in the job. Another thing it highlights though is like we, I don't think we've, I don't know Jordan, if you've done a warcaster, a China Talk episode on House of Dynamite, but the way in which we have From a pop culture perspective, thought about nuclear war since the Cold War, really in the latter stages of the Cold War and then through this period of relative peace after the Cold War has always been this kind of like, again, it's like war game. It's like all the nukes flying at once. What are we going to do about this? This event in 2022 highlighted something very different, right? Which is the potential for local tactical nuclear weapon use, probably low yield, specifically for battlefield effects, as Chris highlights, and probably for the people who are thinking about making those decisions with the hope that it wouldn't lead to further nuclear use or that, you know, that it would sort of end the conflict on the terms that the nuclear weapons user wanted to end them, which is very different, right? It's very different from this sort of spiral escalation and all the nukes fly at once and the world ends. And we like in the. In the outside government community, and frankly, it's motivated some thinking within the government community, have just gotten used to this idea that when nuclear weapons are used, it's probably end of the world. And the reality is like, while any use of nuclear weapons is the use of a strategic weapon, I don't think that there's. I don't think there's a way to sort of use a nuclear weapon and not invite further nuclear weapon use. There's a lot of incentive for us to think through. How can you ensure, if somebody makes that mistake or if that does happen, how do you, like, restore deterrence such that there's not this global nuclear conflict? And to me, that's one of the interesting features in this AI study that, that Chris mentioned that was conducted by a professor at King's College, is if you really get into the heart of like the 50 or so pages of this study, looking at how these different models that were used, and I think it was. I think they used Claude Chatgpt and then one of the Gemini models in their simulation. Where did these models decide? Based on inputs like Herman Kahn's Escalation Ladder and Schelling's books from the Cold War, where do they decide it was worth exchanging strategic nuclear weapons? I think we haven't done enough of that as human beings the past 30 years to think about that effectively, and we need to do a little bit more of it. And this event sort of highlights the need for us to think through this stuff a lot more carefully in the way that Chris has outlined, even if it's not me, because obviously I was just kind of, you know, lucky enough to be there at the time.

A (70:05)

Lucky. Yeah.

C (70:07)

The.

A (70:08)

And it comes back to the, the. The proliferation discussion, right? Because, like, the reason you can do that, like tactical battlefield play, is because it's not US and Russian, you know, or Soviet forces, you know, in the fold, a gap, right. And there's like a higher probability of it not spinning out into World War III. Pranay, like, has this energized, you know, has 2022, 2023, like, energized the field? Is there, is there, are there more students, more funding, more excitement around this stuff?

B (70:41)

You know, it's really interesting. It's interesting. And maybe to the benefit of people who are focused on AI issues, a lot of the. So as you know, a lot of the research that's conducted outside of government is really based on what these sort of philanthropic foundations that work in the national security space are willing to fund. And at least I've noticed, you know, last year I was at mit, and now that I'm at Sandia, I've noticed a lot of the funding has driven towards sort of the AI nuclear nexus without a very kind of clear understanding of what is sort of within that nexus. Like, where do these intersections happen, under what kind of categories of policy, what's sort of the right answer, wrong answer? Some people really are focused on building the international architecture that Chris mentioned. Others are focused on AI integration into NC3. And a lot of this is happening outside of government, where there's not a lot of information, frankly, about what's going on in government. And so it's almost like that trend was happening in the funding space while this event occurred. And I've seen a few people who've taken on projects to focus on, like what is ending a limited nuclear war look like. And there's sort of war gaming associated with that and sort of understanding how those scenarios play out. But by and large, it seems like the nuclear policy community outside of government has really sort of sprinted to try to tackle some of these AI and nuclear issues to sort of varying degrees of questions and quality of analysis and things like that. So it's energizing for me, I guess, and I think it's energizing for other people who have had government experience or they're in the government at the time, but for the community overall, not as much. There are a few exceptions. A mentor of mine, a former colleague at Carnegie, George Perkovich, is trying to write a book, I think publishing soon on this episode, and kind of what it could mean. I think there are others who have kind of Dived into this issue with podcasts or articles or things like that as well. But yeah, I think there's like a pretty fundamental, it's interesting, I think there's sort of a fundamental, there should be a fundamental shift in how we have been analyzing these issues for the past few decades at times of relative peace. And I haven't necessarily seen that shift. I, I, it shifted me and I'm focused more on it. I think there's others who are like me who are focused more on it. You know, vivid. And I wrote this article last year because we think this sort of is, is the clearest indication of there being a new nuclear era, the fact that this event occurred. But I haven't seen this sort of academic community catch up and maybe I'm just not well read enough. And it's out there and I haven't come across it, but, but I don't know, that's my perspective on it.

C (73:21)

One other thing I'd say is I think Pradesh talked about how the nuclear community can kind of adapt to the world of the AI threats. I actually think that there's also the inverse, which is how can the AI community not necessarily adapt to the nuclear world. But look at the lessons learned from the last 75 years of the nuclear community, which is not saying AI is equal to nuclear weapons, they're different. But like nuclear weapons were the strategic technology of the last 75 years and AI is a strategic technology of the next who knows how many, but certainly next 20 years, maybe the next 75 years. Right. So thinking about, there is so much thinking there about how nation states adapt strategic technologies and how they adapt their force posture and what it means for state behavior and kind of proliferation and deterrence and state power. And like a lot of that is actually somewhat applicable if for no other reason than because it's the only example that we have of kind of technology dictating large scale state behavior and kind of strategic relationships. So I think that there's actually a lot of lessons learned for the AI community on that front that should be more utilized. And I think that there's frankly some analysis that are reactionary things that are a little bit lazy on that front when they get to that where people are like, well, we have an IAEA for nuke, so we should have an IAEA for AI. And it's like, okay, but like that, that is not, that is obviously not in the cards right now. Right. I mean, I think, I don't think any US government would just kind of sign up to that. I Think that was the product of a little bit of a unique moment in the post Cold War period where also you had a kind of, you were able to enshrine a regime where there was like two tiers of states, nuclear weapons. And there's a bunch of reasons why we're just not going to get to that. It's not a feasible political outco. But like what do we do in the absence of that? Like what does a, what does a global AI order look like? Because I think a lot of people just kind of disregard it and they're just saying well we can't do that because this is a commercial technology. But I don't think they're really thinking about like well what does that mean? Like what does it mean if we don't have no non proliferation regime for AI? And just we're like yep, every state can just develop the military capabilities they want with respect to kind of strategic precision strike and like really advanced, you know, connectivity and stuff like that. That's probably not a great world for the US milit military or for the US economic and technological competitiveness. So this is probably something that's in our interest and then the question is how do we do it? This is a subject for a much longer form, discussion and detail and probably writing. But fundamentally there's a lot of lessons I think from the nuclear age in terms of infrastructure. The way that we have a non proliferation regime for nuclear weapons is really infrastructure focused. Right? We obviously, yes, we don't want people to have the nukes moving around, but it's the visible material, it's the centrifuges, the enrichment capacity, et cetera. That is how you actually manage the proliferation of the capability to develop the capability. And there's a lot, I think there are parallels there on AI. Now obviously it's much more complicated because it's a dual use commercial technology. But this is, this is moved on from a kind of niche US export control debate to I think that this needs to be thought about in the sense of what does an AI non proliferation regime look like that enables the full commercial adoption of the technology in ways that are very, very important for the US and the world that we can't hold back. But that also makes sure that our rivals are not using this for military purposes. That's not just them signing a form that says yeah, we're not going to give it to our military. Because I mean, can you imagine if that was how we managed nuclear proliferation and be a joke, right? So yeah, I think there's a Lot more thinking to be done in this space and parallels that I think are appropriate and there's a lot of lessons learned from the last 75 years. It should be applied here.

B (77:40)

Yeah, I think there's, there's a ton to delineate of. Like what are these sort of foundational elements of developing AI that's relevant to certain types of military use of AI that creates destabilizing dynamics. Right. So there's like a ton to work back from. I think one starting point we have to think about is can we better understand what the deterrent effects are of having AI enabled military capabilities? I think right now we've observed automated military capabilities being used in relatively local conflicts where one side seems to have a large advantage over the other or the sort of attrition based conflict like we're seeing in Ukraine and Russia. And so it's unclear to me looking at the two conflicts playing out in Europe and then the Middle east, like how are these countries changing their military behaviors because they're worried about the relevance of the other side's AI enabled military capability. So a one way attack drone that can easily find Russian soldiers hiding, for example, how the Russians adapted that, because that to me signals they've changed their way of war. Which also suggests that the Ukraine's Ukrainians ability to have these autonomous kill systems has a deterrent effect. And so as we have a better understanding of the deterrent effect and how military operations are affected by these capabilities being introduced into a conflict, that could lead to some of the architecture that Chris has discussed. Because, because there might be really destabilizing applications of AI and military technology that causes a country to go really high and. Right, right. Consider really escalatory responses. And that might be something we want to control. So that allows you to work back to, okay, what are the types of sort of fundamental technologies that enable the use of AI in that way? And we want, because we want to avoid this outcome that really parallels how countries thought about theater nuclear weapons. Right. And the PNI initiatives in the 1990s. I mean, I mean at some point when the prospects for war between the United States and Russia after the Soviet Union fell apart really diminished over time. And we were in a period of like, okay, how do we manage the fact that we have tens of thousands of these nuclear weapons in countries that are bordering one another that we don't want to use, but we are more worried about accidental use. That led to a series of initiatives by President Bush one as well as Gorbachev and then Yeltsin to sort of withdraw those types of capabilities from the field. And the US dismantled a lot of these capabilities as well. We might have some instructive lessons to draw from a deterrence perspective from the conflicts that are going on right now that allow us to work back to the, okay, these types of applications, these types of fundamental technologies to develop AI for use by militaries in this way aren't really in our interest. So like, what, what should we do to sort of regulate that? That's really hard to do because I think what we're seeing now is everyone's using these types of capabilities for relatively minor and local goals. Right. Killing individuals or taking out certain nodes that are important for the conflict that's, that's coming out. But we don't really have a paradigm by which we want to think about guardrails on those types of conflicts yet.

A (80:46)

And I, it kind of comes back to the, the fall. Was it 22 or 23? It was 22, right? Yeah. The fall, 2022 arc of this question of like, how, how, how badly do you want to. Or like, like, like you want to win the wars you're in, but there is this danger in winning them, like too hard or too fast. And it's kind of the same thing with a sort of nuclear balance, right. Where like, if you're, you know, if your second strike capability is like so overwhelming and you have like total nuclear overmatch, then the other side is not just going to roll over. They'll come up with other stuff or they'll be kind of trigger happy in a way that actually makes the world a lot more unsafe and more likely to start the horrific war. So I don't know if there are answers to this because you can't know how hard you're going to win before the war starts. And it's not like you're trying to lose. But having that dialogue just right seems to me to be an impossible thing.

C (81:56)

I don't know.

B (81:56)

Yeah, I mean, it all comes back to like, sort of fundamental asymmetry of stakes issue that we've been dealing with and throughout the Cold War and frankly throughout history. Right. I mean, like, so the U.S. the idea of the U.S. fighting a war on behalf of an ally that's really far away led to some, like, fundamental changes in how the US formed its nuclear posture and its military posture overall early in the Cold War. Because at some point, you know, a country that's a frontline country, like West Germany at the time is thinking, why, why would we trust that the United States is willing to sacrifice American servicemen, American cities, American economic power on our behalf in a war with the Soviet Union, like, we just kind of can't buy it. I mean, it led to a lot of changes in US nuclear strategy. And it's sort of why we have the strategy we have today where you might have some forward deployed forces, you might need a diverse nuclear force that includes theater nuclear forces or low yield weapons. And then you need this overwhelming capability to potentially go first preemptively to destroy an adversary's nuclear forces. Because that's the only way that an allied country may believe the US is willing to fight that war that you could threaten. If things really went crazy, the US wouldn't have to worry about sacrificing allied cities or US cities. It could actually destroy the adversary's nuclear forces. Now we can talk about like how credible that was, how destabilizing it was, what the after effects of it are, etc. But that fundamentally was like the core of US nuclear strategy for a really long time. Now the version of that that could apply to the artificial intelligence world is sort of, you know, well, if the, if the asymmetry of stakes issue is part of, is solved in part by how the US adapted its military strategy towards the Cold War, essentially what was assuring to that ally in West Germany was also deterring to the Soviet Union leadership who may have had designs, let's say, on exploiting a Berlin crisis and taking territory in West Germany and reunifying Germany under sort of the Soviet flag. This is a hypothetical. I don't think there's any record of that actually being the case in the Soviet Union. But you know, follow me on the hypothetical. Well, then you can imagine that like, okay, so what does the US need to do to further ensure that a country doesn't want to sort of pick that type of fight. And from the asymmetry of stakes perspective, you're talking about, okay, so like, how can the United States adapt the integration of AI into its military policy to further convince an ally and an adversary that trying to pick a conventional fight isn't going to work? I mean, from the AI's perspective, as Chris laid out, or from the US's perspective, as Chris has laid out, if the United States is able to do that more readily at the conventional level of conflict, essentially deter conventional conflict from ever occurring. Some of these risks of nuclear escalation can actually be shifted upward or to the right where, you know, the adversary doesn't want to even start the music on little green men or Gray zone activity, cyber attacks, or, you know, a conventional conflict, because the AI, the US Is able to signal its dominance in AI military integration pretty outwardly in a way that not just convinces the adversary they're going to use, but also convinces the ally that the United States will defend them. Because from a cost perspective, like, we think we can actually win the conflict much more readily than we could before we integrated AI. That's one of those sort of, like, interesting deterrence dynamics that AI integration could. Could sort of bring to the table, as Chris outlined. And then if you're an adversary, you're like, hey, you know what? We really want to regulate how the US Is able to integrate AI so effectively in their military capabilities, because it's giving them a clear strategic advantage.

A (85:31)

Well, Chris, do you. Do you buy that as, like, there are kind of multiple ways to read the US Performance of the first few weeks of the Iran war? Like, on the one hand, like, only we have, like, six Americans who have died, which is a pretty low number, all things considered. But then you have this, like, broader narrative of, like, we're doing a bad job of shooting down the Shaheds. We haven't actually figured out mass precision, and we're wasting all this money and stockpiles and what have you. So, I don't know, it seems pretty far like this. We're not quite in, like, Gulf War One, you know, Gulf War, like, oh, my God, America's like, two decades ahead, but it's not the disaster that this could have been by any means. I don't know, Chris, why don't you take from there?

C (86:23)

I think that there is. There's. Look, there's still, like, multiple factors that influence those things. I think that, you know, the things that you outline, for instance, the fact that we don't have that many people that died. I mean, yes, some of that's the kind of cyber attacks, and then some of that is just the complete decimation of Iran's A2AD infrastructure. And then we get, you know, complete air dominance, at which point, you know, if they're not shooting missiles at the planes and all we're doing is flying planes, like, it's a little bit hard to have people that are. There's a. Kuwait is shooting down more planes than Iran is. So that's literally true. You know, I think that's some of the rationale. There could be an AI element, but not necessarily. Yeah, and then you look, you highlighted this kind of fundamental asymmetry on the interceptor to drone problem. But again, that's Something that's going in. I think that that's just kind of fundamental dynamic going in. I think the way to evaluate it would really be something we don't have a ton of data on now. It's, it's how, you know, how many strikes are we conducting, how effect is really how effectively are our strikes degrading Iran's missile drone production capabilities relative to the counterfactual of if it was all being done by hand because presumably, you know, from all intents and reports were effectuating kind of thousands of strikes very quickly, probably much more rapidly selecting targets than we were before. There's things that would be, you know, previously taking human beings days, that are now taking hours or minutes. And if, if, if that is very, very high accuracy and precision, then it's possible that our degradation campaign is actually substantially more effective than it would have been before. Whereas, yes, maybe since the 90s we've had some precision strike capability, but we haven't had the intelligence fusion that you really need that at wide scale. So like, yes, you can hit a target when you know where it is, but when all of a sudden, not only can you hit a target when you know where it is, but you also know where, exactly where every single target is on a minute by minute basis. That's a fundamental game changer.

B (88:23)

So I don't think we actually know

C (88:25)

the answer to that quite yet. I think there's, there's certainly some reason to believe that, that we, you know, we have been quite effective at that. I mean, at the very least we're expending a ton of munitions. Presumably we're expending them for some reason. Now look, it also has to be said that some of those munitions are not hitting their target. It does seem like that one of them hit a school. Like obviously this is under debate, but like that's an absolutely catastrophic outcome that I actually think is going to have a much bigger impact. There's going to be questions on, like, was that, was that recommended by AI? And then what does that mean? Like, that is something that I think is going to play out in terms of it's going to be in a little bit the aftermath of the conflict. But like, I think that's going to be a big part of this kind of broader, like what is the role of AI and war debate. But like, look, we can't evaluate the counterfactual, but I think it's very possible that actually we are using AI to give us a substantially more powered air campaign and more effective air campaign than we'd had previously. Does that mean that all of a sudden you can effectuate regime change with an air campaign? Like, no, not necessarily, but you might be able to substantially more degrade someone's capabilities for a much longer amount of time. So jury's out. But I think it's possible.

B (89:35)

Yeah, like if you're only limitation.

A (89:37)

Chris, you think Trump's going to blame the school on Dario? Okay, got it.

C (89:42)

Sorry for not.

B (89:44)

No, you had a serious question. I'm glad that I weighed.

C (89:47)

You can keep that in. But seriously, I do think that is going to, that debate is going to rise up. Like eventually there is going to be a DoD report that says what happened, I don't know anything. But obviously we're starting to see indications that inside DoD, basically everyone's saying, yeah, we thought there's independent reporting of that. And then if that was recommended as a target by AI, which like, I don't know, I have no idea, but it seems possible just given a lot of AI is recommending a lot of targets, then like what does that mean now? Like look, AI's recommending a lot of targets and yeah, things happen in war and like that's horrible. But, and, but like maybe people decide like this act, this technology actually isn't quite ready for prime time yet. Maybe that's a fair decision, maybe it's not. But I think we're going to be having those debates if there is an AI nexus of this or would the

A (90:38)

human have lead you to blow up three schools instead of one or something?

C (90:42)

Totally. Right. It's like the Waymo hitting the kid that actually saved the kid's life, right?

B (90:46)

Yeah, well, there's like, you know, the US made human decisions to bomb the Chinese embassy in Belgrade during the Bosnia Herzogena conflict. And we know that there are incidents of friendly fire as well as collateral civilian deaths throughout the global war on terror. So like that, that, that, that was all human beings. Right. And so what will be an interesting and calm comparison that probably will not be possible because of the Internet is comparing sort of the relative improvements to sortie rates. Let's say, let's say the only limiting factor for the US right now in the current conflict has been how quickly can you get planes back and get them back in the air to sortie? Because all of the decision support utilization of AI is enabling that sort of target search find fix that Chris outlined. So you're not wasting a ton of time needing to have the kind of intel support that you would normally have in this type of conflict, particularly as like, the target list probably dwindles radically over time, given the number of munitions used. How do you kind of compare those advantages to disadvantages if we see higher error rates, missed targets, that sort of thing, which, again, I don't think we really have indications that the current air campaign is kind of out of the norm with. At past air campaigns where the U.S. or other countries have committed these types of errors. Because that, that sort of is a thing that happens in war, unfortunately. Making that kind of comparison, I think will be really difficult. But I suspect that if, you know, you're working at the Pentagon on this conflict and you work in the emerging technology space, like, this is something that you're going to be looking at really closely over the course of the next year or so.

C (92:19)

Yeah. And look, I just. I just want to caveat too, that, like, I don't know, thinking about this problem, like, as, like a father, has also made me, like, very uncomfortable, like, this issue of the school. So I don't want to make light of it at all. Like, it's actually, like, deeply sad and concerning at the same time. Like, you know, this is. This is war. And like, the important evaluation is the counterfactual and also the initial decision to enter into the conflict. You know, that's. I don't want to make light of it at all, but, like, we have to also have a little bit of perspective. At the same time.

A (92:55)

I want to come back to this sort of like. Like, to what extent does this, like, confer a perception of American battlefield prowess? And I wouldn't necessarily put the bar as, like, you create regime change in a week, but the fact that the Strait of Hormuz is still closed, the fact that, like, there are still missiles that are going, and even just the visuals of the Shaheeds, like, hitting random Dubai airports, is something that, even though, you know, we are having this, like, pretty remarkable performance, I feel like the, like, the standards of what is expected. If we're comparing, I mean, Gulf War, like, you. You walked in expecting 50,000 casualties, right? That was like the. That was like the. What they were kind of briefing to the press. And then you have this, like, big surprise on the upside of, like, oh, this was actually way easier than we. Than we thought. And I feel like for the sort of narrative coming out of this, like, very much a war of choice to be, wow, look how amazing America is. It probably would have had to be something more kind of dramatic and overwhelming than what we've been able to deliver over the past two weeks.

B (94:10)

Maybe it was Impossible. Right. I think like the, we're sort of talking about the kind of like, operational effects and amazing capabilities of the US Military, which again, clearly on display is this is the only country that can do this type of thing still in the world, I think halfway around the world. And then there's the strategic realities of oil prices, domestic political popularity, allies feeling out of the loop, adversaries watching to learn like, oh, like X amount of economic pressure may make a president change his mind. I mean, those are all relevant lessons as well that you take away. Similar, similar to the costs imposed on Russia for the Ukraine conflict. Right. A lot of the, the initial Russian gains and a lot of the sort of speaking of Putin to his own population have really been focused on some of these really sort of local and operational accomplishments. But if you look at the kind of, like, arc of history as it relates to Russia, its economic state and its ability to sort of consider a possible major power conflict in the future, I mean, all of those indicators are sort of have trended downward pretty rapidly. So, yeah, I think that the, the focus on the kind of operational and utilization of AI in the conflict, for example, is just like one really tiny piece of this. But the impact globally is going to be very different. Right. And I don't think that you, I don't know what level of success beyond the success that's already been had by the United States and Israel, you would have to have to like, overcome all of that stuff.

A (95:34)

Well, Chris, I mean, we were doing these shows about, you know, US China, economic escalatory dominance in the fall. Right. And like, like the fact that Iran has it over the US now, potentially, where like, oil prices started spiking. And then Trump went out and said, don't worry, we're gonna end this war, like, in a few days.

B (95:53)

I'm Elizabeth. Remarkable is Chris common measure during that. I feel like that was a difficult portion period of time for him.

C (96:02)

No comment either. So, like, look, I think something to consider here is like, we're still fighting what you actually have, you know, in, in theater or, you know, the actual kind of physical realities of what you're fighting the fight with really matter. They will always matter, and they matter now. And like, what we're, what we are right now, we're at a place where we have, like, very powerful software that is enabling, you know, decision making that we never could have made before. You know, intelligence fusion efforts that, that we never could have made before, presumably. But we, we haven't seen like, the manifestation on the, the kind of hardware side yet. Right. So like we're still fighting with, with weap designed for the pre air, right? That's exactly why you see that we're using Patriots to shoot down shaheds, right? Like that is a weapon system that was designed to shoot down fighter jets and now we're just forced to use it for the new reality of these like cheaper, more capable systems that we haven't yet had the, the catch up to have the defensive systems that are at the kind of equivalent cost parity like something like that will happen at some point whether it's the kind of more efficient interceptors, obviously Ukrainians have have a bunch on this front, whether it's directed energy systems, whether it's just like much lower cost kill vehicles, etc. But we will adapt there too. So we're fighting at a point when we have this really powerful software, but still the legacy hardware. I think that is why we're in the situation we're in. And look, fundamentally yes, you are going to be limited by that. There is a bunch of different moves and counter moves that our adversaries have that are completely separate from the AI space still. And I don't think anyone would say like AI just allows you to counter everything across the board, no matter what, all the time. Maybe some people would say that in the future, but it's certainly not the case right now. And it doesn't, you know, did we have a plan for countering the kind of closure of the strait? Like I don't know, it doesn't, you

B (98:00)

know, I don't know.

C (98:02)

It doesn't. Yeah. If you do, you probably should tell people. So yeah, you need to do a

B (98:08)

plan for anticipating the closure of the straight arm.

C (98:11)

You need to plan for these things. You can't just put into Claude like how do I do this? And it spits it out and solves it. Right. So yeah, we're not at that point.

B (98:18)

Yeah, there is a little bit of like I try to think about this in, in terms of like how I've used AI chatbots for like helping with writing or research. Right. Where like if you're a PhD level person who's using the tool as your kind of half of a research assistant to help you find primary sources. But because you understand the subject matter well, you know, what are the critical questions to ask? It can be really great, right. And you feel like you have some semblance of control over the types of information you're getting because you've sort of drawn a specific box and ask questions within that box. The type of thing you're trying to accomplish if you just say, if you ask the wrong question or provide the wrong prompt, how do I, how do I destroy every air defense battery in Iran most efficiently with the, the force, the forces that I have in the region? It's probably not going to answer the question of like, here's what could happen after. Some models might, some probably won't. And so if you're not asking the right questions or prompts, which again, I don't know how exactly this capability was used in the current conflict, then you're not going to get, you might get correct answers to your prompts, but you're not going to get correct answers to the prompts that you didn't provide.

A (99:25)

I think it comes back to the hardware versus software thing that Chris laid up, because what the model is going to say is you need some minesweepers and you need some cheap drone interceptors. And AI cannot yet conjure that. Right. That is a multi year thing. And it's not only the fault of the Trump administration that they didn't have all these capabilities going into this war in 2026, but the fact of the matter is it's like, yeah, you still need the things to blow up the other things. And that's not something that software can do yet.

B (100:01)

Yeah, right. Until we, until that prompt actually can then turn around and say, like, and I've just constructed this factory for you to have those drone interceptors that you never thought you needed before. We're pretty far.

C (100:13)

I've hacked all the drones. They're all dead.

B (100:16)

Yeah, right. We can dream.

A (100:19)

Can we close maybe on. So I just ran this essay contest about kind of like economic security and was like, really impressed by the quality of answers I got back. I kind of want to like, do more of these, throw them out into the universe. I think Chris had a prompt earlier about sort of less like Cold War nuclear lessons for contemporary AI problems. You know, if we were going to do the inaugural war talk essay contests, like, what were the. What would be other. The two or three other questions you'd want to, to throw out into the ether and see what came back.

B (100:56)

I can start Chris with one. Yeah, go ahead. I mean, one question I had, and I kind of alluded to it before because the US has essentially said, and other countries have agreed, that there needs to be a human in the loop for nuclear employment decision making. I want to better understand the dynamics of what if. What if that policy shifted to human. Amelie Loop. How would that actually change things from a Deterrence perspective, what would that mean for the sort of a theater nuclear conflict and conventional deterrence dynamics under a nuclear shadow in a region? I think there's. I don't think that's been explored because basically everyone agrees that we shouldn't have autonomous nuclear weapons just flying around. But again, you can envision a scenario in which a US President or another president of a nuclear weapons state says, okay, like, start the nuclear war. I'm giving my authorization to do so. And then you have systems that are making autonomous decisions about finding targets, striking them, et cetera. And so I'd want to better understand the deterrence dynamics there. If the US or another country introduced that type of capability, what would the other side, an adversary of that conflict, think about and sort of like, how are we putting our understanding of nuclear deterrence in a blender on the kind of control side of things, as Chris alluded to? I think there's a lot of work to be done on. Is there a systematic way in which countries that are probably exporters of AI capabilities, whether it's kind of foundational technologies, industrial capability, etc. To develop AI in foreign countries, is do they agree on a set of controls on specific things relevant to AI advancement in the same way that these countries agreed upon limiting missile technologies at the end of the Cold War, or nuclear technologies, as Chris alluded to very early on, after the atom was first split and forming an IAEA safeguards mechanism or what have you that could follow from that? But what's an actual intellectual, intellectual and international architecture that makes sense, given that AI adoption is becoming very widespread and we can't just sort of ban the killer robots. So those are two ideas for essay contests. And I, in part want to read those essays, so that's why I'm suggesting them.

C (103:17)

Yeah, I would definitely second. Second those. The second one. I think that there's. There's a lot of important thinking there and we just, we just have. We just have to think, think through that. I think another thing is just like, how should we. I feel like on the China front, there's a lot of. I feel like a lot of people are either who think about this, are either in two camps of just like, we have to negotiate with the Chinese and make sure that we as the two superpowers, to make sure we have an agreement where it's safe or reliable and no accidents using AI, et cetera, et cetera, etc. And then there's some people that are like just to never do it, never going to happen, forget it. But like Is there a middle ground? And what. How do we get to that middle ground? I think is interesting. I'll give you my personal opinion. I don't think that. I think the Chinese still view arms control generally, and I think they would lump in anything with emerging tech as arms control, with the, the possible exception of that one statement that we got as the thing that the United States did to the Soviet Union to cause the Soviet Union to lose the Cold War. Therefore extremely skeptical of it. And it's very difficult to have good conversations. And much of that is because they never went through a Cuban missile crisis themselves. So there's not the same kind of visceral understanding of the strategic responsibilities of using really advanced technology in a conflict. But, but what are ways that we get them to the negotiating table absent accusing a Cuban missile crisis? I think that's something that we probably have to think through. I would say my personal opinion is like, be so far ahead that the Chinese actually see it as in their interest at bringing us to the table because they're worried that we are so far ahead and only going to get farther ahead and the world is going to get worse and worse for them, which is like a very real politic way of looking at it. But are there other, are there other ways? I think would be interesting. I think that like, both alternatives of just like we need to negotiate something now in good faith and we can never do anything are probably not realistic end states, but I think we do more than it was in the middle.

A (105:40)

You know, coming back to the like, China hasn't had a Cuban missile crisis thing. There's, you know, I've watched a lot of like, not great military related content coming out of like Chinese movies and television and, and yeah, nuclear war being a scary thing is like not a theme I have seen explored at all. And I'm gonna spend a little time looking at the kind of reviews and discourse around the Netflix movie because I feel like that's the closest window we will maybe have into this. But the fact that you haven't had War Games and Terminator 2 and this be like a, a, a theme which is like deeply explored in the domestic cultural context, I think also kind of leads to a little bit of this sort of thing feeling less real than maybe it could or should be for the broader kind of political. And you know, and I'll just say

C (106:40)

like, look on the Russian side, obviously, you know, everything with Russia notwithstanding, and, and you know, their, their horrible actions, like you know, not at all yada

B (106:50)

Yada, his

C (106:53)

in 2012 no. And I'd be like, we're in a fundamentally different world now and obviously, like, negotiating arms control agreements with the Russians is not something that we should be. We are doing at the moment, but should be doing. But there is that understanding, though, like, you can see it. Like, Pranay and I have both sat on the other side of the table from the Russian nuclear negotiators and like, yeah, they'll read their statement from Moscow that's, you know, the US has wronged us in all these ways, but then they'll put it aside and have a real conversation with you about, like, the things that we have to do on safety and responsibility and, you know, preventing accidental use. Because, like, there is genuine interest and belief in large parts of their system that, like, this could end very, very badly for them if they don't, you know, have some degree of responsibility. Like, that is actually a deeply felt feeling among a lot of people, not everyone, and not necessarily President Putin, which is the whole problem right now. But like, like in many people in the Russian system, in the military, in the Ministry of foreign affairs, etc. So like, you know, that does exist there and it might not be rising the top, but it does exist. I haven't just seen evidence that it exists anywhere to the same degree in China. And like that. That is a stark difference that applies not just to nuclear, but kind of a lot of other technological issues.

B (108:12)

Yeah, there's like a. I think to the point that Chris is making, it feels like in conversations we've had with Russian counterparts over our careers, there's at least agreement on the types of risks that emanate from the presence of nuclear weapons and the potential use of nuclear weapons or the loss of nuclear weapons. And what we see right now because of politics and geopolitics is a disagreement over whether those risks should be manipulated and how they should be manipulated.

A (108:35)

Right.

B (108:36)

I don't think that we can say the same about the baseline between the United States and China because we just have not had that shared experience, 60 plus years of negotiations, et cetera. And so there might be some fundamental disagreements on the types of risks that emanate from nuclear war as well. And frankly, from the Chinese perspective, they just think that the US Wants to mitigate risk because risks are bad for the US because we're less risk tolerant. And so if you're China, why not sort of increase risk or be opaque or do all the things that the United States wants to mitigate instead? And a lot of the sort of conversations I participated in Whether they're war games or actual work or research for articles. I mean, the sort of tendency of China to want to end communications at the moment of crisis has been sort of common theme.

A (109:21)

Right.

B (109:21)

And we observed that over the course of our last positions in the Biden administration as well. It's just a different language and it takes time to come to grips with that. And maybe AI and space and other areas where the US and China are much more kind of not equivalent, but sort of, I think in the same competitive category as sort of being world leaders might be easier places to have conversations and then talk about how nuclear weapons inject into those topics as opposed to trying to just go hard at like, hey, freeze your arsenal at 1,000 and then start to reduce it and then maybe we can be friendly again. It's just that approach has not worked and probably will not work moving forward.

A (110:01)

All right, we're almost at two hours. I think we can end it here. What a wartok debut. Thanks for bringing the takes.

B (110:10)

You got War and Talk.

A (110:15)

I learned a lot. This was. This is really wonderful. Oh, sorry, Pranay, I had one final question for you. So, you know, we're at mit, we're at Sandia. We just assigned a lot of folks some nuclear homework. What is the type of work that you can't do without super secret clearances? And what is the type of work where you totally can, and what's your perspective on people without the clearances where they may have kind of unique advantages in contributing to these sorts of dialogues?

B (110:47)

Yeah, I actually think this is such a sort of. We're at the sort of building the foundation of our understanding for AI and nuclear issues. The, that a ton of the really good work can actually be done in the academic space. When we are talking, we're talking about how can we understand the impact of artificial intelligence on nuclear deterrence theory, which, by the way, is something that's already being examined very closely now because most of our nuclear deterrence theory is derived from a world in which the United States and the Soviet Union were the only sort of superpower competitors. And now we're thinking about, well, what if it's China? What if it's a. A larger North Korea arsenal in 10 years, France is increasing their arsenal as well. This sort of multipolarity doesn't necessarily change all of the fundamentals that we've come to research and understand in nuclear deterrence, but it changes some of them. So while this examination is under place, enter AI, right? Enter emerging technology. And I think that given the Kind of academic and theoretical scope of that, of this, that space of study right now that's being led by political scientists, economists, historians, technologists, you name it. You don't actually need to be in government with a clearance to do really meaningful work there because we need a, we need sort of an architecture by which we think about these issues right now. We don't necessarily need to think that in that much detail about the sort of operational command and control actions that will be taken, or at least we don't need to sort of know the details to opine on that stuff. I think internally, for people who work in government who do have clearances are all thinking about the more operational features like how would I actually use artificial intelligence to enhance decision support for a president if he were in that sort of meeting thinking about nuclear weapons employment. All the people who have tried to make sure that our NC3 system is as modern, capable and supportive of, possible for a decision that is completely unimaginable for any president to make. They're going to sort of think of ways that artificial intelligence can improve this work. But again, it's been such a human centric and political decision for decades and decades and decades and I fundamentally don't see anyone being comfortable changing that wholesale. I do think that they'll look for ways for AI to influence that though. So I actually think that this area is ripe for NGO and academic discussions. But as, as we talked about today, there's, there's a lot of debate that maybe is not focused on some of the, the issues that we focused on, which is kind of like, like how would this actually play out in a conflict? How does it actually affect the motivations, interests and stick to itiveness of belligerence in a conflict that deserves a lot more study, which does require I think a little bit of focus on kind of military operations and deterrence theory. But again, people outside of government can do this. And frankly with nuclear weapons, as Chris mentioned, throughout the Cold War, the people who were doing this work were a lot of scholars on the outside. They didn't necessarily need all the access to all of the super secret details to do meaningful work that we still lean on today.

C (113:41)

Very much agree. I think that even more so now, I think that this is, A, the conversations are more open and B, the technology is something that you can acknowledge, at least you can literally play with it on your own computer. I think it actually makes it much more accessible for those kind of public conversations. So I totally agree with that. Just one note, while we were talking, we Obviously talked a little about the kind of the school incident but just generally how AI can, you know, how it will influence warfare. But while we were talking near Times published a fairly long report on the school strike. And it does say all indications are it was an accidental US strike. But it does say officials conducting the investigation examined whether AI models were responsible for the strike on the school. And it said the officials concluded that is much more likely the result of human error and is unlikely the result of new technology. So just interesting that yeah obviously it's part of the debate but like maybe there's a way this ends up is actually more use of this would have actually prevented the catastrophic outcome. Like you know, I don't want to render judgment without before all the facts are in. Not to quote Dr. Strangelove again but it does seem that there is actually there's some positive benefits here that we're seeing in the real world on how it could actually improve prove the humanity of conflict in a perverse way.

A (115:12)

I wouldn't be shocked if that was actually the takeaway from this war. Obviously it's going to be the takeaway with the mass precision stuff. But also on the AI side that you thought you were leaning in. You thought you had your Ukraine lessons learned. No, you didn't. The wars actually changed pretty fundamentally and that hasn't quite quite flowed through all the acquisitions and force structure things that you need to do to kind of grapple with the gravity of changes and kind of new opportunities that are put forward for militaries to spend in more effective efficient ways. All right, now we are literally at the two hour mark and I think that means we all need to eat some lunch. Okay, thanks guys.

C (116:01)

This was great.

A (116:01)

I'll pranay you volunteered yourself to judge the essay contest. So happy to help.

B (116:09)

Happy to help.

D (116:15)

There's a minefield starting in the straight reaching big density, shutting down trade. The blast from your love remind me of us. It keeps me thinking that we almost cleared it all. The blast from your nuke it leaves me breathless. I can't help feeling we could have cleared it all. Blowing in the deep minds are going to fall. Blowing in the deep. You had a straight inside of your hand and you looped it to the be baby. I've got 3,000 warheads to deploy But I've done the math now and everyone's a decoy. Think of me in the depths of your despair. Irradiated water now nobody sailing there. The blast from your love remind me of us. It keeps me thinking that we almost cleared it all the blast from your nuke it leaves me breathless I can't help feeling we going to clear it all. The straight inside of your hand and you n de to the be

A (117:50)

Throw

D (117:50)

your dolphins through every open lane Count your and do it all again Turn your Davy crockets in a scrap the be sabbath was right there on the

C (118:04)

ma

D (118:07)

yeah we could have cleared it

A (118:09)

all

D (118:12)

glowing in the deep

B (118:16)

you had

D (118:16)

the straight inside of your head but you nuked it with a day we could have cleared it all. You had the strange right inside of your head but you played it, you played it, you played it, you blew.

B (118:57)

Powerful doesn't just happen, you have to make it happen so the moment Total Wireless offers a free Samsung S25FE with Galaxy AI. When you switch to total 5G or 5G plus unlimited 3 month plan, you take the network as powerful as you with unlimited 5G data that won't slow you down. Now that's a total power move.

A (119:16)

Visit your neighborhood Total Wireless Store device taxes and fees may apply. Requires new activation on a total 5G

B (119:22)

unlimited 3 month plan or higher. External port in and ID verification available only in total wireless stores.

A (119:26)

Limit to 4 devices per account.