← The AI Policy Podcast
The AI Policy Podcast
Inside Project Maven and AI-Powered Warfare with Katrina Manson
01:01:29
In this special episode, we sit down with Katrina Manson, author of Project Maven: A Marine Colonel, His Team, and the Dawn of AI Warfare.
Loading summary
Transcript76 lines
- [00:00]
A
Foreign.
- [00:11]
B
Welcome back to the AI Policy Podcast. I'm Gregory Allen, and today we've got something that is a genuine delight and special moment for me. We're talking with Katrina Manson, who is the author of the new book Project, a Marine Colonel, His Team and the dawn of AI Warfare. Now, I'm just going to skip right to the end, which is if you listen to this podcast and you do not buy this book, I think less of you. Moreover, if you want to talk about military AI in the year 2026 and you either have not worked for the military or you have not worked for a military contractor working on military AI and you have not read this book, I'm taking away your permission to talk about military A and claim that you're doing so intelligently. If this book does not win Financial Times Best Book of the Year or one of the finalists for Best Book of the Year, I will not think less of this book. I will think less of the Financial Times. So with that, Katrina Manson, thank you so much for coming on the podcast.
- [01:17]
A
Thank you for having me. I guess you read it then. That's good.
- [01:22]
B
I did, I did. I would have read it in one sitting, but I have children and they're feisty. But. Okay, so let's just start a little bit at the beginning, which is with you. So you're a journalist. Who are you? How'd you get interested in this story?
- [01:39]
A
I am a journalist. I am with Bloomberg now for the last three, four years covering AI and national security, cyber tech, emerging tech, defense, those kinds of things. Before that, I was with the financial times for 11 years. I was the US foreign policy and defense correspondent. So I was a Pentagon correspondent from 2017 to 2020 and covering the intelligence community and State Department and National Security Council. So I got to kind of see that Nexus. Before that I was based in Kenya and I spent five years as East Africa correspondent for the Financial Times. And before that I was writing in Congo as a Reuters correspondent in Sierra Leone and Burkina Faso. So I got to America eventually.
- [02:29]
B
Well, wonderful. And we're glad we did. So, as you say in the subtitle, this, this book is about the dawn of AI warfare. So how'd you get interested in this story? What made you decide it was worth a book?
- [02:43]
A
I think I just want to know what's happening next. And when I found out there was this project about the future of war and this claim that was made that it would involve AI and then that it shut down and couldn't be foyered, which is Freedom of Information Act. So there was no obligation for the Pentagon to respond to journalist requests. That made me from that moment on think, well, what's happening and why do they care so much? And why do the Google workers care so much? Who are protesting against this? What does it actually mean? And I became frustrated with the debate because the debate about AI warfare is fascinating, important, powerful. So is the debate about how you protect military operators at war and civilians. So there is so much to care about. Add in this unprotected, predictable black box technology. I just wanted to understand what everyone was thinking. And anyone who is animated by passions, as I discovered everyone to do with Project Maven is animated by passion, whether they're for it or against it, really. I wanted to get under the skin and understand their motivations, one part. And the second part was, well, what has it done? What are the concrete uses of AI in warfare and do they work or do they go wrong? And nothing really about Project Maven as much as has been written about it talks about the actual quality of the algorithm, the way it was put into the workflow, whether operators wanted it or didn't. And that really.
- [04:17]
B
And how it's changed over time.
- [04:19]
A
Sorry.
- [04:20]
B
And how it's changed over time.
- [04:21]
A
Yes, yes. And where it's got to. And you know those things that the campaigners were worried about at the beginning, they said, you know, Google protesters said, we don't want to be involved in the business of war. Okay, that's a kind of ethical position. But the human rights campaigners who were really concerned about Project Maven said, we think bringing in AI at this point could help lead to autonomous lethal weapon systems, which is not what Maven said it was. And so I wanted to pull on that thread and see were they right to be worried about that? Were they over egging it? Did people inside the Pentagon always plan that? And so I think a lot of it for me was also about a little bit of accountability and transparency. If I could get there. It was not simple.
- [05:06]
B
So, as you said, and as I've intimated, this really does go from the dawn of Project Maven, including some of the Eureka moments that gave some of the key protagonists in the story the desire for something like Project Maven through its years of struggle all the way to middle of 2025, where it is easily the most impactful AI capability in the US military. One source you interviewed described Maven smart system, the sort of current incarnation of where we are in the story as quote, the Microsoft Windows of warfighting, which I think is a lovely quote So I want to give our audience, I want to start our discussion by giving the audience a sense of just how impactful Maven has been. And I think you've already done the perfect job of it, which is on page seven of your book. So could I actually invite you now to read from your book this passage on page 7?
- [06:05]
A
Sure. Ten years since Kukor started his effort, the AI decision making systems developed under Maven and some of the Pentagon's 800 other AI projects are used on the battlefield. Maven Smart System MSS, a software platform that develops targets with the help of AI, is now deployed in every branch of the US military and all over the world, incorporating more than 150 data feeds and the work of more than 50 companies. NATO started using a version of the system in the spring of 2025. And I would learn in October 2025 that 10 NATO members were lining up to use it for their own militaries. Maven has already sped up the pace of war. I learned from an official at the National Geospatial Intelligence Agency that with the help of computer vision, the US went from being able to hit under 100 targets to being able to hit 1000. In combination with large language models, LLMs, integrated into the Maven platform, that number has risen fivefold, to 5,000 targets a day. The AI algorithms developed under Maven now deploy in submarines and in space operations. They're in subsea sonar systems belonging to America and two of its closest intelligence allies, the US and Australia. Designed for nuclear deterrence, they're fielded on autonomous drone boats. I learned the AI targeting systems live in at least two highly secretive systems, one aerial and one aquatic, that could surveil, select and kill targets entirely on their own, intended for the defense of Taiwan.
- [07:49]
B
Okay, so anyone who heard that and doesn't understand why this is a vitally important story to tell is clearly clueless. That's where Maven ended up in the middle of 2025. And actually, at the end of this, I do want to talk about where we are in terms of the war in Iran. But its origins, which you tell in this book, are much more humble. It begins with the trauma of one Marine intelligence officer, Drew Cukor, serving in Afghanistan in a series of deployments that began only two months after September 11th. So who is Drew Cukor and what was it about his experience in Afghanistan that planted the seeds for what would ultimately become Project Maven?
- [08:28]
A
He's the Marine colonel who is chief of Project Maven from its inception until late 21. I think I've got that right. Maybe early 22. And he is really described to me by many people before I ever meet him as the driving force of Maven and also something of a difficult boss in that he is very.
- [08:49]
B
Yeah, you describe him as a noun, a verb and an adjective all in one, which is amazing. Just because of his personality.
- [08:55]
A
Many, many people talk about getting C cored and to CU Cor and some points, you know, if you are working hard, sleeping not very much, and being very exacting of everyone who works for you, that certainly was one way he was described. So he is sent into Afghanistan in October 2001 after 911 as the first set of expeditionary Marines. And he sets up and is carrying this hefty computer along with. It's actually Dave Sperk, his colleague, who is carrying it. And the two of them are part of the intelligence cell for those early operations. And he's very angry and frustrated by the inability to really bring intelligence to the front lines. So information exists to some extent, maybe written down about past Taliban hideouts that the US might have known. Of course, the US was very underpowered in terms of intelligence at the beginning of that war. And some would argue, you know, throughout, just because it was so different. It was a very different place to operate in different language, all sorts of different community networks that Flynn later makes extremely public. And it unleashes this evisceration of the failure of US intelligence. And Drew Cukor is really one of those people operating in this intelligence vacuum as an intelligence officer, and he simply just wants data. He wants to the Americans start facing terrible improvised explosive device attacks, and he wants to start logging where they've been, how often they come, who makes them, under what weather conditions, and all this sort of information. Eventually he gets to meet Palantir and he works with them and the US Military bureaucracy to get that Palantir system delivered to the Marines in 2011. So a lot later by then, he's gone up the Pentagon bureaucracy. And this is a system that he says argues save lives almost overnight and becomes a hit because the previous systems which exist. It's not that none of this has been thought about, but the previous data integration systems are shonky. Maybe they just simply don't work very well. The data feeds aren't working, the access to them isn't there. And of course, we all, I'm sure your listeners at least all know the big fight that Palantir has with the army over trying to become a program of recording. So he is really part of that effort to get Palantir to US Troops from the outset.
- [11:41]
B
And that initial experience of his, you know, bringing software and modern software engineering techniques forward, deployed with end users who can complain and actually have a company take that feedback and make changes to try and make the software better, you know, contrasts with his experience basically using Microsoft Word and Microsoft Excel to log intelligence information and having terrible analytic tools, having terrible dissemination tools. And his experience of bringing technology and improving intelligence is a formative one for him that he then brings to later in his life as he's thinking about artificial intelligence. So I want to now jump to the, you know, that's Cukor solving one flavor of the intelligence problem. But if you fast forward to later era of those wars, there's a different flavor of intelligence problem, which is about having too much data compared with the analytical capacity. And this is especially acute in the challenge of processing, exploiting, and disseminating information. So can you talk a little bit about that flavor of intelligence challenge and what was going on at the time?
- [12:50]
A
This really becomes the launching pad for Project Maven. This is the pilot project, if you will, that most people are familiar with. Project Maven ends up being much more than this. But it starts with this problem that they believe they can tackle using AI, which is that the drones the US has unleashed in the global war on terror, the gwat, are collecting information that is not actually being looked at. They're running so many videos the entire time, and they have screeners looking at it, or not looking at it, as the case may be. And there were complaints as high military, senior military officials would go out to Iraq and elsewhere and. And realize that no one was actually then observing, analyzing, taking this data from the drone feeds for any kind of
- [13:36]
B
actionable intelligence, because they're collecting years and years and years worth of video footage per hour or whatever the case may be. And so that the number. The analyst community is just overwhelmed by what it is. And so Cukor and I guess also Will Roper have this insight that maybe AI could be helpful in solving this problem, sort of what. What sparks that insight for them.
- [14:02]
A
So Will Roper already has a project by the time we kind of join him in the story, looking at satellite imagery and trying to use AI on satellite imagery. Ku Kluk comes and that gets through Congress and gets called Project Maven, and he gets some money for that. Ku Klu has been attending these breakfasts in the Pentagon that really reignite the kind of Cold War effort to try and get one up on Russia. This time it's reignited Under Bob Work, then Defense Deputy Defence Secretary, at how the US can catch up with what it perceives as a risk that it will fall behind China's technology. And so they're trying to come up with ways of knitting together emerging technology, the commercial sector, and what the Pentagon perceives as the China Challenge. And Drew Kukel comes up with this idea to bring AI to drone footage rather than satellite footage at this time. And he pitches it, and the Deputy Defense Secretary loves it. And they then develop over a series of months, a pilot. And very interestingly, Dave Sperk again works with some early algorithmic companies and says, can you detect this? Can we run an algorithm of an image and just say, what's in it? This was already stuff that existed in the commercial world, but they wanted to see if they could do it with military objects for which there was much less data available commercially. But of course, the Pentagon inside had lots of data on these objects. But. But it wasn't in the right place. It wasn't in the right format, and they just didn't know if it could work. And together they take Will Roper's name, Project Maven, they apply it to this.
- [15:41]
B
They take his money too. Don't forget that.
- [15:43]
A
That is in the story. Yes. Kuko was always very pleased about that, he told me. And they go to Congress and they start getting more money as well. And Will Roper was a supporter of Project Maven?
- [15:54]
B
Yeah. Okay. So they have this challenge of basically infinite data, very finite analytical capacity, and it's costing lives effectively in the war on terror. And they say, can we make the life of these analysts easier? And I think the original technical capabilities and even the original technical aspirations were quite humble. Right. Counting the number of people in an image, counting the number of cars in an image, these are pretty humble tasks, but they actually are on the job jar of these analysts. If you can make their life easier, you can increase the analytical capacity of the entire workforce. But as you were just hinting, Kukor really wants to bring leading AI companies, Google, Palantir, Amazon, to work with the Pentagon. This was a time when tech workers were pretty skeptical of the Department of Defense. How do Kukor persuade them to come work with the Department of Defense?
- [16:56]
A
Two things just to pick you up on. Quite humble. The reality was quite humble. But the ambition of Kuko from the outset was enormous. He had this thesis that talked about the need for white dots that you can click on and send as a target. He really had spent years developing this idea. So although analyzing an image just for what's in it. And the effort to do that went kind of almost comedically wrong instantly because it was so hard. They got better at it. His scope, his vision was always for something that could automate or could bring intelligence to the front lines.
- [17:38]
B
Yeah, it's a little bit like Jeff Bezos. He's like, I'm going to start with selling books online, but my dream is to ultimately sell everything online. And this is the logical starting point based on the maturity of the technology and the infrastructure where we are right now. So you're totally right. Right. His ambitions were much broader. So one of Cukor's priorities was getting. Well, actually. Sorry, finish answering. I interrupted you before you got to the second part of answering.
- [18:02]
A
The second part is how did he get the commercial sector on board? So he studied the papers, he looked at who he wanted to.
- [18:10]
B
You mean the AI research papers that were published.
- [18:11]
A
He studied the AI research papers. He was reading about Microsoft. He was doing a lot. It wasn't just him. There was a diu, your listeners will know, Defense Innovation Unit. I got that right. And various other parts. But eventually companies started gathering. They also had a prime contractor, ecs, who was also kind of understanding who was at the cutting edge, but he wanted to go to the cutting edge. And one of the companies Kukul wanted was Google. They used Google Earth already. Whether Google Earth knew it, whether Google knew it or not, they were using Google Earth for their military operations to help, to make that legitimate, to bring that in, to make it as a kind of platform. That was part of the vision. And then he needed really good algorithms. So he wanted DeepMind and he didn't get them. He wanted another part of Google, Google Brain. He didn't get them, but he did get Google Cloud. And then at the same time he was trying to find the algorithm makers. And there was this one startup called Clarify who really was just getting going with an award winning leader who had become very interested in computer vision and had started winning these contests for how many images could identify.
- [19:28]
B
Let me just jump in here, folks on this podcast will have heard about the 2012 ImageNet competition, which was basically the marriage of modern GPU technology with some neural network approaches to computer vision. And they blew everyone out of the water. And that was the aha moment for Silicon Valley that like the time is ripe for an AI revolution was that ImageNet competition. And one of the Clarify founders was on the winning team of that 2012 competition. So when Cukor got him, probably felt pretty good about himself.
- [20:02]
A
Yeah. And it took. It took A few visits. So Matt Zeiler, the boss, he began to waver. He wasn't quite sure if this is what he wanted to do at the time. The way he was making income was he was using computer vision for wedding blogs. So his, his algorithms were really good at identifying a bridal veil or a groom's suit or the tears of a wedding cake. So to repurpose that. And he was based in New York. He knew that some of his staff might not be comfortable with that. He didn't have a background in the military. He was, he told me, you know, he. The most he knew about the military was watching old war movies with his grandpa. He had one friend in the military. He's. He was, grew up in Canada. So it really wasn't part of his. It wasn't on his bingo card. And Cukor made these visits to see him. Zeiler was kind of at least entranced a little bit by just how much of a kind of military cliche Cukor seemed. You know, he was meeting a real life Marine colonel. But they also had these long discussions and Cukor put to him. According to Seiler, Cukor did not confirm this account that he kind of laid out the scenario where US Military operators were somewhere in Africa and had got. Were concerned that they were getting attacked and were potentially getting roughed up. And it turned out it was by cattle. And he put to Zeiler that if they just had AI that could be analyzing and observing what those concerns were, AI would help. And for Matt Zeiler, this was a very compelling potential use case that he then used to put to his staff and say, we can help save lives rather than take lives with AI and this is a direction I want to go in. And he did lose some people over that. And then subsequently, after the Google protests, he had further, even more public protests. But he has continued down this road.
- [21:58]
B
Yeah, and you know, I'll just reveal my bias here. To his credit, has stuck with it and I'm grateful to him. So Google was one important relationship. It obviously blows up with Google now basically exiting the contract and working on Project Maven. At the time, it was seen as, you know, a devastating blow to the military's AI efforts. But ultimately Palantir and other companies, Microsoft, aws come in. So what actually was the aftermath of the Google exiting?
- [22:37]
A
If you ask Kukor, he says everything was fine and all it did was bring other companies coming forward to say we're here to help. I think that is, I do think it's fair to frame it as a big fault line, a big rupture. And the way you trace this history is you look at the reaction. The claim that AI was needed to face off against China becomes a much more substantial part of the argument from here on in. You see in Congress a discussion, a public discussion for public consumption, that AI is needed, and if China has AI, the US Will fail in any potential conflict. You see the emergence of Eric Schmidt's group, the attempt to really try and analyze what role AI can play in war at a time that publicly, it seems quite unpalatable. You also have this outreach to and interest from AWS and Microsoft. And then Kukor himself rings Palantir and says, hey, you got a minute? And he goes and pitches what would have been his Google Earth for war to Palantir, which is not necessarily on board at the time they want the contracts, but they don't believe in AI at the time that he pitches it. And they also don't see themselves as a user interface company. And in the book, I report that some of them were put out by this idea. They want to do the data management, the data integration. They don't want to be a fancy user face. And also that's not what AI is. So Kukul begins to encounter problems, even from his own team, about the idea of trying to create a kind of everything app, as it were, rather than develop the algorithms to be good enough, which at that point really weren't delivering. But he does get on board other parts of the commercial world, but you also get this fracture. So you have Google workers saying, we don't want to be part of this. You have this kind of moral backlash that I think does make the Pentagon very wary. And of course, you have the formation of the Jake, which I want to hear from you about your experience in the Jake. But you have Jack Shanahan, who was director of Project Maven, so Coogor's boss, moving over to then lead the Jake, and he goes on a listening tour. And I've learned enough from enough people in this book, that listening tour means, you know, people who are prepared to also soften the blows, perhaps even apologize. Palantir at one point is counseled to go on a listening tour. They call it a listening tour. Others call it an apology tour. And really, this is Shanahan reaching out to try and temper the water on AI, deliver the public acceptability of AI in warfare. And that's, I think, why one reason why the Jake focuses very much on ethics. The development of this ethics code, the development of Responsible AI, which annoys some people in the Pentagon because they want to go faster.
- [25:32]
B
And the use of AI for non lethal military applications such as warfighter health, predictive maintenance, humanitarian assistance, disaster relief. There was all of these workflows that were very much intended to say, like, look, there's a lot of ways companies can help the Department of Defense other than things that go boom or things that, that, you know, are one step removed from going boom.
- [25:54]
A
Yes, they talk about wildfires as well, but if you look at the statements of Jake and Jack Shanahan, from the very outset, he is talking about designing AI into weapons and saying no weapon should be designed in the future without AI in it. So he somehow was trying to straddle both and moving into this much more operational field, further in public, at least than Maven was. And then meantime, the people in Maven were not focused on ethics in, in any way. They were developing, testing and evaluation, trying very hard to do that. But at the time, they were judging algorithms, sometimes just by eye, by looking at the screen and trying to work out which algorithm was better at identifying things than the others. They then embark on this kind of growing up process of trying to improve that. But Maven was more operational at the time and less focused on the ethics and responsibility. And Jake took on this very public focus to say, we need this because China's going to get it and we will be responsible for it. But there actually wasn't very much integration between the two efforts, you tell me.
- [26:56]
B
I mean, number one, I was the policy guy at the Jake, and that ethics portfolio fell to me. And for me, it was perfectly natural that Maven didn't have a big ethics component because they were in the Intelligence, under the Intelligence Directorate. They're not a policy shop. We were a policy shop. So we wrote ethical guidance that applied to them as opposed to them writing ethical guidance. That app applies to us. And there was some sharing of relevant infrastructure. So, for example, the computer development environment, Sunnet, which was a mechanism for labeling images, that infrastructure, was actually important for our humanitarian assistance and disaster relief efforts, even though it had not its origins in Project Maven, but an important growth phase under Project Maven. So, okay, but one little tweak on that.
- [27:47]
A
So, yes, Maven was Intelligence Directorate, but it was focused on getting it out to operations. And I think that's the thing that palpitate at the time.
- [27:55]
B
Philosophy. Right? Yeah, yeah. And, well, let's talk about that field to learn philosophy, because you have, you know, these stories of Colin Carroll, and I think it's Brian Ward, you know, going off to these far flung places, trying to persuade people to plug in these boxes to try out Project Maven, and oftentimes in the early stages being really disappointed. So how did we go from this story of like, the AI is super broken and junky to like, AI is this world conquering colossus flavor of capability in such a short period of time? Like, what is your take on the field deem to learn hypothesis?
- [28:37]
A
First of all, the jury's still out on whether it is this world conquering capability. And I think there's still a lot of debate inside the Pentagon about how good computer vision really is, for example, and the way in which LLMs should reliably be called upon. So we can maybe get to that later. But the effort to get operators to try this out was highly fraught. Many of the services that Project Maven tried to seduce and say, hey, try out this algorithm didn't want to play. And it ended up falling back on past relationships. Colin Carroll had a previous relationship with a commander in Somalia. And so they were able to, partly through that relationship field. The first algorithm to Somalia, it didn't go very well. That was even in the first year. So they worked incredibly hard to get the algorithm out there, to have some system to integrate it into. And then the operators put it on and it flashed a lot. It was identifying too many things. The boxes were flashing up a lot. And if you imagine a second video footage requires multiple still frames, and if the AI detects it on some frames and not on others, which can happen because it's not consistent, you just get a flickering. And so the operators turned it off. Sometimes the detection boxes, I think at one point they were using blobs rather than boxes and the blobs would obscure what was in someone's hand. So you couldn't even then, you know, there were lots of sort of learning efforts going on during the first two, three years that they could get the software updated much more quickly than hardware. But still operators complained that it wasn't being updated quick enough. And they then relied on sending out people to the operational zones to essentially coach the operators. Hey, please try this out. We've got to think about China. Okay, it doesn't work now. I know that, but we're trying to get it ready. And even Kukor described the AI as a bag of potato chips, which took me quite a lot of time to try and unpack. But I was told in the end it meant, of course the algorithms are no good. We don't care about the algorithms so much as the system that this is going to sit inside the way we're going to change operations. And of course, it's competing with existing systems. So NRO has its own system that does intelligence. Kukor told me that he tried to work with them, but in the end, Maven went faster. But you have to imagine every single system has its competitive spirit, its cheerleaders. It has its own budget line. And so Maven was beginning to, I suppose, threaten, in the view of Project Maven, other efforts. And also it was leaning further towards operations. And Kukor, I report in the book, was advising people to say, don't talk about targeting. You know, that gets into operations. We don't have buy in for that yet. Even though it was in his mind.
- [31:32]
B
Yeah. So at least for me in the book, one of the key stories that you tell that made me think of as like, AI went from mostly doesn't work to is actually delivering some pretty extraordinary capabilities was the story you tell of Project Maven in Ukraine. So when Russia invaded Ukraine in February 2022, Maven had already been around for about five years. What state was the program in at that point? And then what happened in the weeks and months after the invasion when the US military decided to bring Maven to bear in support of Ukraine?
- [32:09]
A
Two things. Bureaucratically. Maven was actually in a place of flux. It was supposed to move to nga. There had been this. This kind of almost catastrophic fight within the department about whether it went to NGA or to the Jake or to the Jake successor, cbao by then. Yeah, and that had been very fraught. What would Congress do? It was meant to go to NGA in the end, but it hadn't happened yet. So it was in a period of transition. And Cukor had left by February 2022 as well. So you had Joe Larson leading it and also Cameron Stanley, who is today's CDE ao. So both of them playing a very prominent role. Now, it had been tried out in Afghanistan during the withdrawal as a means to see what was happening at the airport, just to give greater clarity at a time where obviously the US had very little clarity and things went very wrong. But it was at a moment where they could see Christopher Donahue at the time, General Donahue, walking around on base, and they could try and identify, count people in the crowd. That moment when Kabul airport was really being deluged by people trying to leave. Maven became a tool for working out how many people were there so that people could advocate internally to say, hey, there's a situation developing or already developed. At that point, senior people in the Pentagon started getting accounts to Maven. And so, so by the time it was time for Ukraine, the 18th Airborne Corps, which had been for two years running experiments, exercises with Maven, to try and link up a computer vision detection potentially to a Weapon through link 16 and be able to fire against that weapon. So really a massive shift into quite explicitly AI targeting, the one that Kukor had always envisioned. It was the 18th Airborne Corps that took that forward. They were deployed to Wiesbaden in support of Ukraine. They didn't really know what support to Ukraine would look like. So some of the first efforts were to count people leaving Ukraine to get a sense of refugee flows. Some of them were to try and assess the state of the Russian capabilities and where they were. But very quickly after the war, the invasion started. They started trying to use computer vision to assess, well, where were Russian things that Ukrainians might be able to hit. And. And again, it didn't go well. The algorithms had been trained on the Middle east, they'd been trained on desert Ukraine. There was snow, there were tanks. What they did was they moved in satellites, they took new pictures, and remember, all those tanks were lined up on the road to Kyiv. They took that new data, they fed it back to the algorithm makers, and overnight they would be retraining their models, trying to get them better. And then they developed this sort of. For some, it was a gray area. For others, it was clearly allowed. For others, they were terrified, quite frankly, of breaking their own protocols. But they started developing points of interest that they could share with the Ukrainians. Almost everyone I spoke to at some point slips and calls it a target, but they were making sure that they weren't calling it a legal target. They weren't telling the Ukrainians what to hit. They weren't sharing America's classified intelligence data with the Ukrainians. But they were drawing on all of that at using computer vision to identify where objects were and then share it with the Ukrainians. Now that computer vision improved, really only because it was then cross referenced with signals intelligence and other forms of intelligence. And so they could get, you know, a hold on what things might be. But they got really good in that first year, in their view. They began to spot things that they could get to the Ukrainians. Within a matter of seconds, the Ukrainians could hit. And. And the trust, which is the other big part of what it is to experiment with AI and develop it became such that the Ukrainians, in one case I was told about Couldn't tell what they would be hitting from their own intelligence sources. And the Americans could say, trust us, hit it. And so it might just look like a rectangle or a truck to the Ukrainians, but to the Americans, they were able to see that it was. I think, in this case, yeah, it was a tell. Yeah.
- [36:37]
B
So, I mean, to me, the story that you're telling as somebody with a background in the Department of Defense, you know, what you're saying in kind of like a blase way or a matter of fact way is probably the more accurate phrase, is just like miracle after miracle after miracle. Right. So the traditional story of military technology over the past 20 years is they make a bunch of promises, it never works. The program is canceled. That's the default outcome. Then the next, you know, the next tier up is something finally shows up, but it's super duper broken. And they promise they'll fix it, and then they never do. And then if you get above that, it's like, okay, a few years later, something that's halfway useful shows up. So the alternative story that you've just told, where Maven shows up and initially, because their algorithms have training data from the Middle east, which has Sam as a background and is going after the types of activities that go on in Iraq and Afghanistan, not the types of activities that were going on in Russia, Ukraine, doesn't work very well, but the fact that they were able to collect an enormous amount of training data, the fact that they were able to label an enormous amount of training data, and the fact that they were able to then retrain those algorithms and then redeploy them relatively quickly and ultimately get to something really, really powerful. And you say, you know, mostly when combined with signals intelligence. But I think it's, it's, you know, to me, it. You can't understate how valuable directing the attention of the analyst really is in that first part of the story. Right. Because the satellite image might cover hundreds of square miles or more. Right. And so telling the analysts, look at this spot first is a very, very, very powerful thing. Now, ultimately, you know, the computer vision model may have gotten it right and say, hey, I think this is bad guys. And then the computer analyst, sorry, the imagery analyst looks at it and says, yes, it is, or no, it isn't, or it probably is, but I'd love to have other data points to corroborate this information. You know, you're shrinking the workflow of assessing what is actually out there by half, by three quarters, by 90%. This is a Huge, huge increase in productivity. And you're right, it wouldn't be that valuable if there weren't other parts of the system that are important to corroborate what the AI is doing. But as sort of a first directing of the attention of the analyst, it's really powerful. And that's why when you wrote that the peak in terms of passing targets to The Ukrainians was 267 in a single day. That is like an unprecedented story in intelligence sharing. You know, other, other military campaigns where the US has tried to support allies don't look like that. And you know, just based on the, the, the feedback that you cite from Ukrainian officers in terms of their own sense of, of just how good the support they were getting from the Americans is, to me it looks like something really special was going on.
- [39:49]
A
I think it's really interesting because it slows down in subsequent years.
- [39:53]
B
But you make it kind of sound like a political story as opposed to a technology story.
- [39:57]
A
I think it is, I think it is about policy rather than politics in this case or certainly the way it was put to me, that and also so trust and bonding. That first team developed that relationship with the Ukrainians to the extent that they could say trust us hit this, which might not pass muster by policy, but they did it anyway under pressure of war and experience. The new team that came in didn't have the same affinity. It was debatable for some about whether that passing of points of interest was okay or not and whether the two systems were becoming meshed in such a way that America's classified systems would be put at risk. Now it's very hard for me to get super close to that, but from what I've, you know, there's a debate, there was a debate at the time and obviously I spoke to lots of people who said no, we were fine, we did exactly the right thing. But how that intelligence sharing happens for other partners around the world becomes key. And of course the Project Maven team said we could do this for Taiwan. You know, let's brief on how we can support a partner to share without being a direct participant. And then with Iran today, Maven obviously is a direct participant because it's a US led operation. But when it slowed down even under the new team, they became confident. The Ukrainians told me, a Ukrainian officer told me that they were good at getting dynamic targets, which is one of the difficult things to get because that's something that isn't already pre vetted it might be on the move. They were very happy with It. But then the permission for that relationship to continue in some way or something changed and it didn't continue. Ukrainians have found other ways now. They've got their own drones that are getting past, but at the time they couldn't see past six kilometers.
- [41:54]
B
Yeah. So by early 2024, Ukraine had destroyed more than 2,600 Russian tanks and nearly 5,000 armored vehicles. You describe a transporter erector launcher, which is like a mobile missile platform, being destroyed just 18 minutes after detection by American AI. So I have two questions for you here. How much credit do you think Project Maven deserves, especially in the early phase of the war, for the success of Ukrainian resistance? And then I have a follow up question which is to what extent do you think the magic of Maven is in the AI versus to what extent is the magic of Maven just in the graphical user interface and the integration of all of these data feeds, Whether or not AI is a part of analyzing that data.
- [42:45]
A
I don't think anyone wants to hear from me about credit or magic. That's just not, you know, that's not, you know, my analysis on that is not worthy. But I think as a reporter, the data integration and the role it would play was very contested within the Project Maven team. And several people told me we could change out this user base tomorrow. We don't need Palantir, we're agnostic on who we use, but we do need a user interface. Some other people told me it's Palantir that's delivering this. So I think that I didn't. It wasn't clear to me that anyone won on that, except that Palantir and Maven Smart System are getting all the contracts. So they're expanding. Clearly when you've seen other countries, you know, NATO is using it. I do have an example where I think another country tries it out. It is dependent on the data feeds. If you don't have the good data coming in, there's nothing to analyze. And even in some of the early examples of Project Maven where it's not going very well, there are examples of data that's been sabotaged. There's kind of expletives,
- [43:55]
B
not by foreign adversaries, but by disgruntled training data labelers.
- [44:00]
A
Yeah, exactly, yeah. Who got fed up and decided to swear curse all over the training data and just didn't play balls. So that of course has an impact or would have an impact if they'd continued to use that on the quality of the algorithm. So getting the data in such a way that it is well labeled and entering the system and then being able to cross check against others, that clearly is the complicated thing that they've got to grapple with. And Maven is kind of increasingly at that stage. But even in eucom, I discovered that Maven wasn't loading because the networks it was based on were delayed so you could come back.
- [44:41]
B
The network infrastructure crisis, it's an amazing thing to hear.
- [44:44]
A
So in the first few weeks and months they started complaining that they needed more cloud in eucom. And so the project Maven team tried to respond to this and unravel it. AWS got involved, I discover, and it turned out that packets of data were crisscrossing the Atlantic twice or even four times. And so they could lose packets of data that way and certainly slow things down. And then you have. In order to have the classified systems running, you need to use encryptors. For some, they argue that that created a bottleneck. So just to get the information through, you just needed bigger encryptor. To get a bigger encryptor, you need to call someone very, very senior within the intelligence community to get permission to move an encryptor because they're certified by the nsa. So you have all these things that had to come together and weren't coming together.
- [45:33]
B
And so people, a way to think about that, right, is on the entire intelligence workflow, on the entire digital infrastructure workflow, Maven is trying to dramatically increase the analytical capacity by using a great deal more data, by using a great deal computation. And so as they eliminate bottlenecks in one area, they start encountering all the other bottlenecks in all the other areas. And so they basically start demanding, right, that the entire rest of the military digital ecosystem reform itself to be compatible with their needs.
- [46:09]
A
And to your point about credit and magic, the thing that Kuko had always wanted was to deliver intelligence to the people who actually fight wars, to get to bring it down to that battlefield level. And he really, although his whole life is now about AI, it wasn't AI itself. And some people have argued to me through that Ukraine phase that it showed that computer vision doesn't work. So you really can continue to encounter vastly different perspectives on what is happening. And I think that's partly because the US still hasn't got to CJC2 this effort to integrate sensors and shooters. Maven is almost a stepping stone in this effort. And it was quite late. I'd already found it out by the time I wrote the book, but it was only In, I think 23, 24. That I realized that Maven was very key to this idea to creating JADC2. It was. They had decided they should pick a platform and run with it. And you get into all the kind of bureaucracy of, well, do we have vendor lock in, then can we change? How are we going to deal with that? Should we have a consortium that is all the kind of. Of Congress bureaucracy, money controversy? And just that effort to deliver those contracts looks very different from the effort to deliver usefulness to people on the ground. And to deliver one product brings with it its own group of people who will disagree.
- [47:40]
B
Yeah, so now I want to bring us to the present moment, which is the war in Iran. And this includes, you know, not just the work you've done in this book, but also your reporting at Bloomberg as a. As a correspondent covering these types of issues. So what is your sense of how Project Maven is being used in Iran and what type of impact it's having?
- [48:04]
A
I've reported that Project Maven, number one Maven smart system, is being used, that Claude Anthropic's AI tool, the LLMs, is part of that. And the piece that you asked me to read out at the beginning reflects that. LLMs can help with speeding up processes. So it doesn't get to the legal decision or the commander's decision to fire, but it helps with those processes, I've been told, and also can help crunch some of those data feeds and overlay them.
- [48:32]
B
Do you have a sense of the mechanism of LLMs impact? I mean, you wrote that they can increase the rate of targets per day from 1000 to 5000, and that LLMs were a part of that acceleration. But what actually is it that the LLMs are doing? And what part of the life of these warfighters does it make easier that accelerates that?
- [48:55]
A
I was told that it was speeding up the processes. So the processes that you would need to get permissions, not the actual permission itself, but almost that ferrying of paper backwards and forwards. It's the admin side of the targeting cycle that it was helping with, not actually in that sense, the identifying side. I think things have moved on because you also now have reasoning where they are on the different data feeds. You'll have to just give me a bit more time to get decent answers to that.
- [49:21]
B
I don't think we look forward to your next next column.
- [49:24]
A
Yeah, or I'll look forward to your next podcast on that. But I think what I've definitely determined is Claud is being used. It's still being used Maven Smart System is being used. And then what CENTCOM has publicly said is, the spokesperson said to me they're using a variety of AI tools and that these are helping to generate points of interest and help make smarter decisions faster. And then the commander has publicly said that it is helping bring AI, is helping bring down processes from days and hours, sometimes to as little as seconds. And he's taking time to say that in the middle of this war. So their focus, their belief in leaning forward into AI is fascinating. Don't forget that senior Centcom people in 2023 were awarding Maven a grade C in public. So it was a tool they wanted but was still frustrated by, and they may. Well. I don't know what public grade they would award Maven today. It's a question I'd love to ask, but just because this tool exists, I'm sure we would find people who are still frustrated with it, who want much more. But it is a tool, clearly, that I think we can confidently say. CENTCOM says it is helping them speed up.
- [50:40]
B
Yeah. Well, I certainly see a pretty dramatic impact and understand this as a pretty remarkable effort. Now, I'm obviously biased, but I want you to not worry about hurting my feelings. So Maven is the original AI Pathfinder in the Department of Defense. There are these other initiatives going on. We've talked about diu, the Defense Innovation Unit. We've talked about the Joint AI Center. Obviously, the NRO has AI initiatives underway. The various services have AI initiatives underway. Sometimes the they overlap with Project Maven, sometimes they don't. You know, it seems from you picking this book that you thought that Maven was a special story. One reason why you might think that is that it was more successful and more impactful than those other organizations. Maybe you just thought the personalities were more interesting than those other organizations. But I'm curious to get your sense of. Was Project Maven unique and special compared to those other organizations? Was it more impactful than those other organizations? And if so, why do you think that was? And again, don't worry about hurting my feelings. I'm made of steel.
- [52:02]
A
Nobody's made of steel. I think really what drove me to Pickett was because it was so public in one way, but so not public in another. And so, in that sense, so the
- [52:21]
B
most fun to be an investigator of.
- [52:24]
A
Well, I had so many questions I wanted the answers to. I wanted to know how algorithms fare. And actually, a lot of the examples in the book are examples where algorithms are not successful, but it's part of a learning curve. Trying to.
- [52:41]
B
In Silicon Valley, they often advise startups to fake it till you make it. And I think that's a little bit of the story of Project Maven, which is they talked to big game, but eventually they became big game. But there was a long time where they were talking more than they could accomplish.
- [52:56]
A
I think, in terms of kind of success from the Pentagon's perspective, or from Project Maven's perspective, obviously, you have this platform that is named after Maven, that is used by every command that has more than 25.
- [53:10]
B
Maven smart system.
- [53:11]
A
Yeah. Which has more than 25 accounts. Palantir has this big role in it. The licenses keep flowing. So in that sense, Project Maven has created a product that is a success for Project Maven. Now, the key debate inside Project Maven was, should they be creating that platform or should they be developing AI? And someone like Colin Carroll felt very differently. He who was on the team, he thought that the digital interface was a distraction from the pursuit of AI and that actually AI should be on the machines themselves. And so in some ways, that was a failure. And when he writes to Kukor, I find his letter. He's very disappointed in what Project Maven has achieved. And so he loves Project Maven, and he loved the hard work. He loved the dedication, all of those things. He loved how unleashed they were as a team to go up against other people. But you are definitely offering a particular perspective to say Maven is this big success story. It has created one tool that still has competitors. The nro, still has its own intelligence platform. I've also heard that Maven is not succeeding at the operator level, which is exactly the level it was hoped for because of bandwidth problems. And I spoke to Alex Miller, the CTO of the army, about this. He believes in Maven. He wants more AI. He's someone who's leaning forward. He was described to me as a frenemy of Project Maven in the early days, but now in his current role, obviously, the army is leaning very heavily into Maven, but it is still trying to untangle problems. And this effort to link up sensor and shooter brings with it all sorts of questions of accountability, data flow, all the rest of it. So, to me, it isn't as clear a picture as the one you're presenting. And then, of course, we get onto some of the very explicit problems of Maven that I recount in terms of AI getting onto drone platforms. This is still also. I mean, the very thing that campaign is worried about is the thing that I uncover in the book that Maven tried to do so, get AI onto drone platforms to be used, used as automatic target recognition. And that process stumbled. It was fantastically successful at collecting data of Chinese vessels at sea. And so it created algorithms that could try to identify those. If you could imagine AI sitting on side a drone boat or an aerial drone. But what happened is the integration was difficult. The algorithm makers often don't get the kind of feedback that they would really welcome because they're at arm's length from the operators. Numerous times I spoke to algorithm vendors who said, we want the same level of access as Palantir. Now, I'm told that's a common complaint. Everyone wants the same level of access as Palantir. But the argument from the AI perspective was unless we're sitting with the vendors, with the users, we don't know how to tweak our algorithm quite what they want. And it gets lost in translation. So I think there are still, for those who support this technology, a huge number of lessons to be learned about how to get that technology out to people and develop it in a way that it can be relied on. And then there were just technical problems, like a splash from the ocean could interrupt the tracking ability of the algorithm. Now, that's not AI's fault, but there are all sorts of logistical hurdles that stand in the way of smooth running AI that can select detectives.
- [57:01]
B
So one person whose story you tell in the book is Admiral War Whitworth.
- [57:07]
A
Yes.
- [57:08]
B
Who starts out as, I think it's fair to say, one of the biggest skeptics of Project Maven and ends up in charge of it and one of its greatest true believers. So what do you make of that evolution and how it happened?
- [57:21]
A
It's so fascinating, his story, I mean, he's someone who was so involved in the targeting cycle, so exactly the. And on the intelligence side. So exactly the kind of people that Drew Cukor saw himself as going up against the J2s, really to deliver intelligence to operations rather than to intelligence. He argued that if he had stuck with Intelligence, Maven would just be another failed intelligence project. Admiral Whitworth was very concerned about accountability, about who would defend this in front of Congress if AI contributed to a targeting error and also whether it was going around taking shortcuts around the targeting cycle. He'd already been briefed on Maven and given them quite a tough examination before it was clear that he would run NGA and that he would get Project Maven. So when he did get it, many of the Maven folks were concerned that he might even Kill it off. What he told me about his conversion is that he found that Maven was able to update and respond to the realities of war quicker than anything he'd ever seen. And it was that pliability of the software, that ability to respond to what U.S. operators needed, that really converted him. And he also worked hard, I think, at nga, to develop a way of assessing models. So by then, especially as it's spreading the reliability of a model in this technology that makes mistakes and hallucinates and all the rest of it, he wanted to sort of characterize the ways in which these models may be successful or may fail and make sure users understood that. So he did some of that, I suppose others might call it the growing up of Maven. I must say, there are also complaints that then Maven slowed down, but the proliferation of it certainly expanded wildly under his watch. And he regularly spoke to commanders to say, here's what I've got. That's new for you this month. What would you like? And finally, Maven had had terrible difficulty cracking into paycom, which is this great irony, given the whole of Maven was meant to be to go up against, you know, provide AI that could be useful in a fight against China, should there ever be one. Admiral Whitworth, I think it's fair to say, under his watch, had the most success at breaking into Indo paycom. And now indopacom is a great cheerleader of Maven and Admiral Paparo, who leads it, hosts AI summits and wants more
- [59:53]
B
from industry and is a big cheerleader of the use of LLM technology as you write as well.
- [59:59]
A
Yeah. I'm told he's very fond of Claude Told.
- [60:02]
B
Yeah. Well, Katrina, I could easily keep you on for another hour, but maybe that would be unfair to you and hopefully not to our audience. Hopefully, they would eat it up. I think they would. But what they should really do is go buy your book, because this is by far the most exhaustive form of the story. I think one thing that's just worth calling out is at multiple times, Cukor and other people on the Project Maven team who were famously reluctant to talk to the press do say, oh, this will all end up in a book someday. And here is the book. Here is the book. So congratulations to you on writing it. You have, I think you say, more than 200 interviews or people you interviewed who were a part of this story at various stages. There's a ton of details and, as I said, the minimum bedrock of knowledge for anybody to intelligently participate in the military AI conversation. Going forward. So, Katrina, thank you so much for coming on the AI Policy Podcast.
- [61:02]
A
Thanks. Thanks for having me.
- [61:06]
B
Thanks for listening to this episode of the AI Policy Podcast. If you like what you heard, there's an easy way for you to help us. Please give us a five star review on your favorite podcast platform and subscribe and tell your friends. It really helps when you spread the word. This podcast was produced by Sarah Baker, Sadie McCullough and Matt Mand. See you next time.