AI in Defense

John Dulin is the founder and CEO of Modern Intelligence, where they are building AI foundation models for defense. He was previously at Numerai and Freenome.

(00:00) Introduction
(00:08) The State of AI in Defense
(03:17) AI Impact on Tactics in Conflicts
(09:24) Real-Time Threat Assessment
(18:21) Naval Operations and Maritime Surveillance
(22:28) Startups and Maritime Technologies
(25:58) Naval Dominance and Global Power
(31:20) Multi-Intelligence in Defense
(35:20) The AI Arms Race and the Future of Defense
(38:59) Rapid Fire Round

John's favorite book: Wilhelm Meister's Apprenticeship (Author: Johann Wolfgang von Goethe)

--------
Where to find Prateek Joshi:

Newsletter: https://prateekjoshi.substack.com 
Website: https://prateekj.com 
LinkedIn: https://www.linkedin.com/in/prateek-joshi-91047b19 
Twitter: https://twitter.com/prateekvjoshi 

Transcript

Prateek Joshi (00:01.571)
John, thank you so much for joining me today.

John Dulin (00:05.442)
Thank you, Pratik. I think it's gonna be a lot of fun.

Prateek Joshi (00:08.403)
Perfect, let's get right into it. What's the state of play when it comes to AI in defense?

John Dulin (00:20.022)
That's a big question and the answer is quite surprising. So I think if you step back and look at defense today, the Pentagon is spending what is approaching a trillion dollars a year every year over the next decade to really rebuild the entire American military because we've lived our whole lives with this assumption that

America and her allies have the largest and most technologically advanced military in the history of mankind. And that's been true, however, that's been based off of largely four decade old platforms, planes, ships, vehicles, built in the 80s and 90s, and then modified through today. And the goal of that $10 trillion in the next 10 years is really to

rebuilt the American military from having few and expensive platforms that have people in them that are manned and are putting people in harm's risk and go to having 10 times as many unmanned and expensive platforms. And the goal of that is not just to take American soldiers, sailors, Airmen and Marines out of harm's way, but it's also just to watch and respond.

to threats across the globe, whether it's a potential Chinese invasion of Taiwan or a Russian invasion of Ukraine, at the pace of the 21st century. So that means building the world's largest mesh network of sensors, radar, sonar, video feeds, infrared feeds, that the world has ever seen. And that is where you need AI. Because when you build that 10x, you cannot 10x the number of people you have staring at video screens.

telling their commander or telling the guy on the ground what they see, which is exactly what they do today. And as I've learned more about defense, I've been really struck by how much of that work of watching sensors, processing sensors is manual today. And I think it's partly a testament to the traditional hardware focus of defense, but also

John Dulin (02:46.038)
a focus on the great AI technological talent of like the last decade, working in an industry rather than for defense. And so what we have today is a lot of work in front of us to build that AI and in modern intelligence's opinion, the foundational AI to watch all of those sensors all at once, be multimodal and see the battlefield because it's gonna be moving too fast and too big for people to do it.

Prateek Joshi (03:17.555)
amazing. If you look at like world history, obviously, it's important to maintain peace and it's important to be peaceful. But the history of mankind has always been the world is rife with conflict and it just happened. It's just the nature of humans and sometimes after until after the war happens and then you realize what's at stake. You don't really know who's on the right or wrong. I mean, you know, but

John Dulin (03:36.055)
Yes.

Prateek Joshi (03:46.619)
the world knows much later. So, you know, if you're going all the way back to the Syrian Empire, Persian Empire, the Roman Empire, Byzantine Empire, Ottoman Empire, and today, if you look at today's world, the nature of warfare keeps changing and evolving. Like there was a time when you showed up with large armies and swords, then there are empires who showed up with warships and then aircraft. So if you look at

today's world. How is AI impacting the tactics in these conflicts? Obviously, it's important to be on the right side, but you just have to sometimes you have to defend. So how is AI impacting those tactics today?

John Dulin (04:32.094)
Yeah, this is a great question and I love it because I feel like you knew that I played a lot of Age of Empires and like civilization growing up.

Prateek Joshi (04:42.568)
Ha ha ha.

John Dulin (04:44.842)
And then I got a physics degree. And so if I think like a physicist for a minute, just from like first principles, I think the simplest answer is the ability of AI and the compute and networking that we have today and both the American military and our adversary, or our adversary's militaries are about to have, um, what that's doing.

is shrinking the time scales on which war happens by 10x, and then expanding the distance scales by 10x. And Ukraine has actually been a tragic and horrible but good example of this. Because when you have people

whether they're watching a video feed from a drone or whether they're just on the ground doing reconnaissance on a battlefield, like a scout or a spotter or like a reconnaissance plane looking for the bad guys, looking for the targets, looking for the Russian tank or the Russian artillery piece in Ukraine. If that's all manual and you have people who have to

You know, like configure complicated radios, talk on radios, go back and forth on radios, make sure they're seeing the same thing, talking about the same thing, then doing math, looking at maps to figure out the exact latitude and longitude of where the target is, which happens a dozen times, hundreds of times every single day in Ukraine today, which is largely this long range artillery battle. If you have

a Russian artillery piece and Ukrainian artillery piece, and those two teams are trying to find and hit one another before the other does. The first impact you see is speed. So now if you have quadcopters, and if you can use AI to process the radio feeds and put a pin exactly where that Russian artillery piece is, the moment it sees it,

John Dulin (07:08.518)
And have that latitude and longitude and have that in real time. And then rather than have that be a conversation that four or five or six people are going back and forth on a radio, but have it be a data packet. That goes into command and control software that can then go to an utility piece, you've now shrunk the time that it takes for two adversaries, whether they're shooting at each other with

artillery pieces in Eastern Europe or with long range anti-ship missiles in the Pacific, you've shrunk the time that it takes for them to see, understand, decide and act on striking each other from 30 minutes to 10 minutes to five minutes to 30 seconds. And so that's the first big step is the timelines are compressing. And when you're integrating AI into this process.

which is actually a term for the Chris Burrows wrote a great book about Chris Burrows, the chief strategy officer on role called The Kill Chain. This entire process of which AI is only apart with networking and decision making is called The Kill Chain. How fast is your kill chain? How robust is your kill chain? How quickly can you make decisions before the other guy does? The first step is the kill chain is just getting faster and faster. And that speed is part of the arms race. And then I think the second part is the distance scales.

Now that there are more commercial sensors, commoditized sensors that are being put out in the battlefield and plugged into these decision-making networks, now you may not just have 10 planes with limited ranges, searching 100 miles, 500 miles. Now you might have a thousand drones searching 10,000 square miles.

And now with hypersonics, we're building the long range weapons to act on those distance scales they see. So you are seeing each other and then threaten each other on wider and wider ranges. And so you need all of these drones and all of this sensor processing to see that range, and then you're going to race to act faster. So it's speed and distance and AI helps with both.

Prateek Joshi (09:24.992)
I mean, that's a brilliant way of explaining how it works. And I love that way of thinking. And it kind of gives a glimpse of how people have to make critical defense decisions, often in uncertain scenarios. And AI can be reliably infused into the process. And you can make it.

make it faster, as you said, cover more distance. And more than anything, you can just process more data and extract information upon which you can take a decision. Now, if you look at real-time threat assessment, right? And threats can happen in about a billion different ways. So much data coming from all direction, most of it is usually noise. So how can AI help?

in a real time threat assessment, especially when the battle is happening right now. Every second counts. So how does AI help here?

John Dulin (10:32.062)
Yeah, you used a great word, noise. I think the noise floor in data, you know, in, in machine learning, I previously worked in biotech and in defense in like the radio or RF background where you're operating is, is a really, really valuable concept to. Use to think about how are we trying to use AI to.

um, move the noise floor, um, of what a human looks at and, uh, move it down by combining lots of sensors so that like the whole is greater than some of its parts. You're not just looking at an infrared sensor or synthetic aperture radar sensor or an RF sensor. You get to see what all of them are seeing at the same time. And

One way AI helps move that noise floor lower for people making targeting decisions, making command decisions is, I think it, first of all, just allows them to draw more counterfactuals. Like, okay, I had this one really low resolution sensor, identify 10 boats that we think are suspect in the South China Sea. Maybe they're a Chinese spy vessel, right?

but it's low res, so it can't get like a unique vessel classification. But now I can also pull in RF and I can see, okay, which of these have a transponder that looks like a normal commercial vessel and which of these have a, you know, unique RF signature that looks like it's talking to Beijing or like listening to something, right? And then, okay, now I have tipped off that these two ships out of those first 10.

the one I care about. Now I can bring in a higher resolution sensor to go look at that and go look at those two ships and then understand, right, which one do I really want to watch. So you want to be constantly lowering the noise floor for the decision maker. And one way to think about this that I like actually is it's like a funnel.

John Dulin (12:54.934)
It's a decision-making funnel. It's sort of like a sales funnel where you're like qualifying and disqualifying deals, except you're doing that with targets. And you want to close the target. You want to close the, you know, the kill chain loop. But the more data you bring in, the more you can filter them. But then also the better AI you have, the faster it can help you sort through that. And so two big focuses of modern we've had are one.

working on multimodal sensing from day one. So not just doing computer vision, but computer vision, infrared, synthetic capture radar, soon, acoustic and RF, all of these are modalities to bring it together. And then actually make computers that can do even more of that counterfactual sorting before a human gets to it. So not just...

One of the first big problems we solved to help, I think, with this noise floor and this funnel was not just solving classification, but solving re-idea, remembering you need targets across time. So not just saying like speedboat, freighter, when you were watching the South China Sea to see what the Chinese Navy is doing, but building a model that actually is building a

multimodal vector embedding in your sensor space. Essentially, you're getting unique ReID vectors of the targets that your sensors are seeing. And so now you remember black speedboat, three upper motors, green speedboat, two upper motors, red speedboat, one motor, two passengers, funny looking radar antenna on top, red Chinese flag to help give users more tools to sort through that noise floor.

Prateek Joshi (14:51.683)
I think it's a good stopping point to quickly talk about modern intelligence. So for listeners who don't know, can you quickly explain what you do?

John Dulin (15:05.41)
We're building the foundational AI platform for defense. And that starts with perception because it's the simplest, it's the most common and it's the most important immediate problem. And can also be bounded to like the most simple problem one defense customer has right in front of them, which might just be, you know, I have one Air Force intelligence plane and it hasn't had AI on it for two decades.

but it has a mission computer, it has all these sensors. Can I buy an AI and a Docker image to run on it and re-ID terrorist suspects or re-ID drug vessels? So we've started by building the multimodal AI platform to take in all of this raw sensor data all at one time from

any defense sensor, whether it's built by Anderil or Lockheed or Raytheon. And then distill it into all the locations of all the bad guys, distill it into a more valuable understanding of who's who and an ability to remember them across time with that ReID vector embedding. And then share it to all the defense systems in the world. And part of why...

we felt like this needed to be built in our conversations with customers and users and looking at the defense market was. There's a lot of fantastic companies building commercial hardware to bring into defense today to build that 10X in sensors and platforms that we need. And then there's a lot of other great companies, Palantir, like Rebellion, Booze, building the enterprise services to tile that data together.

However, fundamentally these companies are focused either on being hardware companies or services companies and unlocking the value that AI brings to totally change what's possible with a duty customers mission with fusing every multimodal sensor they have re IDing and building that building that baseline of a unique target. And then even more advanced perception features we have like

John Dulin (17:34.166)
one-shot learning, predictive analytics of like where a target's going to go and what's it going to do. The hardware companies and the services companies fundamentally are not AI companies, they don't get to focus on that and they're more focused on building their hardware, building their services, which is amazing, but AI in defense, especially 10 years from now, is going to be such a hard problem.

and such a big problem, and it's going to be spread across all of these different sensors that all of these different companies are building, we need a foundational AI for defense. And then we need a foundational platform that can plug into all of these different systems to make the whole greater than some of its parts. And that's what we're building at Modern.

Prateek Joshi (18:21.467)
Amazing. I want to shift the conversation to one of the biggest and most important component of defense, which is naval operations. And obviously there's a ton of nuance here. And I recently did an episode on underwater robotics and it's fascinating. And also, how surprising how complicated, how much more complicated it is than

like building a land robot and working on land. So let's start with, when you look at naval operations, let's start with surveillance. What is involved in maritime surveillance and what are the key technologies involved here?

John Dulin (18:53.272)
Yes.

John Dulin (19:13.002)
It has a lot of the same principles of any one of those kill chains or like search funnels that I talked about, which is also sometimes called an offense, like fine fix finish, like a targeting cycle. Um, however, it requires different tools and then it happens on a much larger scale. Um, in naval it's difficult because.

The scales are so large, especially in the Pacific, that you can only see so far. Whether you're a ship or a plane, your radar might only reach out to the horizon. Your vision especially only reaches out to the horizon. And a plane searching an area can only cover so much ground so fast before it runs out of fuel. And so,

You need to bring in other sensors that can cover a lot of ground really fast. And then what's called tip and cue those other assets, like ships and planes to go take a harder look at something, which is space-based assets, satellites, um, that do RF detection that do synthetic aperture radar that do EO. And many of them at lower resolutions can scan, you know, like huge, um, like a hundred miles trips.

of ocean. And then of course, there's radar and sonar and radar, especially like long range high frequency radar, you can see very far, it can't see over the horizon. But then you're getting radar pings, which are just dots, or at least like low resolution dots, you can try to combine those with other like multi in sources to understand them better, which is again, how you need AI's help.

And then of course there's sonar, which you absolutely need to see anything that you're hunting for if it's underwater, but also can be really valuable for casting a wide net or quietly listening for things over a wide area with things like sonar buoys and like picket lines. So it's still a lot of those principles of different sensors working together to

John Dulin (21:35.998)
see really far, decide what you're interested in using like multi-intelligence, then scope down, look at something harder, scope down, look at something harder, scope down on the thing you actually wanna look at. And you just need to use all of those together.

Prateek Joshi (21:54.055)
And when it comes to the needs of the government, obviously there are so many things that they're using, they're building. And if you look at startups that, let's say there are founders who are really excited about maritime technologies and they wanna build something useful and sell it to the government. What are the big gaps or rather, what is the government more likely to buy?

when it comes to startups in terms of maritime technologies.

John Dulin (22:28.046)
Um, it's a good question. I think there's a lot of great companies like SailDrone is pretty far along now, like Sironic, um, region for maritime logistics, who have pretty good sites on a lot of the fundamentals, um, with SailDrone that's super long-term endurance maritime domain awareness because SailDrone builds these, these drones.

sailboats with solar panels and they can stay out there for a really long time and they just watch. They're not super fast, they don't have weapons on them, but they can watch everything. And then Saronic is taking that a step further with more like speed boats that aren't as quite as long endurance as Sail Drone, but they went about a whole lot and deployed them fast.

and then they can of course react faster than a sail drone and their goal is absolutely to get to like kinetic payload someday. And there's a lot of the great satellite companies like Umbra, ISI, Hawkeye 360, who are building out the satellite constellations that the government customer will need to watch the maritime domain. All of those you know ultimately fall into

more sensors, more data, all of which modern, you know, wants to process someday, wants to partner with. And I think the really most difficult technical problems that are both in acquisitions and technology, like the hardest things to do right now, are underwater.

both sensing platforms, weapons systems, UUVs, unmanned underwater vehicles. Under-Rills Dive is, of course, a wonderful platform. And they're going to be building and deploying a lot of those with the Australian Navy. But it is a really big need. And it's also probably not a need that will be solved by any one.

John Dulin (24:49.894)
There will probably be some that are like massive and launch others. There will probably be like others which are smaller and launched almost like torpedoes. And there's a lot of really tough problems there that modern is having conversations with partners we've had who are interested in, in modern processing sooner data of just how do you send data back from and to things underwater. And that's

partly going to be innovation on communications, but it's also going to be a lot of innovation on software.

Prateek Joshi (25:25.555)
Amazing. And this is more of an open-ended question. If you look at all the naval powers, at least especially when people figured out how to build like big ships, any country or empire with the biggest like naval fleet usually tends to dominate a lot. Obviously land armies, air warfares, there's all been here. What gives naval technology such

dominance and indiolence. Like why is it that if you are dominant in naval operations, you tend to usually have the upper hand?

John Dulin (25:58.594)
Yeah.

John Dulin (26:06.466)
This is the mayhem question, Alfred Thayer Mayhem. You know, sea power is everything. And the sea power versus land power conversation is also a good conversation for age of empires and civilization players. It's fundamentally a massive, massive majority of global trade, oil, freight, natural gas.

commodities travels by sea. And the importance of that is greater than it's ever been before because of how globalized the economies are today. And the nation state with not just the best Navy in quality, but the biggest Navy in quantity that can be everywhere and do everything to ensure maritime security.

and promise globally, or at least in the areas it cares about, that maritime trade here is safe because I say so, becomes a fundamental pillar of the global economy. And then of course, they get a large say in what that economy looks like. This was what the British Empire did. And after World War II, and then especially after the end of the Cold War, this is the role that the US Navy has played.

And this has formed the foundation of our partnerships with the Gulf Arab states because their entire economies rely on oil. The rest of the world relies on their oil. And they need to take that oil through waterways that sometimes have mischievous other nation states playing in them, like the Strait of Hormuz, the Red Sea, like we just saw this past weekend. You know,

Huthi rebels shooting anti-ship missiles and US commercial vessels and destroyers. And being able to be the power that tells the Saudis that you'll always be able to ship oil because we're gonna protect it. Or tell us Taiwan, you're always gonna be able to ship silicon because we're gonna protect it. Is a really, really powerful position. And it's also why

John Dulin (28:35.318)
A Chinese invasion of Taiwan is such a critical thing for the United States to deter for our national security. There's a really great blogger who lived in China a long time and now writes about geopolitics that his name is Tanner Greer, who had a blog post laying out like, why would Japan be motivated to join a defense of Taiwan scenario?

And the answer was pretty clear. It's that 80% of all shipping that goes to and from Japan, an island nation, goes past Taiwan. And if China takes Taiwan, well, now they're that force that can tell Japan what oil they do or don't get, what food they do or don't get, what chips they do or don't get. And so that's always a really important thing. But also it is something that doesn't just rely on quality alone. It relies on quantity.

And that's why one of the most important things the U.S. needs to solve today, both in its big shipyards, but also hopefully in its small shipyards with companies like Staronic and SailDrone need to solve building lots and lots of maritime assets.

Prateek Joshi (29:49.607)
Yeah, and for hundreds of years, if you look at the early Dutch, they built amazing ships, built ships that can go all the way to Indonesia and they monopolized the spice trade, a fabulous wealth after that. As you said, British Empire, the dominance, I think most people would agree the naval dominance is a single thing that

held them together for like so long, because anytime somebody disagreed with anything, they showed up with warships. And that's a pretty good position to be in. Like, what are you gonna do? And obviously it's a way to maintain peace and order, obviously, in what they did. Similarly, like, I don't know, one of the castes, one of the key reasons why World War I started building up is Germany said, hey, British, they have their ships. They look pretty good. We'll start doing that. And then British got nervous. They're like, hey, we don't want...

competition and that, and obviously the many reasons, but yeah, it's like such a flash point, the naval dominance, it's pretty interesting. All right, coming back to the way in which data flows through and you alluded to earlier to this point, can you explain the term multi-int? And also when it comes to defense, like what...

are all the types of data that they're dealing with, that they as in the people in charge of defense.

John Dulin (31:17.835)
Yeah.

Yeah. Multi-int is one of those great, um, DoD words. Um, you don't know what it is at first, then you learn it and you love it. Um, multi-int is a fancy word for, um, kind of like an AI we refer to multimodal. However, the int is the, the DoD abbreviation for intelligence. So all decision-making, um, all planning.

and all targeting decisions in the deity need to come from intelligence. And there are countless communities and countless experts and technologies in the deity that form these different areas of expertise for different types of int. So there's geoint, which is like geo...

spatial intelligence, meaning normally imagery or data collected from space. So you had a satellite, you collected it, it's geo-int. If you want to be more specific, then there's im-int, you know, image intelligence, if that imagery from a satellite was a photo. But that can also apply to video feeds from drones. Then there's sig-int, which is signals intelligence, listening to people's...

RF emissions, their radios, their cell phones, their encrypted messages, things the NSA does. There's human, which is human intelligence. We have a source or maybe the intelligence community has a source that is telling us information that we can then corroborate with all of these other ints. And multi-int is always more valuable.

John Dulin (33:17.954)
to a defense decision maker than like a single intelligence because making a decision is, it's sort of like building a case. Like there's always fog of war, but going back to that like flannel of decision making that has counterfactuals in it that you can constantly draw. The more ints you have,

GEOINT, SIGINT, MAZINT, IMMINT, HUMINT. The more of those you have, the stronger case you're building for a commander to make one decision or another and the more confidence they can have in it. But each of these, and I think this is part of how AI will help in the next decade, each of these fields are, they've had hundreds of billions of dollars invested in them over decades.

And they have really, really exquisite capabilities, especially in classified settings, and tons and tons of knowledge, but they are often very focused on like the specific int, that office or that science and technology researcher, or that analyst is focused on. And part of, I think, how AI can help is, we need to be making more multi-int decisions.

faster and bringing in like a wider and wider and wider birth and AI can also help with that.

Prateek Joshi (34:51.703)
Right. I have one final question before we go to the rapid fire round. And it's about this AI arms race that is building up. If you look at the global AI supply chain, there are so many critical choke points, like starting all the way from ASML, like almost a monopoly that supplies the equipment TSMC needs to kind of build those ships. And then there is

TSMC in Taiwan, and that's almost a monopoly in advanced chips. There's Nvidia, insane dominance in creating the GPUs that's powering the AI world, and obviously many little companies around the world. And also you need rare earth metals to build some of this, that it just, yeah, so there are countries where you, where it's just, you cannot control it, it cannot engineer, you either have rare earth metals in the ground or you don't. So with AI in the mix.

What does the future of defense look like?

John Dulin (35:58.422)
Well, I think the first and most important thing is securing the technologies which we have secured today. So that's ASML, that's TSMC, that's a whole collection of smaller companies with expertise around things like packaging or, you know, subcomponents that TSMC and ASML need. So the first and simplest and most important thing we can do today is use our

shared values, but also economic and political weight with the Netherlands to protect that IP, protect IP like even IP cores for silicon chip designs from like ARM, from diffusing our most advanced technologies into China too quickly. I think that is really, really important. I think sometimes I think of our adversaries like China and Russia as-

I think of them as like technological and intellectual property heat sinks, where like, you know, the West and the US is where like, a lot of this innovation is happening. And so we're like this furnace burning, creating all this technological heat. However, if we're not protecting it, then our adversaries are just like these big heat sinks that are like wrapped around the furnace and they're just absorbing it. Right?

So I think we need better like thermal insulation there. And then, you know, a key part of that is doing everything we can to deter an invasion of Taiwan. The easiest way to protect Taiwan's independence, to protect the security of the Pacific. And then of course, the technological security of semiconductors is just for that to never happen, right? But I think while we're doing everything we can to

Prateek Joshi (37:29.779)
I'm gonna go to bed.

John Dulin (37:57.054)
insulate that heat sink and deter more aggression, we still need to be acting as if that will not last forever. And we have the capital, partly thanks to the CHIPS Act, and we have the talent to build those technologies here. And I think we have to, because if we're wrong and we can't deter an invasion of Taiwan,

You know, we might not have an opportunity to onshore all those technologies fast enough.

Prateek Joshi (38:35.751)
That is a wonderful way of thinking about it. And I agree, I think it is becoming more and more critical and yeah, it's, we've got to secure. The entire supply chain needs to be secured. All right, with that, we are at the rapid fire round. I'll ask a series of questions and would love to hear your answers in 15 seconds or less. You ready?

John Dulin (38:59.726)
Let's do it, I'll watch the clock.

Prateek Joshi (39:01.391)
Alright, question number one. What's your favorite book?

John Dulin (39:07.53)
I'm an odd one. Wilhelm Meister's Apprenticeship by Goethe. It's a beautiful coming-of-age novel written, I guess it was the end of the 18th century. And it tells the story of a young boy who has his first love, loses it, has his first passion, which is actually theater, which was one of my passions at one point, and then his growth. And it just captures, I think, how young men grow really well.

Prateek Joshi (39:37.231)
Amazing. What has been an important but overlooked AI trend in the last 12 months?

John Dulin (39:52.478)
I think it's still early for non-text vector embeddings. And this is a lot of what modern is focused on.

John Dulin (40:06.73)
Vector embeddings are this incredibly powerful tool, but right now everyone's attention is focused on them for text and LLMs. However, part of what we've seen is you can do pretty magical and insane things. And I need to maybe read the Gemini paper after this, but you can do pretty magical and insane things to solve hard problems with.

unstructured data vector embeddings. And part of what we've seen also is current vector databases aren't ready for that yet. And so we're going to need to build them.

Prateek Joshi (40:44.407)
What's the one thing about defense tech that most people don't get?

John Dulin (40:53.058)
Bringing AI into the battlefield doesn't just make warfare safer because it's automating a process where you might have human error. It also can make it safer because you can actually design AI in a structured way. Like you can't always design a human when they have a lot of cognitive load, when they're tired to have the safe biases.

John Dulin (41:23.482)
We had a bias for false positives for our analogene of cancer tests because we want we wanted to be better than colonoscopies and say people doing colonoscopies first and you can do that with AI in defense if you decide there are situations where it's better to have a false negative here because we really want to err on the side of not striking a civilian you can put that bias into the model.

and acknowledge that these are fundamentally probabilistic models and not always right, but then reflect it with positive bias.

Prateek Joshi (41:56.547)
What separates great AI products from the good ones?

John Dulin (42:05.462)
Picking the right 80-20 problem to automate with AI. There's a lot of problems where 80% of the work is 20% of the problem, and 20% of the problem is 80% of the work. And I think product leaders in AI companies,

John Dulin (42:26.366)
either win or lose by how they pick that. Because if you pick right, then people just do the last 20%, which they're happy doing. But if you pick wrong, people are gonna spend so much time fixing your AI's problems that they're gonna hate what you build.

Prateek Joshi (42:40.807)
I love that way of thinking. It's a wonderful framework. All right, next question. As a founder, what have you changed your mind on recently?

John Dulin (42:51.926)
Mmm.

John Dulin (42:58.314)
I think 10 things I need to pick, but the clock is too late.

John Dulin (43:04.999)
I think...

John Dulin (43:10.39)
in...

defense specifically. It's not giving people, it's not giving people clear roles early, because people should know what their responsibilities are, how they're gonna grow. But it is.

John Dulin (43:37.398)
basically not having teams at all when you're less than 20 people. Like when someone asks what team I'm on, it's like, there's one team that's less than 20 people.

Prateek Joshi (43:46.607)
Right. Actually that's a good one actually, because in the early days, you're on team, in this case team modern, but you're on one team because it's less than 20 people. That's a good one. Yeah.

John Dulin (43:47.423)
Yeah.

John Dulin (43:58.59)
Yeah, yeah. You always want to have a sense of giving people clarity and clear responsibilities. And like, you can do that with an individual, but like for teams and like, how you structure meetings and stuff. And you're like, Oh, I don't want to have too many meetings. Like, let's not pull everyone in. It's like, actually, no, just everyone all the time, but do everything all at once. Yeah.

Prateek Joshi (44:18.835)
Right, right. All right, next question. What's your biggest AI prediction for the next 12 months?

John Dulin (44:28.91)
Who? Um.

John Dulin (44:42.969)
I think...

Open AI is going to keep winning. I think they're doing a really good job there. I don't know if that's controversial. And I think in defense.

John Dulin (45:01.506)
we are going to see more AI completely change what defense missions can be accomplished with pretty wild capabilities, especially to like see things and hear things that people miss, independent of like a specific hardware platform, like independent of a specific drone. And I know a few things there.

Prateek Joshi (45:31.177)
Right, right, final question. What's your number one advice to founders starting out today?

John Dulin (45:31.342)
Okay.

John Dulin (45:40.946)
Um, I think it's know who you are, especially in AI. Um, you know, two years ago, everyone could raise a ton of money to do a ton of R&D for a long time, or what they thought was gonna be a long time. Um, and you could spend more time discovering your problem. Um, I think in AI...

You have an advantage if you already know the problem you're solving, you already understand it, you're solving it, and you're just going to be super lean and get something in the user's hands as fast as possible. And all those principles are still true. And if you can do that and be cashflow positive as fast as possible, like do that in this market. Um, but I think it's also still possible if you have a vision in AI to raise a lot of capital for the moonshots. However, I think you need to be really clear with yourself and your investors on who you are.

Are you the moonshot company and like how clear do you have expectations for investors on like, you know, when you're gonna have revenue? Um, or are you the like lean MVP as fast as possible company? And once you decide that. Don't confuse the two. Like once you're the lean MVP company and like, you see like one like cool, you know, research project you could do. Um, like don't get distracted by that. And like actually once you're the moonshot company and you.

feel like, ah, if I do this commercial deal, I'll get like a million ARR. Maybe you don't need to do that until your moonshot is like well and truly landed. I think it's like, just know who you are.

Prateek Joshi (47:20.815)
Right. And that's a wonderful, wonderful piece of advice because I see many, many founders who are like, Hey, I'm going to build this consumer company. And they're like, you are, you are an enterprise person through and through you just not, it's just not too. Okay. I think I'm talking in terms of business model fit, but basically like if you're a consumer person, just do that. Or if an enterprise person do that, if you're a defense person, just like, and also it also, uh,

translates to who you are as a leader, as a person. If you're a person who can manage well, who can hire. Like, I think that's a great point. I think just know who you are. Many people don't spend enough time, I feel like, just discovering who they are and then going all in. They try to be someone else just because or it's easier to hire or fundraise, but I think it's a great point. John, this has been a brilliant, brilliant discussion. Loved everything about the specific episode. So I'm really excited to.

release this. So thank you so much for coming on to the show and sharing your insights.

John Dulin (48:21.266)
Me too. Thanks man. If I ever come back, we'll have to get a buzzer for those quick questions. A couple of them were longer than 10 seconds. Yeah, but it was a lot of fun. Yeah.

Prateek Joshi (48:30.675)
Perfect.