Palmer Luckey: Autonomous Weapons Are Ancient and Why Anduril Won't Build Humanoids | Moonshots Live | EP #295

0:04Peter Diamandis Wow! It's so much brighter out here. It is! The future is bright! There's lens flare in the future. Yeah... So, welcome everybody, to the Oscars of Optimism. Of course, we're opening up the summit with the king of autonomous weapon systems.

0:26All right. So I want to start with the fact that we're awarding three X-Prizes this week. Palmer, congratulations for winning our Wildfire X-Prize. Thank you so much. This one's been a long time in the works. Let's get the slide up here. So Palmer won$2.2 million for Anderle. As if you need it. A million of that from Lockheed, which was kind of interesting. We do a lot of work with Lockheed. We also compete with them, but in the defense space, we compete with their fighter jet division, and we partner with them on some of their payloads. So we do a lot of stuff together. But me and you have both been trying to get this autonomous firefighting stuff going for a long time.

1:15For a while. I mean, you got this prize announced back in 2019. It's been a long time. Andrel's, one of our first two products was an autonomous firefighting tank that we started building in 2017, finished in 2019. It's insane that we let fires become conflagrations before you put them out. And it's kind of insane that people are saying, oh, the technology's finally made this possible. And I'm not going to correct him too often in public, but it's, I mean, this has been possible for over a decade. It's just that it's become politically impossible to ignore the problem now. I'm going to watch a one minute video clip of what your team achieved and if you'd walk us through it here.

1:56Sure. This was an example of what we were doing at the test site where we were detecting fires, classifying fires. What you don't really see in here is that you also need to reject false fires. That was part of the test as well. So you have to know who's the guy smoking a cigarette? Who's a guy, what's a campfire? What's a road flare? And then what's something that shows the characteristics of real fire? That's what you need to flag to people. That's what you need to flag to emergency services. And then you need to matchmake the type of fire to the right asset that's in the area. And so you want to basically use the most cost-effective thing that can respond and get there in time.

2:35And if something's going faster, you might need something different. If it's in woods versus on a brush hill, you might need something different. And so it's about detection, classification, rejection, matchmaking, and then going and hitting the target. And the last thing I'll say is that with most of what we do, our work with the military, you go through many of these same steps. But if you find a target, let's say it takes you an hour to get to somebody versus 10 minutes. It usually doesn't make as much of a difference. When you're talking about fire, though, the difference between 10 minutes and an hour response, totally different tools you're going to need to respond with.

3:17A bomb is a bomb whether or not it takes a day or an hour or a minute to get there. But when a fire is metastasizing, you need to pick the right thing that can put it out before it gets too large. Amazing. You know, we're going to be announcing something in a month's time that you're funding a new XPRIZE, and it's going to be secret until then, but we're going to tease it now. Another$10 million prize. Would you share with us what we're going to do? Sure. You know, we should keep the details secret because there's some good twists, but suffice to say that I'm not interested in joining the race to push AI forward with my money.

4:06I think that the artificial intelligence side of models is pretty well covered by plenty of people. I think we've not been doing enough understanding biological intelligence, particularly beyond people. And so this is going to be a prize around understanding how animals think and making so that we can communicate with them 10 times better than the current state of the art for the current smartest species. And not just trained animals, but wild animals in their own language. Amazing. And I think that there's actually a lot that we're going to be able to learn. I mean, if you look at large language models, they're kind of this, you know, they're based on humanity's way of thinking and communicating distilled.

4:47And I think that there's plenty to learn from all of the other species that are on this planet. And I wouldn't be surprised if as applications for AI, particularly AI where you need to be more efficient, more power efficient, more time efficient, a different point of view, two peds are better than one, whatever metaphor you want to use. Love it, love it. All right, I want to kick off the first official question here. P-Doom, you're in the AI industry, the defense industry. What's your P-Doom? I'm just not worried about AI dooming. I guess I'm not worried about AI doing the killing. I'm very, very worried about evil people using moderately competent AI to do evil things.

5:35And I am kind of frustrated by a lot of people who are new to the dooming hobby. It's sort of like how we had to deal with, you know, the war in Ukraine happens, and all of a sudden, all these people who have had the dumbest opinions you've ever heard about the military for their whole life, and people who believed we lived at the end of history, that we had moved past conflict, that we lived in a world where you could just write a nasty letter in the UN, forgetting that that only works if it's backed up by credible threat of violence. And then all of a sudden, they change their minds. You're seeing this with AI.

6:13People who have not thought about this problem for more than a couple years or maybe even just a couple months are saying, oh my God, this is the biggest threat to humanity. These are people who have spent almost no time actually analyzing what the largest threats to humanity are. They haven't spoken to any of the experts in this area. They certainly weren't buying freeze-dried food and building prepper bunkers, you know, like me and many of my buddies who have indulged fantasies about what doom could look like and also thought realistically about it. And so, yeah, the AI doom is way down the list for me.

6:50I'm a big optimist. I think it's going to massively reduce the cost of resource extraction, processing, transformation. You're going to see every part of society become subject to the same economies of scale and efficiency that automated industries, that automated without AI have seen. So think like agriculture or textiles. You can buy two weeks of calories for less than one hour of federal minimum wage labor. That's crazy. Let's give it up for Palmer here. You're going to see that for everything.

7:24Peter Diamandis:Dave. So I got a really pressing question on my mind. I'm kind of a feeds and feeds and chips and AI foundation models guy. And so when you're putting systems in the field, weapons in the field, you can put the intelligence inside it. And this has always been a big problem for the Air Force. Like, if we're going to start flying drones instead of pilots, who's going to unleash the weapons, and how's that decision going to get made? And the downside of transmitting it back to headquarters is you have latency. The upside is you have a lot more compute. What you can put locally in a device, in a robot, in a field, in a plane, is limited, but it's really, really fast.

8:01Peter Diamandis:So what's your philosophy on, like, you know, I think a lot of people would say, well, are we gonna have self-contained, decision-making weapons wandering around enemy territories, or are we gonna have control at the headquarters? How does it work? More or less. The real concern isn't so much the latency or the speed of computation. The problem is that if you rely on a central location to do your compute, you now need to have command and control that has good, high bandwidth communication back and forth between you and potentially thousands or tens of thousands of active assets. And if your enemy jams your communications or intercepts your communications or hacks your communications, you're done.

8:44Or worse, they just skip targeting all the stuff in the field and they go for the place where it's all being controlled. We saw this. There was this documentary that George Lucas was working on. And the Trade Federation, they had their central AI command and control blown up. And that was how the Battle of Naboo actually resolved, was all the battle droids lost. Oh, yeah, they all fell over. That's right. They all fell over. And so I think that that's the practical reality that's driving the need to have autonomy living, not back home, but forward. There's other areas, like you go underwater, you can't talk to them.

9:23And sometimes you could talk to things, but you don't want to. Having radio emissions actually exposes you necessarily. And so for better or for worse, you're gonna need to do this. I actually don't think it's as much of a shift as people think. There's a lot of people who look at it and say, oh my God, Pandora's box is being opened, you're gonna have weapons that are making their own kill decisions. And they don't recognize the reality, which is that that's been the case for decades. You look at the Aegis system on destroyers, it has a fully autonomous mode. You can flip a button and it will shoot down anything that comes within whatever range you programmed it with Without any human confirmation You look at CWIS and C-RAM for shooting down incoming missiles incoming mortar incoming artillery fire incoming missiles or incoming speed boats Same thing you can put them in fully autonomous modes and I mean it going back to into Vietnam We had missiles you could launch from an aircraft They flew over the horizon tried to identify the right radio emitters that they believed corresponded with a surface-to-air missile site and then struck them without any, you know, the person had to decide, I'm going to send it in this area, they're going to look for this signal, and I know that there's this probability they're going to get it right.

10:34That decision is going to remain with people. But autonomous weapons are widely, widely, there's a whole other talk we're not going to have today, but I will, I do have a rant about this, where actually autonomous weapons have existed for thousands of years. If you look at, and defined as kind of weapons that can execute the intent of what you program to do independent from the operator or the engineer. You look at, for example, pitfall traps that allow people to walk over them, but then capture the commander on his horse when he walks by and it's heavier. You look at certain poison traps that would kill people who had certain characteristics, not kill people who had other characteristics or not kill animals.

11:16You look at many types of landmines and sea mines and even torpedoes. And so it depends on your exact definition, but even going into myth, where people are creating golems, like can you go out and fight on their behalf or defend their city? I guess my argument is we've had autonomous weapons for a very, very long time. It's just a new set of technology that's making them better.

11:41Peter Diamandis:So technical question on that, the foundation models that go into those remote systems, right now, Yeah, well, so right now the open source models are all Chinese. So I'm pretty sure you're not going to be putting Chinese models into remote weapons. Are you going to build Android foundation models or how are you going to get AI in those devices? I mean, we've been doing it since 2017. So we're way ahead of everybody. We're a decade beyond that question. I think that we also don't, the foundation models that people are building are not the types of models you want to be doing this. They're largely non-deterministic.

12:17They're non-auditable. It's a very different category. The thing that we are building, unfortunately, has very few consumer applications and vice versa. There, of course, are LLM use cases for defense that make a lot of sense, but it's not guiding my missile or classifying radar feeds. I mean, we can't build models that ingest text and then spit out text. I need to process raw radar data feeds. I need to process sonar data from passive listening stations on submarines. It's pretty highly specialized. And I will take some credit here. All these people who are talking about how AI is going to change everything, they are mostly recent converts.

13:01Andrel Industries, you might notice the acronym is AI. And that was not a mistake.

13:07Peter Diamandis:Alex, jump in. Yeah, I think maybe dovetailing on our discussion of driving bench, we're now in an era when it's possible to take frontier models, drop them into a robotic embodiment, and get pretty stunning results. So I'd love to ask you a question maybe about the scaling hypothesis as it applies to autonomous weapon systems or autonomous systems in general. If you have to extrapolate out is there a Moore's law, is there an experience curve for what, as obviously the capabilities of underlying frontier models, whether they're from frontier labs proper or from AI, as those progress, do we see any equivalent of a Moore's Law for the autonomous systems themselves?

13:50Peter Diamandis:And maybe the second half of the question would be, do you think ultimately we achieve some sort of convergence? Andrel makes a variety of embodiments and a variety of different form factors. Do you see ultimately some general purpose category killer robot that is equivalently the generalist robotic embodiment, just like we now have generalist models that are capable of doing substantially any task? No, I don't think so on the defense side. I think that for manufacturing and for hospitality and medicine, you might see this because you get economies of scale that are really material for those types of businesses.

14:30And a humanoid robot may not be the optimal form for a burger griller, and it may not be the optimal form for a firefighter. It may not be the optimal form for a nurse or an assembly technician. Salima's smiling. But even if it's not optimal, the fact that it is shared with humanity in its form makes it a really, really good form factor. Now I can make one robot that does all of these things, and I would probably rather have a, let's say, robot that's 5 % slower or less reliable than the optimal burger-flipping robot if I can share commonality with that nurse. In defense, that's not really the case.

15:11Because you are in this constant race to not just barely beat your adversary, but to massively overpower them, to try and make it where, the goal is not to just barely win a conflict on the margins. It is to fight unfair, to have overwhelming threat of violence, underpinning whatever diplomatic overtures or negotiations you have. And everyone's always pushing the limit. So if I have a jet that has a reliable 5 % every time edge over the enemy, that's worth paying a lot of money for, right? And I think the loss in performance you would see through lack of specialization probably makes it non-viable.

15:53Like to go to the extreme, like, you know, I build robot submarines and those have a specialized form that needs to dive to 6 ,000 meter depth and sit on the bottom of the ocean for months at a time. And then I have other stuff that's on the surface and, you know, they're robotic systems that, you know, roll up a hill and they loiter. And then I have other things that are, you know, aerial vehicles that are doing very specialized tasks. And I just can't imagine a convergence that would make sense. unless you're telling me that we're going to get something cool, like humanoids that have onboard fusion reactors, and they can dive to 6 ,000 feet depth.

16:26So, Palmer, why aren't you building those? I mean, if anybody, you should... Look, you could potentially build such a thing, but then very quickly you realize that specialization would allow you to build something even better. You say, okay, I've built something that can dive to the bottom of the ocean and fly to the moon. Man, it could get to the moon a lot faster if it didn't have to also dive to the bottom of the ocean. So I think defense is going to be this one where kind of like, you know, you look at, for example, the difference between cars, where everything has shared platforms within a manufacturer, you know, manufacturer, and they also share a lot of the same standards and interfaces.

17:03I think that you compare them with, like, motorsport, where it's all completely different, and it's continuously pushing forward. I think the defense side is probably always going to be this weird branch of exotic, highly specialized robotic systems. Palmer, I expected if anybody was going to be building humanoid robots, it would be you. No, no, the opposite. You said absolutely not. I've opted out because I believe in them. If I didn't believe that humanoid robotics had such a powerful future, then I would probably build them. But I want Anderil to build things that wouldn't exist otherwise. That's what I generally want to do.

17:37That's a good lesson.

17:38Peter Diamandis:You want to maximize your marginal impact. Correct. All the other people are already doing a great job. I want to use their robots and do things with them. I think humanoids are going to be an incredible general artificial intelligence interface between AI and remote systems and a lot of legacy platforms that were designed for a person to pull the lever, push the button, read the screen. We're going to be able to take all those and automate them without replacing them entirely. So you have a lot of things like radar systems that cost millions, tens of millions, even hundreds of millions of dollars.

18:12Nobody wants to man them during a hot conflict because they're the first thing that's going to go. But if I can put humanoid robots in them, I've now been able to automate them without that. So, yeah, I'm not going to build humanoids. This is also why Andrel is not building, like, small quadcopter attack drones. Too many other companies doing a great job. There's a lot of areas. And we're not going to do space launch. Like, we do space products. We operate a lot of our own satellites and a lot of our own sensors in space. But I don't want to be a launch provider. That's for all the other billionaires to...

18:41Yeah, yeah. Let them handle that.

18:42Peter Diamandis:That's fantastic. Salim. You know, when you're operating Android Industries, you're decentralizing defense, right? The Pentagon and ministries of defense globally have been the most centralized entity in the history of humanity. You're sitting on the edge dealing with that. you must have a massive immune system problem where people inside the entity are going, no, we've done that before. We can do it better. They either try and suck you in or they try and spit you out. How are you navigating that interface? Because you're doing it in a way that I've never seen anybody succeed. Do it. I think that this is a problem that's very, very hard at the beginning.

19:26Peter Diamandis:I said the wrong thing. Wow. No idea. Well, until someone gets on the intercom and tells us to go. The AI is a contract. What were we talking about? What was it?

19:44Seriously? I knew the intercom was going to come on. What were we talking about again? Centralization of Pentagon versus decentralized. I'm going to keep talking. Just keep speaking. So we've gotten through this a bit. I think early on, there was that concern that Anderil was not going to be able to get through that. Our hardest problem raising money. Skippy, would you please shut this down? I've been through enough of these to know how it's going to work. There we go. I imagine. One of our early problems was that our investors said, Palmer, potential investors, they said, look, we believe that you'll be able to build things that work.

20:19We believe you'll be able to build products that are great, but we don't believe the government is capable of actually purchasing them at scale. Because, I mean, you talk about centralization. you know about the issues for people who don't. Until just a few years ago, you had about 80 % of MDAPs, Major Defense Acquisition Programs. So basically anything that matters. 80 % of MDAPs went to five companies and that's it. And I think you've seen incredible consolidation, centralization. Anduril has been successful because we use, we're less of a contractor and more of a product company. We use our own money to build products, then we sell them.

20:59And we've gotten to the point now where people in government have done very, very well betting on Anderil. And then miraculously, we finish on schedule, we finish on time. And then the next program manager says, I know how I can make my job easy. I just give the contract to Anderil. Got it. So it's got to take guts to enter the defense industry, right? The most incestuous, you know, self-licking ice cream cone on the planet. Yeah. And what insanity gave you the balls to do that? The insanity was the reason that I got into it. The first page of our pitch deck said, Anderil will save taxpayers hundreds of billions of dollars a year by making tens of billions of dollars a year.

21:43The idea being, we're not just making products that we're selling, but we are displacing the systems that cost, needlessly, ten times what they should. And yeah, thank you. And so it does seem crazy, but you know, I can't put myself on the same pedestal as him, but one of my mentors and idols growing up, and then also early in my career was John Carmack, the founder of it's software. John is great. John is incredible. And for people who don't know, he's basically the inventor of the modern 3D game engine, first person shooters, also started a rocket company that competed in the - He was competing in the XPRIZE.

22:21That's right. And I remember... And he was one of the, if not the only one, to be doing reusable vertical takeoff and landing. Exactly, he was doing vertical takeoff, vertical landing. Which nobody else was even trying to do, and he did. I remember this name on the side of his vehicle, NVIDIA. How do you pronounce that thing? If I had only known. That's right. He was using... He was building his own sensor fusion algorithms. He was processing them on modern, or on commercial style graphics card hardware, using his game programming experience. Anyway, the point is... It was called Armadillo. Armadillo Aerospace.

22:54Yeah. That's right. And the thing, and yeah, they were doing vertical takeoff and landing long before SpaceX was, I mean, doing anything. Yeah, before it existed, for sure. That's right. But anyway, the point is, he ended up becoming the CTO of Oculus VR. I managed to poach him away from id, but after he left, he went to work on AI actually even before I did. And so he started his own company to work on AGI. And I said, why are you doing this? The odds of you succeeding, and to be clear, I still believe AI was going somewhere, but he said, I'm going to start a company that's just going all in on the end state.

23:38I'm not trying to incrementally get there. It's just all in on AGI. And the point that he made was, Palmer, even if I've got a 1 % chance of succeeding, which maybe is an overestimation, but suppose I have a 1 % chance of succeeding, and suppose that I could pull it off. The risk-adjusted outcome kind of demands that I do it, and that was how I felt about defense. If there was a 1 % chance that I could help solve this problem, then it was one that would be massively impactful to work on versus, let's say, starting another gaming company, getting into space launch, building a car company. All of the billionaires, they end up doing the same thing.