Raw transcript — "Palmer Luckey: Autonomous Weapons Are Ancient and Why Anduril Won't Build Humanoids | EP #295"

Auto-generated (YouTube auto-sub, en) via yt-dlp, cleaned with vtt-to-text.py. Note: the auto-caption track shows signs of machine translation/re-translation (stilted phrasing, some garbled proper nouns) — treat as a rough transcript, not a verbatim record. Timestamps in [HH:MM:SS] mark VTT cue boundaries, not necessarily topic changes.

Recorded live at Moonshots Live 2026 (Sep 25, 2026). Panel: Peter Diamandis, Salim Ismail, Dave Blundin, Dr. Alexander Wissner-Gross ("the mates"). Guest: Palmer Luckey (Anduril Industries, ex-Oculus VR).

Wow, it's much brighter here. That's right. Well, you know, the future is bright. There are lens flares in the future. Yes. So, congratulations to everyone at the Oscars of optimism. So, of course, we open the summit with the king of autonomous weapons systems. OK. Let me start by saying that this week we are awarding three X Prizes. Palmer, congratulations on winning our X Prize competition on wildfires. Thank you very much. We worked on this for a long time. Come on, let's bring the slide up on the screen. So, Palmer won $2.2 million for Anduril. As if you needed them. Um, a million out of a million of that amount is from Lockheed. Which is quite interesting. We work a lot with

[00:01:01] Lockheed. We also compete with them, but in the defense sector, you know, we compete with their fighter division, but we cooperate on some of their payloads, so we do a lot of things together. But you and I have both been trying to get this autonomous firefighting thing going for a long time. You announced this award back in 2019. A lot of time has passed. One of Anduril's first two products was an autonomous firefighting tank, which we started building in 2017 and completed in 2019. It's crazy that we allow fires to become natural disasters before we put them out. And it's kind of ridiculous that people say, "Oh, you know, technology finally made this possible." And I don't know, I won't be correcting them publicly too often, but I mean, it's been possible for over a decade. Yes. It has simply become politically impossible to ignore this problem now. Now let's watch a one-minute video about what your team has achieved, and can you comment on it? Of course. So, this was an example of what we were

[00:02:00] doing at the range, where we were detecting and classifying fires. What isn't really visible here is that you also need to filter out false fires. This was also part of the tests. You have to distinguish between who is smoking a cigarette, where there is a bonfire, where there is a road flare, and where there is something that has the characteristics of a real fire. This is what you should point out to people. This is something that emergency services should be notified about. And then you need to choose the right extinguishing agent depending on the type of fire in the area. So, you want to use the most effective tool that can react and be on time. If the object is moving faster, something else may be needed. Depending on whether it is a forest or a hill covered in bushes, different solutions may be required. It's all about detection, classification, screening, selection of means, and then — reaching out to the target and defeating it. And lastly: in most of our projects, particularly for the military, you go through the same steps. But if

[00:03:00] you've found a goal, whether it takes you an hour to get there or 10 minutes often doesn't matter much. When it comes to a fire, the difference between a 10-minute and an hour-long response requires completely different tools. You know, a bomb is still a bomb, regardless of whether it takes a day, an hour, or a minute to deliver it. But when a fire spreads, you need to choose the right means to put it out before it gets too big. Surprisingly. We are going to announce the funding of the new XPrize in a month; until then, it's a secret, but now we'll lift the curtain a little. Another prize of $10 million. Will you share with us what we will be doing? Of course, we have to keep the details secret, because there are interesting nuances, but I'll say this: I'm not interested in

[00:04:00] spending money on racing with artificial intelligence. I believe that the field of artificial intelligence is already well understood by many people. I think we don't do enough studying biological intelligence, especially outside of humans. So this prize will be dedicated to understanding how animals think, so that we can communicate with them ten times better than the current level allows for the most intelligent species. And not just with trained animals, but with wild ones — in their own language. Surprisingly. I think we will be able to learn a lot of new things. If you look at the major language models, they are based on a distilled way of thinking and communicating of humanity. I believe we can learn a lot from all the other species on this planet. And I wouldn't be surprised if this finds applications in AI, especially where greater efficiency is needed, you know,

[00:05:01] energy efficiency, time efficiency, a different way of looking at things. One head is good, but two is better. Whatever metaphor you want to use. I like it. Okay, I'll start with the first official question. Probability of death from AI (P(doom)) — you are in the AI industry, in the defense industry. What is your assessment of the probability of death from AI? I'm just not worried about AI destroying us, I mean I'm not worried about AI killing. I'm really, really worried about evil people, who use moderately competent AI to commit evil deeds. And I'm a little disappointed in a lot of people who just got into this hobby of predicting doom. It's like how we dealt with the war in Ukraine, when suddenly all these people who had the most absurd ideas about the army their whole lives, and people who believed that we were living at the end of history, that we had overcome conflicts, that we lived in a world

[00:06:00] where you could just write an angry letter to the UN, forgetting that it only works when backed up by a real threat of violence. And then they suddenly change their mind. You see it with AI. People who haven't thought about this problem for more than a few years or maybe even a few months are saying, "Oh my God, this is the biggest threat to humanity." Those who have spent almost no time analyzing what the biggest threats to humanity are. They did not speak to any experts in the field. They certainly weren't buying freeze-dried food or building bunkers like me and many of my friends, who indulged in fantasies about what the end of the world might look like, and also actually thought about it. So yes, the likelihood of death from AI is way down the list for me. I am a great optimist. I think this will significantly reduce the cost of extracting, processing, and converting resources. You will see every part of society become

[00:07:00] subject to the same economies of scale and efficiency that industries automated without AI have already seen. Well, think about agriculture or the textile industry. You can buy a 2-week supply of calories for less than 1 hour of federal minimum wage. This is madness. Let's give Palmer a round of applause. You will see it in everything. Dave. So, I have a very, let's say, urgent question. I specialize a bit in systems, chips, and basic AI models. And so when you put systems in the field, weapons in the field, you can put intelligence into them. And that has always been a big problem for the Air Force. For example, if we start using drones instead of pilots, who will use the weapons and how will this decision be made? And the downside of transmitting data back to headquarters is, in fact, the delay. And the plus is that you have much more computing power. What you can put locally in a device, in the field, in an airplane, is limited, but it is

[00:08:00] very, very fast. So what's your philosophy on, you know, I think a lot of people are going to ask, "So what, we're going to have autonomous decision-making systems roaming around hostile territory?" Will we have, you know, control from headquarters? More or less. The real concern is not so much about latency or computational speed. The problem is that if you rely on a central location for computing, you now need to have a management system that provides reliable high-speed communication between you and potentially thousands or tens of thousands of active objects. And if the enemy jams your communication, intercepts it, or breaks it, you're finished. Or worse — they simply won't aim at the equipment in the field, but will go to where it's all controlled. You know, we saw it. There was a documentary that George Lucas was working on

[00:09:00] , and the Trade Federation blew up their central AI control, and that's how the Battle of Naboo ended: all the battle droids lost power. That's right, they all fell, and that's why I believe that this is the practical reality that dictates the need to have autonomy not in the rear, but on the front lines. There are other realms, like underwater, where you can't connect to anything. And sometimes you can connect, but you don't want to. Using radio radiation inevitably exposes you. So, whether you like it or not, you're going to have to do it. I don't really think it's as big a change as many people think. Um, there are a lot of people who look at this and say, "Oh my God, Pandora's box is opening. You will have weapons that will make their own decisions about killing." But they don't realize the reality, which is that this has been going on for decades. You look at the Aegis system on destroyers. It has a fully autonomous mode. You can press a button and it will shoot down anything that comes within a programmed radius, without any

[00:10:01] human confirmation. Look at SeaWiz and SeaRAM for shooting down missiles, mortar shells, artillery fire, or approaching boats — same thing. You can put them in a completely autonomous mode. And, I mean, ever since Vietnam, we've had missiles that could be launched from an airplane; they flew over the horizon, tried to identify radio emitters that corresponded to the positions of anti-aircraft missile systems, and then hit them without human intervention. The person just decided: "I'll send her to this area." They will be looking for this signal, and I know there is a certain probability that they will get it right. That decision will be up to the people, but autonomous weapons are a separate big conversation that we won't have today, although I have a lot of thoughts on the fact that autonomous weapons have actually been around for thousands of years. If we consider it as a weapon capable of carrying out the intent of whoever programmed it, regardless of the operator or engineer, look, for example, at pit traps that allow people

[00:11:00] to pass through but trap a commander on horseback as he rides by because he is heavier. You look at certain poison traps that killed people with certain characteristics without harming other people or animals. You will look at many types of land and sea mines and even torpedoes. So it all depends on your exact definition, even if you think of myths where people created golems that could go out and fight on their behalf or defend their city. I guess my argument is that we've had autonomous weapons for a very, very long time. It's just a new set of technologies that makes it better. So, the technical question about this is: the basic models that are embedded in these remote systems, or in pitfall traps — now the open source models are all Chinese. So I'm sure you won't introduce Chinese models into remote weapons. Are you going to create your own basic Anduril models or how are you going to implement AI into these

[00:12:00] devices? Listen, we've been doing this since 2017. So we are way ahead of everyone. We are already decades ahead of this issue. Uh, I think the basic models that people are creating now are not the models that are suitable for this. They are largely non-deterministic. They cannot be verified. This is a completely different category. What we build, unfortunately, has very few consumer applications. And vice versa. Sure, there are use cases for large language models in defense that make sense, but they don't guide my missiles or classify radar data. That is, we cannot create models that receive text and output text. I need to process raw radar data. I need to process sonar data from passive listening stations on submarines. This is a very specialized field. And I take some credit for that. All these people talking about how AI will change everything are mostly recent neophytes. Anduril Industries,

[00:13:02] you may have noticed that the abbreviation is AI. So it was no accident at all. OK. Alex, join in. So, I think, continuing our discussion about driving, we're now in an era where you can take advanced models, implement them in robots, and get some pretty amazing results. So I would like to ask you a question about the scaling hypothesis in relation to autonomous weapons systems or autonomous systems in general. Extrapolating, does Moore's Law exist? Is there an experience curve for how the capabilities of basic advanced models evolve, whether from labs or AI, or are we seeing some equivalent of Moore's Law for autonomous systems themselves? And maybe the second part of the question is: do you think we will eventually achieve some kind of convergence? You at Anduril create different incarnations in

[00:14:00] different form factors. Do you ultimately see some kind of universal killer robot category that would be analogous to a generalist robot, just as we now have generalist models that can potentially perform any task? No, I don't think so about the defense sector. I think it's possible for manufacturing, hospitality, and medicine because there's an economy of scale that's really important for those types of businesses. And you know, a humanoid robot might not be the optimal shape for frying burgers. And this may not be the optimal form for a firefighter. This may not be the optimal form for a nurse or assembly technician. So you just smile. Even if it's not optimal, the fact that the shape is similar to a human's makes it a very, very successful form factor. And now I can create one robot that will perform all these tasks. And I would probably be better off choosing a robot that is 5% slower or, say, less reliable than the

[00:15:00] optimal burger-frying robot if I can unify it with a robot nurse. In the defense sector, this is not entirely true. Because you are constantly participating in races where you need to not just barely outrun your opponent, but significantly outpace him. The goal is not to win the conflict only minimally. The goal is to act dishonestly in order to have an overwhelming threat of force that underpins any diplomatic overtures or negotiations. And everyone was always pushing the boundaries of what was possible. So, if I have a jet that consistently has a 5% advantage over the enemy, it's worth paying big money for, right? And I think the loss of productivity due to lack of specialization probably makes this approach unviable. Taking this to the extreme, for example, I'm building robotic submarines that are specialized in shape to

[00:16:00] dive to depths of 6,000 meters and stay on the ocean floor for months. And I also have other systems that work on the surface, these are robotic installations that drive up to the hills and wait. There are also other things, like aerial vehicles that perform very specialized tasks, and I just can't imagine a convergence that would make sense. Unless you're saying we're going to get something cool, like humanoid robots with built-in fusion reactors capable of diving to depths of 6,000 feet. Why don't you build these? I mean, if anyone were to do it, you... it's... You could potentially build something like that, but you'd quickly realize that specialization would allow you to build something much 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 dive to the bottom of the ocean." So, I think, I think defense is going to be an area where you look at, for example, the difference between cars, where everything has common platforms within the same manufacturer, they also use a lot of the same

[00:17:00] standards and interfaces. I think you're comparing it to motorsport, where everything is completely different and everything is constantly moving forward. I think the defense sector will probably always be this weird industry of exotic, highly specialized robotic systems. Palmer, I expected that if anyone was going to build humanoid robots, it would be you. No, no, you are wrong. You said — absolutely not. I gave up on it because I believe in them. If I didn't believe that humanoid robotics had such a powerful future, I would probably build them. But I want Anduril to create things that wouldn't otherwise exist. That's what I generally want to do. I want to maximize my marginal impact, but they can do the right thing. All the other people are already doing great. I want to use their robots and do things with them. I think humanoids will be an incredible common AI interface between AI and remote systems, as well as many legacy platforms designed for a human to push a lever, a

[00:18:00] button, or read a screen. We will be able to take all of this and automate it without completely replacing it. So you have a lot of things like radar systems that cost millions, tens of millions, even hundreds of millions of dollars. No one wants to serve them during a heated conflict because they are the first to be hit. But if I can put humanoid robots there, I can automate them without that risk. So, yeah, I won't be building humanoids. This is also the reason why Anduril doesn't build small quadcopters, for example. Too many other companies do this perfectly well. There are, there are many, many industries. And we will not engage in space launches. We create space products. We have many, we operate many of our own satellites and many of our own sensors in space. But I don't want to be a launch service provider. This is for all the other billionaires. Yes, yes. Let them do it. This is fantastic. Salim, you know when you run Anduril Industries, you decentralize defense, right? The Pentagon and defense ministries around the world were the most centralized structures in human history. You are on the

[00:19:00] front lines and dealing with this. You probably have a huge problem with the organizational immune system, where people say, "No, we've done this before." "We will do it better." They are either trying to pull you in or push you out. How do you overcome this barrier, because you do it in a way that, in my opinion, no one has ever managed to do before. I think this is a very, very difficult problem at the initial stage. [interruption — an XPRIZE announcement over the intercom] What were we talking about? Pentagon centralization versus decentralization. I will keep talking until someone tells me to stop. So, we've made some progress on this. I think there was a fear at the beginning that Anduril wouldn't be able to get through this

[00:20:01] . Our most difficult problem was raising money. I've been to many such meetings and I know how it works. One of our early problems was that our investors would say, "Palmer, potential investors." They said, "Listen, we believe you can create things that work." We believe that you can create great products, but we don't believe that the government is capable of buying them on a large scale, because you talk about centralization — a few years ago, about 80% of MDAP, the main defense acquisition programs, fell to five companies, and that's it. So, practically everything that matters. We observed incredible consolidation and centralization. Anduril has succeeded because we are less of a contractor and more of a product company. We use our own money to build products,

[00:21:00] then we sell them, and now we've gotten to the point where people in government have really won by betting on Anduril, and somehow we're completing everything on schedule, on time, and then the next program manager says, "I know how to make my job easier. I just gave the contract to Anduril." So, it takes courage to enter the defense industry, right? The most nepotistic, closed-off system on the planet. And what madness gave you the courage to do that? It was this madness that made me do this. The first page of our presentation said, "Anduril will save taxpayers hundreds of billions of dollars a year while earning tens of billions a year." The idea is that we're not just building products to sell, we're replacing systems that unnecessarily cost 10 times more than they should. And yes — it really seems crazy, but I can't put myself on the

[00:22:00] same level as him, but one of my mentors and idols in my childhood and early career was John Carmack, the founder of id Software — the inventor of modern 3D game engines, first-person shooters, and he also founded a rocket company that competed in the X Prize, and he was one of the few, if not the only one, to do a vertical takeoff and landing which no one else had even attempted. And I remember this name on board his devices — Nvidia. He built his own sensory integration algorithms and processed them on modern commercial graphics hardware, using his experience in game programming. The rocket company was called Armadillo Aerospace. And this thing — they were doing vertical takeoff and landing long before

[00:23:00] SpaceX even started doing anything, before it even existed. And, anyway, the point is — he eventually became the CTO of Oculus VR. I managed to lure him away from id, but after he left, he started working on AI even before me. And so he founded his own company to work on artificial general intelligence (AGI). And I said, "Why are you doing this? You know the chances of success," and just to be clear, I still believe that AI is going somewhere, but he said, "I'm going to start a company that's just going all in for the bottom line. I'm not trying to gradually get to this point. This is simply a bet on general-level AI." And he emphasized that: "Palmer, even if I have a 1% chance of success, which is perhaps an exaggeration, but let's say I have a 1% chance, and let's say I can do it, the risk-weighted outcome seems to require me to do this." And that's exactly how

[00:24:00] I felt about the defense sector. If there was a 1% chance that I could help solve this problem, then that would be an extremely important task, unlike, say, starting another gaming company, launching space rockets, or building a car company. After all, all billionaires end up doing the same thing. We don't have enough people who are on defense or donning the Batman suit. It's just cars and rockets, and they just go in circles, over and over again. So, last time I interviewed you for the Moonshots podcast, you mentioned that you were part of a billionaire chat group, a sort of "men's club." Are they working on anything noteworthy at this point? I criticized them — I said, "My God, you have so much capital, solve the world's biggest problems." I think the "billionaire guys" are doing much better now than they were a few years ago. They are becoming more active. And for those wondering who we're talking about: I've talked to a lot of people, especially some retired billionaires. And I said to them: "What are you

[00:25:01] doing? You have — because money is just one of the resources they have, but there's a resource that people don't talk about, especially with retired billionaires, which is their authority, which would allow them to rally people, teams, and capital around them. Um, they have authority because people trust them as leaders capable of leading such an organization. They can put forward the same ideas and say the same words as someone else, just as smart or smarter, and they will still achieve more because of it. So, I'm trying to convey to them that you have a responsibility not only as people who have significant capital, but also as those who have this underutilized resource — the ability to do more. If you don't do this, then what do you even do? Anyone can sit on a yacht. Few people can create another multi-billion dollar company that does something truly important. And the good news is that most of them — I think many

[00:26:00] — felt a desire to solve some of the existing problems. So nihilism is on the decline. We need more "tea boys" than there were "bee boys." Although, as I said, most of the "tea boys" seem to be quite busy. That's the point. I don't even have anything against "bee boys" who don't do enough. Some people misunderstand my words. They say Palmer wants other billionaires to work on defense, energy, or biology. I don't mean that people need to be distracted by a bunch of things. Many people can only do one thing. You know, that's what I did for a long time. But I can't stand those who do nothing at all. This is simply disgusting to me. Maybe a message to the audience: you can join the "B boys" group on Slack or somewhere on Telegram if you're creating something new and unique. The group was actually shut down, and we are now discussing that it is time to create "B boys 2.0". But

[00:27:01] I think people were posting inappropriate memes too often. There was a time when there weren't any girls there, and then this article came out in Business Insider whining about the subject — they say, "Oh, there's this group and they don't invite women there." And I say, "It's not that women aren't invited. It's just me and my friends." So, we need to restart that group. So, you're now starring in a Facebook ad. Yes, I am in Facebook advertising. VR is already out — how does it relate to the vision you wanted to achieve? Look, it does a lot of things, but it's not exactly what I would do if I were the one waving a magic wand and

[00:28:00] — listen, the new Meta VR headset is much closer in many ways. So I've long believed that we need to separate the computing unit and battery from the head. It shouldn't be like this, more like moved to the front. So, or [as] the Oculus Quest [did]. Oculus Quest was the last thing I worked on before I got fired. And the prototypes at that time were a series codenamed Santa Cruz, and we actually showed it publicly before I was fired. And the Santa Cruz had the battery and the computing unit at the back of the head, and the lightweight part that was adjacent to the soft tissues in front was as lightweight as possible. And then, at some point after they got rid of me, they crammed it all into one big massive unit in the front and ruined the ergonomics; they forgot why the top strap was needed. This worries me the most. This is what bothers me the

[00:29:00] most right now, because the Oculus Rift CV1, which was the first modern consumer VR headset, had more or less perfect ergonomics back in 2016; and we showed it in 2015, and literally no one has done it as well, and it's been over ten years. Everyone, absolutely everyone, ruins everything, even Apple, and I just can't understand it. How could it be that an ergonomic design led by a 19-year-old with no experience is better than anything anyone has come up with since then? In any case, the new Meta [Ray-Ban Display / VR] glasses are the closest thing to that. And they are very light. They took the computing unit outside. They took the batteries outside. They have a really high-quality end-to-end video mode. They have a feature that I invented for CV1 called "transform fabric", although they don't call it that. They actually use fabric, a soft flexible black matte fabric that absorbs light: firstly to minimize glare and secondly

[00:30:01] to seal the cavity between and around the lenses. This means they can be moved and the fabric simply stretches to accommodate the lens position for different interpupillary distances, instead of using sliding mechanical interfaces and plastic. And they brought back transform fabric for the first time since my time. So I'm a huge fan, and it's a third of the price of an Apple product and a slightly higher pixel density. I think that's quite extraordinary. And it's eight times lighter. Let's end this. We have 1,500 entrepreneurs, visionary innovators, and tens of thousands of viewers watching live. What is your advice? What did you learn in high school that allowed you to succeed so quickly? I was very lucky — well, I guess I was

[00:31:00] unlucky, I did it myself. But — well, I am lucky. It's even in the last name. I created an online community called Mod Retro when I was 14 or 15 years old. It was an online forum for people who modified vintage game consoles and computers, creating portable devices, improving their performance in various ways. We had several thousand active participants, and a whole team of volunteer moderators and administrators, and for some reason I was at the head of it all. And I learned a lot of valuable lessons about how to build an organization, how to promote it, how to get people to believe in what you're doing. I learned a lot about solving the problems you face when running a company, with people. It's a story as old as the world. You have two people. They are both brilliant. They are both very good at what they do. And they hate each other. And they disagree about something. You need to somehow

[00:32:00] resolve the conflict or separate them in a way that benefits them, so that both can contribute to the big project, but so that they don't hate you, lose faith in the common cause, and, of course, so that they don't kill each other. Sometimes this means you need to reconcile them with each other. Sometimes this means putting one here and another there. And sometimes you have to make sure they don't know you're doing it while you're doing it. And once it turns into a three-body problem, then — but that, like a hundred other things I learned while running Mod Retro, came in handy when I started Oculus and in my subsequent work at other companies. And that's why I firmly believe that children should find jobs. I started working myself at 14, so I also learned how the real world works. I'm a big fan of

[00:33:00] people doing things on a scale bigger than themselves. In other words, don't just do a project, do it with someone else. Do projects together with other people. Create openly, learn the strengths and weaknesses of this approach. This was extremely helpful to me. Has creating anything changed in the age of AI? What is your advice? As we've been saying on the podcast for a long time: forget about a master's degree, and even college is another question. You didn't go to college. Do you know what high school students should do? Find something you love and use AI to program it. What is your final advice on this? You know, one of the reasons I'm not as upset as some people are is because — [an on-screen interruption, panel declines to discuss it] — okay, great. There are many

[00:34:01] people who love their job and love what they do. And they quite rightly mourn the end of the era when, for example, code was written by hand, from start to finish, as a true craft. And I understand that. It's the same feeling that many artists had when the camera was invented. They complained that it ruined the art by allowing anyone to just press a button, and such changes are always painful for masters of their craft. I've been burned by this for a long time. When you're a founder, you don't work directly with technology anyway. Everyone else is enjoying themselves. And, you know, I like electrical engineering, I like optics, I like mechanical engineering. I don't do any of that, because the founder's job is to hire people to replace you and do what you thought you could do yourself, or what they do better than you. I often tell people: if you want to work with technology,

[00:35:00] work for someone else. If you want to constantly deal with nonsense, start a company. So I'll tell people this: AI doesn't really change much in the part of the job where there's a lot of nonsense. You are still legally responsible. You are still responsible for all interpersonal matters. You can delegate some things, but ultimately the decision is yours. It's your choice. The same applies to attracting investment. The same goes for the overall product architecture. And when people make bad decisions, it's you who chooses when to undo, change, or replace those people. I don't see AI replacing this part of the job anytime soon. Not because it can't, but because the efficiency gains in other areas are so much greater that we'll have a good 10 years of accelerated development of all other business processes before it's time for this. So, listen, AI is changing a lot of things, but I'm telling you: a lot of you are in for a bitter disappointment if you're an amateur programmer or an amateur engineer and you think you're going to start a company and do it all

[00:36:00] by yourself — no. A harsh lesson awaits you. Being a founder has always felt like AI took your job. It was just taken away by other people. Ladies and gentlemen, let's give Palmer Luckey a round of applause.