YT YouTube2Text

YouTube2Text → Transcripts → Palmer Luckey: Autonomous Weapons Are Ancient and Why Anduril Won't Build Humanoids | EP #295

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

by Peter H. Diamandis · 7,241 words · 1,050 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:05Wow, it's so much brighter out here.
0:07>> It is. It's, you know, the future is
0:10bright.
0:10>> There's lens flare in the future.
0:12>> Yeah. So, uh, welcome everybody to the
0:15Oscars of optimism that, of course,
0:17we're opening up the summit with the
0:20king of autonomous weapon systems.
0:22[laughter]
0:26All right. So, uh, I want to start with
0:30the fact that we're awarding three X
0:32prizes this week. Palmer,
0:34congratulations for winning our, uh,
0:38Wildfire X-P prize.
0:39>> Thank you so much.
0:40>> Yeah, [applause]
0:42>> this one's been a long time. And in
0:44>> it has Let's get Let's get the uh the
0:46the slide up here. So, Palmer won $2.2
0:49million for Ander
0:51>> as if you need it. um a million of
0:55>> a million of that from Lockheed,
0:57>> which was kind of interesting.
1:00>> We do a lot of work with Lockheed. We
1:02also compete with them, but we in the in
1:04the defense space, it's uh you know, we
1:06we compete with their fighter jet
1:08division and we partner with them on
1:09some of their payloads. And so we do a
1:11lot of stuff together. But me and you
1:12have both been trying to get this
1:13autonomous firefighting stuff going for
1:15a long time. For a while.
1:16>> I mean, you got this prize announced
1:17back in 2019. It's been a long time. And
1:21one of our first two products was an
1:23autonomous firefighting tank that we
1:25started building in 2017, finished in
1:272019.
1:28>> It's insane that we let fires become
1:30conflrations before you put them out.
1:32>> And well, and and it's kind of insane
1:34that uh people are saying, "Oh, that you
1:36know, the technologies finally made this
1:38possible." And I I'm not going to
1:39correct him too often in public, but
1:41it's I mean, this has been possible for
1:43over a decade. It's just that uh
1:45>> let's watch it's become politically
1:47politically impossible to ignore the
1:48problem. out. I'm going to watch a one
1:50minute video clip of what your team
1:52achieved and if you'd walk us through it
1:54here.
1:56>> Sure. So you this was uh an example of
1:59what we were doing at the test site
2:01where we were uh detecting fires,
2:04classifying fires. What you don't really
2:06see in here is that you also [music]
2:07need to reject false fires. That was
2:09part of the test as well. So you have to
2:10know, you know, who's a guy smoking a
2:11cigarette, who's a guy, what's what's a
2:13campfire, what's a road flare, and then
2:15what's something that shows the
2:16characteristics of real fire? That's
2:18what you need to flag to people. That's
2:21what you need to flag to emergency
2:22services. Uh and then you need to
2:25matchmake the type of fire to the right
2:28asset that's in the area. And so you
2:30want to basically use the most cost
2:31effective thing that can respond and get
2:35there in time. And if something's
2:36growing faster, you might need something
2:37different. If it's in woods versus on a
2:40brush hill, you might need something
2:41different. And so, you know, it's it's
2:43it's about detection, classification,
2:45rejection, matchmaking, and then going
2:47and and hitting the target. And the last
2:49thing I'll say is that with most of what
2:51we do, uh, our work with the military,
2:55you go through many of these same steps,
2:58but if you find a target, let's say it
3:00takes you, you know, an hour to an hour
3:03to to get to somebody versus 10 minutes,
3:06it usually doesn't make as much of a
3:08difference. When you're talking about
3:10fire though, the difference between 10
3:12minutes and an hour response, totally
3:15different tools you're going to need to
3:16respond with. You know, a bomb is is is
3:19a is a bomb whether or not it takes a
3:21day or an hour or a minute to get there.
3:23Uh but when when a fire is is is
3:25metastasizing, you need to pick the
3:27right thing that can put it out before
3:29it gets too large.
3:30>> Amazing. you know, uh, we're going to be
3:33announcing something in a month's time,
3:36uh, that you're funding a new X-P prize
3:39and, uh, it's going to be secret until
3:41then, but we're going to tease it now.
3:44Uh, another $10 million prize. Would you
3:48share share with us what we're going to
3:49do?
3:50>> Sure. you know, uh, the, you know, we we
3:54should we should keep the details secret
3:55because there's there's some there's
3:57some good twists, but suffice to say
3:59that, uh, I'm not I'm not interested in
4:01joining the race to push AI forward with
4:06my money. I think that the the the the
4:08artificial intelligence side of models
4:10is is is pretty well covered by plenty
4:12of people. I think we've not been doing
4:14enough understanding biological
4:16intelligence, particularly beyond
4:18people. And so this is going to be a
4:20prize around understanding how animals
4:23think and making so that we can
4:25communicate with them 10 times better
4:27than the current state of the art for
4:28the current smartest species. And not
4:30just trained animals but wild animals in
4:32their own in their own language.
4:34Amazing. [applause and cheering]
4:36And I think that there's actually a lot
4:37that we're going to be able to learn. I
4:38mean if you look at uh large language
4:40models, they're kind of this, you know,
4:42they're they're based on uh humanity's
4:45way of thinking and communicating
4:47distilled. And I think that there's
4:50plenty to learn from all of the other
4:52species that are on this planet. And I
4:53wouldn't be surprised if it has
4:55applications for for AI, particularly
4:57AI, where you need to be uh more
5:00efficient, you know, more power
5:01efficient, more time efficient, uh you
5:04know, a a different point of view, two
5:06heads are better than one, whichever uh
5:08whatever metaphor you want to use.
5:10>> Love it. Love it. All right. I'm going
5:11to kick off the first official question
5:13here. Poom, you're in the AI industry,
5:16the defense industry. What's your P
5:18doom?
5:20I'm just not worried uh about AI dooming
5:26the I guess I'm not worried about AI
5:28doing the killing. I'm very very worried
5:30about evil people using moderately
5:34competent AI to do evil things.
5:36>> Exactly.
5:36>> And I am
5:39>> kind of frustrated by a lot of people
5:41who are new to the dooming hobby. Um,
5:45and and it's it's it's sort it's sort of
5:47like how we had to deal with uh you the
5:50war in Ukraine happens and all of a
5:51sudden all these people who have had the
5:53dumbest opinions you've ever heard about
5:55the military for their whole life. Uh,
5:58and people who believed we lived at the
6:00end of history, that we had moved past
6:02conflict, that we lived in a world where
6:04you could just write a nasty letter in
6:07the UN, forgetting that that only works
6:08if it's backed up by credible threat of
6:10violence. And and all of a sudden they
6:12change their minds. You're seeing this
6:12with AI. People who have not thought
6:15about this problem for more than a
6:18couple years or maybe even just a couple
6:20months uh are are saying, "Oh my god,
6:22this is the biggest threat to humanity."
6:24And these are people who have spent
6:25almost no time actually analyzing what
6:28the largest threats to humanity are.
6:30They haven't spoken to any of the
6:31experts in this area. They certainly
6:33weren't buying freeze-dried food and
6:35building prepper bunkers, you know, like
6:37like like like me and many of my my my
6:40buddies uh who who have indulged
6:42fantasies about what doom could look
6:45like and also thought realistically
6:46about it. And so uh yeah, the AI doom is
6:49way down the list for me. I'm I'm I'm a
6:51big optimist. I think it's going to
6:52massively reduce the cost of resource
6:55extraction, processing, transformation.
6:57You're going to see every part of
6:58society become subject to the same
7:01economies of scale and efficiency that
7:03automated industries that automated
7:05without AI have seen. So think like
7:08agriculture or uh or textiles. You can
7:11buy 2 weeks of calories for less than
7:16one hour of federal minimum wage labor.
7:19That's crazy.
7:20>> Let's give it up for for Palmer here.
7:22>> You're going to see that for everything,
7:24>> Dave. [applause]
7:26So, I I got a really um pressing
7:28question on my mind. I'm kind of a feeds
7:30and feeds and chips and AI foundation
7:32models guy. And so, uh when you're
7:34putting systems in the field, weapons in
7:36the field, uh you can you can put the
7:39intelligence inside it, and this has
7:41always been a big problem for the Air
7:42Force, like if we're going to start
7:43flying drones instead of pilots, who's
7:45going to unleash the weapons, and how's
7:46that decision going to get made? And the
7:48the downside of transmitting it back to
7:50headquarters is you have latency. Uh the
7:53upside is you have a lot more compute.
7:55uh what you can put locally in a device
7:57in a robot in the field in a plane is
7:59limited but it's really really fast. So,
8:01what's your philosophy on like, you
8:03know, I think a lot of people would say,
8:04well, are we going to have
8:06self-contained decision-making weapons
8:08wandering around enemy territories or
8:10are we going to like are we going to
8:12have, you know, control at the
8:13headquarters? How does it work?
8:14>> More or less. The the real concern isn't
8:16so much the latency or or or the speed
8:19of computation. The problem is that if
8:21you rely on uh you know central a
8:25central location to do your compute, you
8:27now need to have command and control
8:30that has uh good high bandwidth
8:32communication back and forth between you
8:34and potentially thousands or tens of
8:36thousands of active assets. And if your
8:38enemy jams your communications or
8:40intercepts your communications or hacks
8:42your communications, you're done. uh or
8:45worse they just skip targeting all the
8:48stuff in the field and they go for the
8:50place where it's all being controlled.
8:52You know, we saw we saw this I there was
8:54a there was a there there was there was
8:56this documentary that George Lucas was
8:58working on and and the Trade Federation
9:00they had they had their central AI
9:02command and control blown up and that
9:04was how the battle of Ne of Naboo
9:06actually resolved was all all the battle
9:08droids lost.
9:09>> Oh yeah. They all fell over.
9:10>> That's right. and they all fell over.
9:11And so I think that that's that that's
9:13the practical reality that's driving the
9:15need to have autonomy living not back
9:18home but forward. There's other areas
9:20like uh you go underwater, you can't
9:22talk to things
9:24>> and sometimes you could talk to things
9:26but you don't want to. Having radio
9:27emissions actually exposes you
9:30necessarily and so you for better or for
9:33worse you're going to need this. I
9:34actually don't think it's as much of a
9:35shift as people think. Um, I there's a
9:38lot of people who look at they say, "Oh
9:39my god, Pandora's box is being opened.
9:41You're going to have weapons that are
9:42making their own kill decisions, and
9:44they don't recognize the reality, which
9:46is that that's been the case for
9:47decades. Uh, you look at the you look at
9:49the the you look at the Aegis system on
9:51destroyers. It has a fully autonomous
9:54mode. You can flip a button and it will
9:55shoot down anything that comes within
9:57whatever range you programmed it with uh
10:00without any human confirmation. Uh you
10:02look at SeaWiz and CRAMM for shooting
10:05down incoming missiles, incoming mortar,
10:07incoming artillery fire, incoming
10:09missiles or incoming speedboats. Uh same
10:12thing, you can put them in fully
10:13autonomous modes. And I mean it going
10:15back to in into Vietnam, we had missiles
10:17that you could launch from an aircraft.
10:19They flew over the horizon, tried to
10:21identify the right radio emitters that
10:23they believed corresponded with a
10:24surfaceto-air missile site, and then
10:26struck them without any, you know, the
10:28person had to decide, I'm going to send
10:29it in this area. they're going to look
10:31for this signal and I know that there's
10:32this probability they're going to get it
10:34right. That decision is going to remain
10:36with people. But uh autonomous weapons
10:38are uh widely widely there's a whole
10:41another talk we're not going to have
10:43today. But I I will I I do have a rant
10:45about this where actually autonomous
10:46weapons have existed for thousands of
10:48years. If you look at and they're
10:50defined as kind of weapons that can
10:53execute the intent of uh what you
10:55program to do independent from the
10:57operator or the engineer. You look at,
10:59for example, pitfall traps that allow
11:01people to walk over them, but then
11:03capture the commander on his horse when
11:04he walks by and it's heavier. You look
11:06at uh certain poison traps that would uh
11:09that would uh kill people who had who
11:12had certain characteristics, not kill
11:14people who had other characteristics or
11:15not kill animals. Uh you look at many
11:18types of landmines and sea mines and
11:21even even torpedoes. And so it depends
11:24on on your exact definition, but even
11:26going into myth where people are
11:27creating golems that can, you know, go
11:29out and fight on their fight on their
11:30behalf or or defend you defend their
11:32city. I I guess my argument is we've had
11:35autonomous weapons for a very very long
11:37time. Uh it's just a new set of
11:39technology that's that's making them
11:41better.
11:41>> So technical question on that. The
11:43foundation models that go into those
11:44those remote systems right now.
11:47>> Yeah. Well, so the the right right now
11:50the open source models are all Chinese.
11:52So I'm pretty sure you're not going to
11:53be putting Chinese models into remote
11:55weapons. Are you going to build and
11:57foundation models or how are you going
11:59to get AI in those devices? I mean,
12:01we've been doing it since 2017. So we're
12:04way ahead of everybody. We're we're
12:05we're a decade beyond that question. Um
12:09I think that we also don't the
12:11foundation models that people are
12:12building are not the types of models you
12:14want to be doing this. They're non
12:16they're largely non-deterministic.
12:17they're non auditable. It's it's a very
12:20different category. The thing that we
12:21are building unfortunately has very few
12:24consumer applications
12:25>> and vice versa. There of course are LLM
12:29use cases for defense that make a lot of
12:31sense, but it's not guiding my missile
12:33or or classifying radar, you know, radar
12:36radar radar feeds. I mean, and we we're
12:38also we can't build models that ingest
12:40text and then spit out text. I need to
12:42process raw data radar data feeds. I
12:45need to process sonar data from passive
12:47listening stations on submarines. That's
12:49it's it's pretty highly specialized. And
12:51I I will take some credit here. Uh all
12:54these people who are talking about how
12:55AI is going to change everything, they
12:58are mostly uh recent converts. And
13:02Industries, you might notice the acronym
13:03is AI. Um so and that that was not that
13:06was not a mistake.
13:07>> Okay,
13:08>> Alex, jump in. Yeah, I think maybe
13:12dovetailing on our discussion of driving
13:13bench, we're now in an era when it's
13:15possible to take frontier models, drop
13:17them into a robotic embodiment and get
13:19pretty stunning results. So, I'd love to
13:21ask you a question maybe about the
13:23scaling hypothesis as it applies to
13:25autonomous weapon systems or autonomous
13:27systems in general. If you have to
13:29extrapolate out, is there a Moors law?
13:32Is there an experience curve for what a
13:34as obviously the capabilities of
13:37underlying frontier models whether
13:38they're from Frontier Labs proper or
13:41from AI as as those progress do we see
13:45any equivalent of a Moors law for the
13:48autonomous systems themselves and and
13:51maybe the the second half of the
13:52question would be do you think
13:54ultimately we achieve some sort of
13:56convergence you'll it makes a variety of
14:00embodiment in a variety of different
14:01form factors
14:02Do you see ultimately some general
14:05purpose category killer robot that is
14:08equivalently the generalist robotic
14:11embodiment just like we now have
14:12generalist models that are capable of
14:14doing substantially any task?
14:16>> No, I don't think so on the defense
14:18side. I think that for manufacturing and
14:21for hospitality and medicine, you might
14:23see this because you get economies of
14:26scale that uh that that are really that
14:28are really material for those types of
14:29businesses. And you know, a humanoid
14:31robot may not be the optimal form for a
14:34burger griller, and it may not be the
14:36optimal form for a firefighter. It may
14:37not be the optimal form for a for a
14:39nurse [laughter] or an assembly
14:40technician.
14:41>> Selma is smiling.
14:42>> But even if it's not optimal, the fact
14:45that it is shared with humanity in its
14:47form makes it a really, really good form
14:50factor. And now I can make one robot
14:52that does all of these things. And uh I
14:54I would probably rather have a let's say
14:57robot that's 5% slower or or uh you know
15:00less reliable than the optimal burger
15:02flipping robot if I can share
15:04commonality with that with that nurse.
15:08In defense, that's not really the case.
15:10Uh- because you are in this constant
15:13race to not just barely beat your
15:15adversary, but to massively overpower
15:17them, to try and make it where you the
15:19goal is not to just barely win a
15:21conflict on the margins. It is to fight
15:24unfair, to have overwhelming threat of
15:28violence underpinning whatever
15:29diplomatic overtures or negotiations you
15:32have. And and and everyone's always
15:34pushing the limit. So if if I if I have
15:36a jet that has a reliable 5% every time
15:40edge over the enemy, that's worth paying
15:42a lot of money for, right? Like I I I
15:44and and I think the the the
15:48loss in performance you would see
15:49through lack of specialization probably
15:52makes it nonviable. like to to to go to
15:54the to the extreme like uh you know I
15:56build robot submarines and those have a
15:58specialized form that needs to dive to
16:006,000 meter depth and and sit on the
16:03bottom of the ocean for months at a time
16:05and then I have other stuff that's on
16:07the surface and uh and you know they're
16:09robotic systems that you know roll up a
16:11hill and and they loiter and then I have
16:12other things that are you know aerial
16:14vehicles that are doing very specialized
16:15tasks and I just can't imagine a
16:17convergence that would make sense unless
16:19you're telling me that we're going to
16:20get something cool like like humanoids
16:21that have like you know onboard fusion
16:23reactors and they could dive to 6,000 ft
16:26depth.
16:26>> So Palmer, why why aren't you building
16:28those? I mean, if anybody you should
16:31[laughter]
16:31it's look, you could potentially build
16:33such a thing. But then very quickly you
16:35realize that specialization would allow
16:37you to build something even better. You
16:39say, okay, I've built something that can
16:40dive to the bottom of the ocean and fly
16:42to the moon. Man, it could get to the
16:44moon a lot faster if it didn't have to
16:46also dive to the bottom of the ocean. So
16:47I think I think defense is going to be
16:49this one where kind of like uh you know
16:51you you look at for example the
16:53difference between cars where everything
16:54has shared platforms within a
16:56manufacturer you know manufacturer and
16:58they also share a lot of the same
17:00standards uh and interfaces. Uh I think
17:04that you compare that with like uh
17:06motorsport where it's all completely
17:08different and it's continuously pushing
17:09forward. I I think the defense side is
17:12probably always going to be this this
17:14weird branch of exotic, highly
17:16specialized robotic systems.
17:18>> You know, Palmer, I expected if anybody
17:20was going to be building humanoid
17:21robots, it would be you.
17:23>> You said you said absolutely not.
17:24>> I' I've opted out because um I because I
17:27believe in them. If I didn't believe
17:29that humanoid robotics had such a
17:31powerful future, then I would probably
17:32build them. But I want Andreal to build
17:34things that wouldn't exist otherwise.
17:36That's what I generally want. You want
17:38to maximize your marginal impact, but
17:40they
17:40>> correct? All the other people are
17:42already doing a great job. Like I I want
17:44to use their robots and and and do
17:47things with them. Like I think humanoids
17:48are going to be an incredible general
17:50artificial intelligence interface
17:52between AI and remote systems and a lot
17:56of legacy platforms that were designed
17:57for a person, you know, pull the lever,
18:00push the button, read the screen. Like
18:02we're going to be able to take all those
18:03and automate them without replacing them
18:05entirely. So, you have a lot of things
18:07like radar systems that cost millions,
18:09tens of millions, even hundreds of
18:11millions of dollars. Uh, nobody wants to
18:13man them during a hot conflict because
18:15they're the first thing that's going to
18:16go. But if I can put humanoid robots in
18:18them, I've now been able to automate
18:20them without that. So, yeah, I'm not
18:22going to build humanoids. Uh, this is
18:23also why Andre is not building like
18:25small quadcopter attack drones. Too many
18:28other companies doing a great job.
18:29There's there's a lot of a lot of areas
18:31and and we're not going to do space
18:32launch like we do space products. We
18:34have a lot we operate a lot of our own
18:35satellites and a lot of our own sensors
18:37in space. Uh but I don't want to be a
18:39launch provider. That's for all the
18:40other billionaires to
18:41>> Yeah. Yeah. Let let them handle that.
18:42That's fantastic. Sm um you know when
18:46you're operating Andreal Industries,
18:48you're uh decentralizing defense, right?
18:52The Pentagon and Ministries of Defense
18:55globally have been the most centralized
18:57entity in the history of humanity. Uh
19:00you're sitting on the edge dealing with
19:02that. uh you must have a massive immune
19:05system problem where people inside the
19:07entity are going, "No, we've done that
19:08before. We can do it better. They either
19:11try and suck you in or they try and spit
19:13you out." How are you navigating that
19:15interface? Because you're doing it in a
19:16way that I've never seen anybody succeed
19:19do it.
19:22>> I think that this is a problem that's
19:24very very hard at the beginning.
19:26>> Oh, I said the wrong thing. [laughter]
19:28>> Wow. No idea. Well, until someone gets
19:31on the intercom and tells us to go a
19:33million dollar.
19:34>> Congratulations.
19:36>> Um, what were we talking?
19:44>> Seriously, I knew the intercom was going
19:46to come on.
19:47>> What were we talking about again?
19:49>> Centralization of Pentagon versus
19:51decentralized
19:52>> I'm going to keep talking to.
19:55Um,
19:55>> so we've gotten through this a bit. I
19:57think early on there was that concern
20:00that Anderl was not going to be able to
20:02get through that. Our our hardest
20:03problem raising money.
20:04>> Skippy, would you please shut this down?
20:06[laughter]
20:08>> I've been through enough of these to to
20:09know know how it's going to work. There
20:11we go.
20:11>> I imagine.
20:12>> Um, one of our early problems was that
20:14our investors said, Palmer, you
20:16potential investors. They said, "Look,
20:18we believe that you'll be able to build
20:19things that work. We believe you'll be
20:20able to build products that are great,
20:22but we don't believe the government is
20:24capable of actually purchasing them at
20:26scale." Because I mean you talk you talk
20:27about centralization you know about the
20:29issues for people who don't uh until
20:32just a few years ago you had about 80%
20:36of uh MDAPs major defense acquisition
20:39programs so basically anything that that
20:41matters uh 80% of MDAPs went to five
20:46companies and that's it
20:48>> and uh I mean you've seen incredible
20:50consolidation centralization and has
20:53been successful because we use we're
20:54less of a contractor and more of a
20:56product company. We use our own money to
20:57build products. Then we sell them. And
20:59we've gotten to the point now where
21:01people in government have done very very
21:03well betting on and then miraculously we
21:05finish on schedule. We finish on time
21:08and then the next program manager says,
21:10"I know how I can make my job easy. I
21:13just I just give the contract to Ander."
21:15>> Got it.
21:16>> So it's got to take guts to enter the
21:18defense industry, right? The most
21:20incestuous, you know, self-licking ice
21:22cream cone on the planet.
21:24>> Yeah. and and and what insanity,
21:28you know, gave you the balls to do that?
21:31>> The insanity was the reason that I got
21:34into it. Uh the first page of our pitch
21:36deck said Andro will save taxpayers
21:39hundreds of billions of dollars a year
21:41by making tens of billions of dollars a
21:43year. The idea being we're not just
21:44making products that we're selling, but
21:46we are displacing the systems that cost
21:49needlessly 10 times what they should.
21:52And yeah, thank you. [applause]
21:56And so it does seem crazy, but you know,
21:58um I don't I I I can't put myself on the
22:01same pedestal as him, but one of my
22:03mentors and idols growing up and then
22:06also early in my career was John
22:08Carmarmac, the founder of Software.
22:09>> John is great.
22:10>> John is incredible. And for people who
22:12don't know, he's basically inventor of
22:14the modern modern 3D game engine
22:16firsterson shooters. Also started a
22:18rocket company that competed in He he
22:20was competing in the X-P prize and I I
22:22remember
22:22>> and he was one of the if not the only
22:25one to be doing reusable vertical.
22:27>> Exactly. He was doing vertical takeoff
22:28vertical.
22:29>> Nobody else was even trying to do
22:30>> and I remember I remember there's this
22:32this name on the side of his vehicles
22:34Nvidia. How do you pronounce that thing
22:36if I would only know?
22:37>> That's right. He was use and so yeah he
22:39was using uh he was using uh yeah he was
22:42building his own sensor fusion
22:43algorithms. He was processing them on
22:45modern or on a you know like commercial
22:47style graphics card hard hardware using
22:49his game programming experience. Anyway,
22:51the point is
22:52>> it was called Armadillo Armadillo
22:54Aerospace. That's right.
22:56>> And the thing and yeah, they were doing
22:58vertical takeoff and landing long before
22:59SpaceX was uh was was was I mean doing
23:03anything.
23:03>> Yeah. No, before it existed for sure.
23:05>> That's right. And uh so but anyway, the
23:08point is he he um I I he ended up
23:11becoming the CTO of Oculus VR. I managed
23:14to poach him away from ID, but after he
23:17left, he went to work on AI actually
23:20even before I did. Um and so he started
23:23he started his own company to work on uh
23:25on on AGI. And um and I said, "Why why
23:29are you why are you doing this?" you
23:30know that the odds of you succeeding and
23:31to be clear I still believe AI was going
23:33somewhere but he said I'm going to start
23:35a company that's just going all in on
23:37the end state like I'm not trying to
23:39incrementally get there it's just all in
23:40on AGI and the point that he made was
23:43Palmer even if I've got a 1% chance of
23:46succeeding which maybe is an
23:47overestimation but suppose I have a 1%
23:49chance of succeeding and suppose that I
23:52could pull it off uh
23:55the uh the risk adjusted uh outcome kind
23:58of demands that I do it. And that was
24:01how I felt about defense. If there was a
24:031% chance that I could help solve this
24:06problem, then it was one that would be
24:08massively impactful to work on versus,
24:11let's say, starting another gaming
24:12company, getting into space launch, uh,
24:15you know, building a car company, uh,
24:17that that which it it all of the
24:19billionaires, they end up doing the same
24:20thing. You don't have enough people
24:22getting into defense or or putting on a
24:24bat suit. It's just it's just cars and
24:27rockets and and they just go in circles
24:29like that for for for
24:31>> So, you know, when I interviewed you on
24:32the on the Moonshots podcast last time,
24:34you mentioned that you're part of a
24:36billionaires boys club chat group.
24:38>> Yeah. Yeah. [clears throat] Oh, yeah.
24:39>> Are they working on anything worthwhile
24:41at this point? I mean, I g I gave them
24:43for saying like, "God, you got all
24:44this capital. Solve the world's biggest
24:46problems."
24:47>> I think the B boys are doing a lot
24:48better now than we were a few years ago.
24:51>> Okay.
24:51>> Um
24:52>> they're they're stepping up.
24:53>> Yeah. And for people the thing we're
24:54talking about is I I was giving a lot of
24:56people a hard time especially some
24:58retired B boys and uh I was telling them
25:01what are you what are you doing you have
25:02cuz money is one of the resources they
25:05have but one of the resources people
25:06don't talk about especially for the
25:08retired B boys is that they have
25:10credibility which would allow them to
25:12gather people around them teams around
25:14them and capital like they could it's
25:17just sitting there unutilized if they're
25:19doing nothing. um they have the
25:20credibility in that people trust them as
25:22a leader to be able to lead an
25:24organization like that. They they could
25:26have the same ideas and the same words
25:29as somebody who's equally smart or
25:31smarter and uh they'll still do better
25:33for that. And so my point to them is you
25:35have a responsibility not only as
25:36someone who's who's got a lot of capital
25:38yourself, but who's got this
25:40underutilized resource of the ability to
25:42do more uh uh if you're not doing that
25:45then then are you really like what are
25:48you even doing? Anyone can sit on a
25:50boat. Few people can start another
25:52multi-billion dollar company that does
25:54something that matters. [applause] Yeah.
25:56And the good news is most of them I
25:58think uh I I I think a lot of people
26:00have uh have gotten the itch to to solve
26:02some of these problems that are out
26:03there.
26:04>> So we've got
26:04>> nihilism is is is on is on the down.
26:06>> The solution is we need more tea boys
26:08and fewer B boys. [laughter]
26:10>> Yeah, that's an interesting point.
26:12Although
26:14like I said, I think most of the tea
26:15boys I think they seem like they keep
26:16pretty busy. That's your point.
26:18>> I don't I don't even I don't have a
26:20problem with B boys who don't do enough.
26:21Like some people they mis mistake what I
26:23what I say like oh Palmer wants other
26:25billionaires to work on defense or
26:27energy or or biology and you like my
26:31point is not that people need to
26:33distract themselves by doing more
26:34things. A lot of people can only do one
26:36thing uh you know and that was what I
26:38did for a long time. Uh it's what I hate
26:41is the people who are not doing anything
26:43that's that that is that is like a
26:45repulsive to me. So maybe message to the
26:47audience, you get to join the B boys
26:49Slack group or whatever this is,
26:52Telegram if you're doing something new
26:54and differentiating.
26:54>> The group the group actually got shut
26:56down and uh we've been we've been
26:58talking about now we need to do a B boys
27:002.0. Um but I think I think people post
27:03>> maybe some B girls as well.
27:04>> I think people posted the wrong memes a
27:06few too many times.
27:07>> Boom. I I look there there was there was
27:10a time where there were no there were no
27:12be girls and then there was this
27:14Business Insider article uh whining
27:16about it and uh the reality is like it's
27:18just like me and some guys like you know
27:20it's not like there's this isn't like
27:21the Illuminati conference [laughter] you
27:23know it's like they're like oh there's a
27:26group and you know but no women have
27:27been invited it's not that no women have
27:29been invited it's just like me and some
27:31some of my boys but
27:33>> [laughter]
27:33>> um but actually there there was a there
27:35was there was a there was a girl at it
27:37by the So, uh, we need Yeah, we need to
27:39we need to restart that. We need to
27:40restart that group.
27:41>> So, you're in Facebook commercials now.
27:44>> Oh, yeah. Yeah. Yeah. I'm so VR is out.
27:48Uh, how does it compare against your
27:50vision of what you wanted to achieve? Is
27:52it?
27:53>> Yeah.
27:54>> Look, it does a lot of the things that
27:56uh it's not exactly what I would make
27:57were I the guy who's waving the magic
27:59wand and and and
28:01>> you got lots of magic wands
28:02>> spinning the CAD. But uh um look, the
28:05new MetaVR glasses are a lot closer in a
28:08lot of ways. Uh so I've long been a
28:10believer that you need to decouple the
28:12compute and the battery from the head.
28:15You should it should not be that and and
28:17and I'm not even saying the head
28:18totally. It's more the front. So before
28:21Oculus Quest, so Oculus Quest was the
28:23last thing that I worked on before they
28:25fired me. And uh the prototypes at the
28:27time were a prototype series code named
28:30Santa Cruz. And uh we were actually
28:33showing that publicly before they fired
28:35me. And uh Santa Cruz had the battery
28:37and compute on the back of the head and
28:40then the light, you know, the part on
28:41the soft tissue on the front of your
28:42head was kept as light as possible. And
28:44then at some point after they got rid of
28:46me, they they jammed it all into one big
28:48giant lump in the front and they ruined
28:49the ergonomics. They forgot why you need
28:51a top strap. Uh but you know, it it
28:55bothers them, right? It bothers me most.
28:57Look, it bothers me mostly at this point
29:00because uh the Oculus Rift CV1, which
29:02was the first modern consumer VR
29:05headset. It got the ergonomics more or
29:07less perfect
29:09>> in 2016. And you know, we were showing
29:11it off in 2015 and literally nobody has
29:15done as good of as it and it's been over
29:17a decade. Every single person screws it
29:20up, even Apple. And I I just I don't I
29:23can't comprehend it. How how could it be
29:25that the ergonomic design led by a
29:2819-year-old kid with no experience is
29:32better than anything that anyone has
29:34come up with since. Anyway, the new
29:36Metavir glasses are the closest thing
29:39and they're very lightweight. They've
29:40offboarded the compute. They've
29:42offboarded the batteries. They have
29:43really high quality pass through. Um
29:45they it has a feature that I invented
29:47for CV1 called transformo fabric which
29:50they do not call it. The basically using
29:53uh fabric a a soft flexible fab black
29:57black black lightabsorbing matte fabric
30:00to one minimize reflections but two uh
30:02basically seal the cavity between and
30:04around the lenses. That means that you
30:06can move them uh and the fabric just
30:08stretches to meet where the lenses are
30:10for different interpupilary distances
30:12rather than having like sliding
30:14mechanical interfaces and plastic. And
30:16they brought Transformer fabric back for
30:18the first time since my my era. So I'm
30:20I'm I'm a huge fan and uh you know it's
30:23a third of the price of Apple's product
30:25and uh somewhat higher higher pixel
30:28density. I I think that's pretty
30:30extraordinary. Oh and and and 1/8 the
30:32weight.
30:33>> Amazing.
30:34>> Amazing. Um,
30:36>> let's close on let's close on this.
30:38We've got, you know, 1500 entrepreneurs,
30:40moonshot visionaries, and tens of
30:43thousands watching on the live stream.
30:46Uh, what's your advice?
30:49What, you know, what did you learn in
30:50high school that made you succeed at the
30:54speed you did? We were talking about
30:56this on the plane.
30:57>> I mean, I I was really lucky to uh I
31:00guess maybe not lucky, I did it. Uh, but
31:02>> Well, you are you are lucky. That's
31:04right. I am [laughter] lucky.
31:06>> Similar name.
31:07>> I started an online community called Mod
31:09Retro when I was 14 or 15. Um, it was an
31:12internet forum for people who were
31:15modifying vintage game consoles and
31:17computers, building handheld units,
31:19improving the performance in a variety
31:21of ways. We had a few thousand active
31:23members and uh and a whole team of
31:25volunteer moderators and administrators
31:28with with me me at me at the top of it
31:30for some reason. And uh I learned a lot
31:33of valuable lessons about how you can
31:35build an organization, how you market an
31:37organization, how you get people to
31:39believe in what you're doing. I learned
31:41a lot about dealing with the types of
31:44problems you deal with when you're
31:45running a company, you know,
31:46>> with humans.
31:47>> Well, I mean, you you have two, it's a
31:49tale as old as time. You've got two
31:50people. They're both brilliant. They're
31:53both very good at what they do, and they
31:55hate each other, and they disagree over
31:57something. you need to somehow mediate
32:00the conflict or separate them in a way
32:02that allows you to get the value that
32:04that allows both of them to contribute
32:06value to the greater project uh but not
32:09have them hate you, not have them lose
32:11faith in the broader thing, but also you
32:14not not kill each other. Sometimes that
32:16means you need to reconcile things
32:17between the two. Sometimes it means that
32:19you need to put one guy over here and
32:21one guy over here and sometimes you need
32:22to make sure that they don't know that
32:24that's what you're doing as you do it.
32:25That's [laughter] just I'm not saying
32:26that's the key to running a successful
32:28business is the management of the two
32:30problem. But it's you and once
32:34you make it a threebody problem now
32:36[laughter]
32:37but uh but but it's it's that and a
32:39hundred things like it that I learned
32:42running Mod Retro that I was able to
32:44take into starting Oculus to carry
32:46forward into what I'm doing in my other
32:48companies. And uh that that's why I'm
32:50I'm a strong believer that that kids
32:53should should should get a job. And I I
32:55also had a job starting when I was 14.
32:56So I learned how the real world worked,
32:58too. But I I I'm a huge fan of people
33:01doing things at a scale that's larger
33:03than just themselves. In other words,
33:05don't just do a project. Do a project
33:07with someone else. Do a project with
33:09other people. Build in the open. Learn
33:12the the the the good parts of that, the
33:14bad parts of that. It it it's been
33:16extremely useful to me. And has building
33:18changed in the age of AI? What advice do
33:20you have? Again, you know, we've been
33:22saying on the pod for a while now, you
33:24know, forget about grad school. You
33:26know, even college is a is a maybe you
33:28didn't go to college. You know, in high
33:30schoolers should what? Find a a purpose
33:33they care about and use AI to go and
33:36code it. What's your your closing advice
33:38here on that?
33:40You know, one of the reasons I have not
33:43been as upset as some people about uh
33:48No, that's funny. There's something on
33:49the screen. It's telling us to not talk
33:51about something. [laughter]
33:53Just wanted to make sure you saw that.
33:54>> I I I did. We're not going to talk about
33:56my company.
33:57>> All right. Great. Great. Um [laughter]
34:00the uh there's there's a lot of people
34:03who who love their craft and they love
34:05what they do and uh and they are
34:08rightfully mourning uh the end of for
34:11example coding things by hand top to
34:14bottom, start to finish as anything but
34:15like a craft endeavor. Um, and and I and
34:18I understand that it's it's the same way
34:20that a lot of artists felt when the
34:21camera was invented. They they whined
34:23and complained about how it ruined art
34:25and it made it so that anybody could
34:26push a button and and you know, these
34:28things are always painful for the people
34:30who are masters of their craft. Um, I
34:32already got screwed in that way a long
34:34time ago. When you're a founder, you
34:35don't get to actually work in any of the
34:37tech anyway. All the other people get to
34:38have all the fun. And so, you know, I I
34:40like electrical engineering. I like
34:42optical engineering. I like mechanical
34:44engineering. I don't get to do any of it
34:45because the job of a founder is to uh
34:48hire people to to displace them and and
34:51and to do what they thought they could
34:53do or what they like they do like doing
34:56better. Uh the job I've said many times
34:58to people if you want to work on
35:00technology uh work for somebody else. If
35:02you want to work on non-stop
35:05start a company. And and and so and so
35:07what I would say to people is AI doesn't
35:10actually change the part of the
35:12job very much. you're still on the hook
35:14legally. You're still on the hook for
35:16all the interpersonal stuff. And you can
35:18you you can delegate some of it, but at
35:19the end of the day, it's your call. It's
35:21your decision. It's the same thing for
35:23fundraising. It's the same thing for
35:24general product architecture. And when
35:26people make the wrong decisions, it's
35:28your choice as to when you need to
35:30overrule them, override them, replace
35:32them. I don't see AI replacing that part
35:35of it anytime soon. Not because it
35:37couldn't, but because the gains
35:39everywhere else are so much larger that
35:42I think we've got a good 10 years of
35:44speed running all of the other parts of
35:46the business before they before they
35:48cover that. So, look, AI changes a lot.
35:50But I I'm telling you, there's a lot of
35:52people you are in for a rude awakening.
35:54If you are like a hobbyist coder or a
35:56hobbyist engineer and you think you're
35:58going to start a company and you think
35:59you're going to be doing all that stuff,
36:01no, you're going to learn a hard lesson.
36:03B being a founder was was always like AI
36:07took your job. It's just that other
36:08people took it.
36:09>> Yeah.
36:10>> Ladies and gentlemen, give it up for
36:12Lucky.
36:19[music]

About this transcript

This page contains the full transcript of Palmer Luckey: Autonomous Weapons Are Ancient and Why Anduril Won't Build Humanoids | EP #295 by Peter H. Diamandis, generated from the public captions YouTube serves with the video. The transcript has 7,241 words across 1,050 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.