When Will Robots Live With Us? Inside 1X's NEO with Dar Sleeper, VP of Design & Product

(Intro Narrator at 00:00:00) Today, we're talking to Dar Sleeper, design product and marketing vice president at 1X, about their NEO robot and when robots will live with us. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:19) Do you have a cool name for that room? Is it the control room or something?

(Dar Sleeper at 00:00:22) This one I'm in right now? Yeah, we call it the Sanctuary. Yeah, we actually—it's funny.

(Dar Sleeper at 00:00:28) And I guess now we're good to record, but I guess we're already recording. The Sanctuary was an idea at 1 a.m. at an In-N-Out Burger. We were in the middle of shooting our video to give the world an update on our world model, and it was the fourth day of shooting. The world model shoot was a ton of fun because basically it was the first time that you could just ask NEO to do anything and it would try it. It's kind of a mini ChatGPT moment in the sense that there wasn't a world before where you could just prompt a robot and then get an output.

(Dar Sleeper at 00:01:06) This zero-shot intelligence just didn't exist. Everything was VLA trained before, for the most part. So most of anybody in robotics spent most of their life like, okay, if we want the robot to do something, you go give it a hundred examples or so, and then you try to train a model and then you iterate on the model and blah, blah, blah. But this shoot was super fun because we basically just went to Target and bought everything we could think of that we would want a robot to try to play with, from shirt irons to toilets and toilet seats to rolling dough.

(Dar Sleeper at 00:01:37) And we ended up shooting just for a really long time because we were just having so much fun with it. But on one of the nights, we were at In-N-Out Burger and we were thinking about how important it is to just share a bit of an update with the world about our factory. You know, I think it's funny how people assume that a lot of this is very strategically planned, like, "Oh, we're gonna throw this punch, then that punch, then this punch." And to a degree, some of it is.

(Dar Sleeper at 00:02:00) You don't have to do a lot of planning when you're kind of tearing through the world trying to bring a robot to market. But for the factory—you see the factory—we were heads down just cranking on getting this factory line in Hayward up. And, you know, me and Finn, who's the guy that shoots all the videos here, we were sitting at In-N-Out eating French fries and burgers, and we'd spent the day before in the factory. And the vibe was amazing.

(Dar Sleeper at 00:02:26) Seeing robots being built and coming off the line, going from raw metal to a robot, is in and of itself so crazy. But we're really childish, the two of us. You know, I'm on the design side of the world, and he's all film. And so we're kind of frustrated that it didn't feel like the future that we imagined when we were kids, where robots were being built—not because the robots weren't being built, but because we totally imagined the end of it being this railway where robots go down the line and then they start walking off and they go into their birthing chamber and they get packaged up and sent out in the world.

(Dar Sleeper at 00:03:03) So I talked to the VP of manufacturing here, Vikram, who's killer in the factory world, who just thinks very utility-based. And I was like, "Hey, what happens at the end of the line?" He starts going through the crash course on how the end of the line works, and I was like, "Okay, so what if you just took—" because they needed a craftsmanship check, a place to really put NEO under a light that was reflective of what it would be in the consumer's home and go through tolerance checks, check all the fabrics, do proper craftsmanship checks. And he's like, "You know, right now we're able to do that in a pretty small space. But as we start to scale up, hit that S-curve of production, we're gonna need a real high-volume craftsmanship check space." I was like, "That is the perfect place for us to try to put some 1X touch in this factory and create this childish review of what a robot coming into the world looks like." So this is NEO.

(Dar Sleeper at 00:03:55) This is NEO's last stop before he goes out in the world.

(Joel Beasley at 00:03:59) The Sanctuary and the Birthing Center? Is that what we're calling it?

(Dar Sleeper at 00:04:03) We call it the Sanctuary, but people just end up calling it the Birthing Center too. It's kind of natural. It's kind of where he's born.

(Joel Beasley at 00:04:13) This is awesome. You guys are opening up the dystopian fiction book and you're like, "Let's bring this into reality." I love it.

(Dar Sleeper at 00:04:21) Yeah, yeah.

(Joel Beasley at 00:04:23) So right before we had this, I was talking with my wife, and I told her I was like, "Oh, I'm talking to this guy, Dar. They have robots in the house." And she goes, "Can it do laundry, like all the laundry?" And I said, "I don't know. I will ask him." How much work would it be from where I am in reality right now—my coordinates in Nashville, Tennessee—for me to purchase one of these bots, have it come to my house, and do my laundry?

(Dar Sleeper at 00:04:52) Yeah, so the first shipments to consumers haven't gone out yet. And I will say Nashville probably won't be the first market to receive their robots. So I'll start with that, just a planned admission of unfortunately, today, I cannot get you your robot in Nashville, but soon. So that's, I guess, the preface. We built the product experience to be pretty scalable. And I'll kind of—I mean, I have to go on a rant to really ground you in why this answer is a bit nontrivial, or I guess not as simple as yes, it does your laundry. It's because we basically had to build a product where it scales with the frontier of AI. Because, you know, if robots today were able to do a full load of laundry—so by that, I mean laundry is a super complex task.

(Dar Sleeper at 00:05:35) And it's one that everybody asks about, and it's also one of the ones that's kind of this AGI-complete moment because it has to go walk up the stairs—which we've solved—go to your room—solved. So that's saying now you've got RL locomotion to do reinforcement learning to learn how to walk up any type of staircase, walk down any type of staircase. That's solved. Getting to a room, that's solved. This is just basic mapping. There are different ways to do it. Some more exciting new developments that we've recently folded into the product that make mapping and navigation way cooler. And then you gotta get to the room and you gotta go grab the laundry basket—solved.

(Dar Sleeper at 00:06:12) You know, this is a basic pick-and-place. You're reaching out, you're grabbing things with your hands, you walk it back down to the laundry room—solved. And then when you get into these really intricate moments where there's top-loader laundry machines—the variance of laundry machines is pretty extensive—using a world model, you could try to open the laundry machine. Given the world model has a lot of different laundry machines in its dataset at the video pre-training level, it will likely be able to do it. Whether it succeeds or not is still based on how much robot data exists to fine-tune that world model to be able to close that gap from sim to real, kind of to go from all the data in the video set to performing a task in real.

(Dar Sleeper at 00:06:55) But doing it reliably may not be quite ready yet. And then you go down the list of tasks cascading from there, which is—a full laundry load is, again, extremely complex. And you have to know what the preferred settings are. That's a relatively solvable problem. You tell your robot what you want. You tell your NEO what you want. You then have to go from washer to dryer, then from washer to dryer, dryer sheets—tiny pinches. So the automation of hyper-dextrous tasks like tiny pinches against dryer sheets—I keep going, explaining how this is difficult. But the way we structured the product is that it scales with autonomy. And by that, I mean we made it an experience where NEO uses its best autonomy to achieve as much as it can. And then if it finds itself outside of its autonomous abilities—which it's able to assess by running its own evaluation of "Was it successful or not?"—

(Dar Sleeper at 00:07:45) And once it gets to that point, if you are during a window where you would like to have—this is very similar to Waymo, how there's somebody that, if anything ever is not able to happen within the wheelhouse of what is autonomously capable by the Waymo, there can be an operator that hops in and corrects the situation and completes that one task. So if the customer is in a window, or say you're at work and you would like your NEO to have that supervision that a Waymo has—it's kind of called shared autonomy—then at that point—

(Joel Beasley at 00:08:17) Your operator could take over and get that done for you.

(Dar Sleeper at 00:08:19) So then the short answer becomes yes, NEO can do your laundry. But the good news is, just with the world model being something that we thought was maybe longer out than we initially thought it was when we launched the product, it's taking so much progress. There's a high chance that these things kind of just get solved before you in Nashville get your robot. So the answer could just be plain yes, and you don't even have to worry about your operator window.

(Joel Beasley at 00:08:41) Does anyone have any of these in a residential home today?

(Dar Sleeper at 00:08:47) Yeah, so we've had limited testing within our company, and kind of the current goal right now is to expand it further throughout our company. So we want to make sure that everybody at 1X loves it as much as me and Bernt and all the leadership at 1X, who—I mean, we kind of drink our own Kool-Aid, so we have to make sure that, yeah, basically there's levels to this. You go from leadership who will die for this product to employees who love this product, and then you go to their brothers and sisters.

(Joel Beasley at 00:09:14) A misdemeanor for the product, but they're not gonna die for it.

(Dar Sleeper at 00:09:17) Yeah, yeah, yeah.

(Joel Beasley at 00:09:18) Yeah, we'll grant that product. You're deploying throughout the company as the test users, essentially, the beta testing throughout your company.

(Dar Sleeper at 00:09:31) This is basic dogfooding. You know, every consumer product should at least go through this. If you're not willing to sell it to your own brother and take your own brother's money for it, there's absolutely no reason why—and not to mention a stranger's a good litmus test too, but I like to think of it, would I tell them about it?

(Joel Beasley at 00:09:50) Answer to family at family events.

(Dar Sleeper at 00:09:53) You know? Yeah, yeah, yeah. Yeah. Look at this asshole, or coolest guy ever, one of the two. I mean, my younger siblings and my younger cousins think I'm really sick, so that's great.

(Joel Beasley at 00:10:07) Okay, so only employee, founder group, executives right now have them operating in their home?

(Dar Sleeper at 00:10:15) Yeah.

(Joel Beasley at 00:10:16) When are they gonna start shipping? Is there a date, or are you just waiting until it's the quality that you accept from you and your company employees?

(Dar Sleeper at 00:10:25) So it's plainly both. We have our internal date, which I will not share here on this podcast, for when these go to our first customers. With that said, there is such a tremendous responsibility to—yeah, we absolutely are committed to keeping right by our promise. Our promise is that we ship the first NEO in 2026. So we start shipping NEOs in 2026. 2026 is a long year, which is good. But to be—

(Joel Beasley at 00:10:50) December 31st, he ships one unit to his brother.

(Dar Sleeper at 00:10:54) "Oh, he did it."

(Joel Beasley at 00:10:54) By the way, I'm a founder. That's what we do.

(Dar Sleeper at 00:10:58) Yeah, yeah. I think the really important part here, though, is there's absolutely no reason to send out a product into the world—for both our customers and us—that isn't something we're deeply proud of, that we're deeply excited about, within all of those litmus tests that I provided. From every layer of the company being excited about it to friends and family of the company being excited about it to moving outwards to very early customers even. That's more of a question of, before going from early adopters to scale, what types of hoops do you jump through? What launch gates exist? But there's no reason why we should ship anything out that's not the best product on Earth.

(Dar Sleeper at 00:11:37) That's super important, not just for our company—because, you know, you ship a shit product, you're fucked. But if you ship an incredibly amazing product, great for the company. But also, more importantly, it's really good for humanoid robots. We take—we don't just take this as a responsibility for ourselves, for our employees, for our customers, but we do kind of assume the burden of the industry. Because the industry is very much one that has been a long time in the making. A lot of the innovations that we're riding off of—everybody says shoulders of giants, but I don't think that's ever been more true than in humanoid robotics.

(Dar Sleeper at 00:12:11) Because, we've done a lot of innovation on these tendon drives and the tendons and the motors that we had to create to make this work. So I give a lot of credit to the founder, Bernt Børnich—true visionary. A lot of credit to the team for bringing it to life. But for the most part, we are riding off the success of decades and decades of roboticists. I mean, da Vinci drew the first tendon-driven robot in the 1400s. So with that said, I think it's super important that we don't disappoint on this entrance into the world. So that's my really long-winded way of saying the most important thing is that we do justice to humanoid robotics because it is the coolest and most fun product on Earth. So that's what we care about when it comes to shipping.

(Joel Beasley at 00:12:56) Oh, I'm on board. Now, who is the least patient? When you interact with all different groups of people, from investors to customers, who is the least patient group?

(Dar Sleeper at 00:13:08) The people that are like, "We want it now." It's the people who—so, you know, I give—again, we made a really good point to be as authentic and real as we could when launching this product. We didn't say we're selling a thing. We didn't say it's gonna do everything you want autonomously from day one. This is a very dangerous move that I see plenty of Silicon Valley companies fall on the sword of—promising the world and not being able to deliver.

(Dar Sleeper at 00:13:33) I won't name any names, but, you know, we've seen plenty of these. And so anybody who really got down with what the beginning of this process and journey looks like, those people are super patient. I'll kind of go through the list. Investors—again, I wouldn't say they're patient or impatient. They just really care about the company making extreme progress. If you're a deep tech company in this type of space, you just gotta show that you have what it takes and the velocity of development speed to rip a new market open. And I would say that the least patient person, backing up into that, is by far and large the person who saw NEO on Instagram, got super stoked—they're like, "Humanoid robots!"—and your imagination runs.

(Dar Sleeper at 00:14:24) Yeah. Their imagination runs a fucking marathon, and you end up at a place where you just see this robot solving every problem you have from day one. Those folks are pre-order holders that I get the angriest tweets from. That's where I actually stopped.

(Dar Sleeper at 00:14:45) I stopped. We have one of my favorite things to do in the world is design everything, not just robots. So we design Neo. I completed the designs with my team for this one, for the industrial design package that's coming out to customers. That was completed almost two years ago now. And so in the meantime, I'm always itching. And so we have some of the best merch in the world created, you know, jeans, highest quality jeans that fit the best and I love wearing.

(Dar Sleeper at 00:15:13) We have sweaters that are human sweaters that look like Neo's sweater. You gotta—

(Joel Beasley at 00:15:18) Send me one.

(Dar Sleeper at 00:15:19) I will. I will. Yeah. Just send me your address. I will.

(Dar Sleeper at 00:15:21) I think Kendall is somewhere in the audience of this podcast. She'll get you your Neo sweater.

(Joel Beasley at 00:15:27) Let's do it.

(Dar Sleeper at 00:15:28) But I, and that's all to say, I stopped dropping merch for the company because, you know, people would be like, "I don't want a sweater. I want a fucking robot." And then I'd be like, "Well, I don't know how much you know about the world, Sonny boy, but it's a lot harder to ship a robot than a sweater." And if you want me to stop shipping merch, I'm the design guy.

(Dar Sleeper at 00:15:45) If the engineers were shipping sweaters, I'd be pissed. Right?

(Joel Beasley at 00:15:50) They would be very efficient.

(Dar Sleeper at 00:15:51) Yeah. Yeah. Yeah. They wouldn't look great, though.

(Dar Sleeper at 00:15:54) Very easy business. But yeah, it's gotten to the point where some people are so mean—this audience that I talked about that's so impatient—they're so mean that I'm just like, "You know what? No more merch for anybody else besides this company."

(Dar Sleeper at 00:16:06) Anything we make is just for the employees. There you go.

(Joel Beasley at 00:16:09) Yeah. That's the comment trolls. Who knows? Most of them are probably bots anyways.

(Dar Sleeper at 00:16:14) Yeah, but I'm sensitive, so, you know.

(Joel Beasley at 00:16:16) I get it. I have to stop looking at the comments, and it's like eating dessert for me. I'll say I'm not doing it, but once in a while, I'll do it.

(Dar Sleeper at 00:16:24) Yeah. Yeah. Yeah.

(Joel Beasley at 00:16:26) You know? Do you have one running around your house?

(Dar Sleeper at 00:16:29) I've had one for a lot of months. I am—it is—I'm actually getting a hardware upgrade now. So we iterate on hardware pretty fast here. Obviously, when you move to scale and you start doing things, when you start doing things like die casting and taking things to really do hyper bomb cost reduction on the robots, you have to freeze designs. But our manufacturing setup, to be able to—we have a great new product introduction line where new hardware models could be spun off in four weeks.

(Dar Sleeper at 00:16:59) So from CAD to robot being, walking off the NPI line and then having its path to be integrated into our main factory line, it's four weeks long. So we're constantly doing hardware iterations. And we went through, like—hey, I can't give specifics on versions or when or why—but yeah, we've had one of the most exciting hardware revisions lately, so I'm actually getting my hardware upgraded right now. But I've spent many months with Neo.

(Joel Beasley at 00:17:22) So then you can do a manual control of the robot. Right? That's what the operators are going to be doing?

(Dar Sleeper at 00:17:28) Yeah. That's not the main experience. The main experience is all AI. So it's like you talk to your robot. Do it.

(Dar Sleeper at 00:17:34) Yeah.

(Joel Beasley at 00:17:34) Yeah. So I've got a line of questioning.

(Dar Sleeper at 00:17:36) I want to get there. Yeah.

(Joel Beasley at 00:17:37) Yeah. I wonder how you got—

(Dar Sleeper at 00:17:39) My phone app is nearby. But you can port into a robot. And I'm saying this because I use the developer version, which I can access all the ones inside of our network in this factory in San Carlos. But yeah, you can steer your robot around from a joystick. You can text it and say, "Go open that door." It's pretty fun.

(Joel Beasley at 00:18:02) Okay. Can you teach it skills with that technology? I'm trying to get at: have you been able to jump into the robot, teach it specifically how to do your laundry—

(Dar Sleeper at 00:18:14) Yeah.

(Joel Beasley at 00:18:15) And then have it repeat that task or teach it 10 times, and then it can repeat the task as long as the things are reasonably within the same area?

(Dar Sleeper at 00:18:23) Yeah. So the cool part here is we're not so far from a world where robots can teach themselves. And that sounds like a crazy claim, but this was actually one of the original—we have a whole flow design for that that we built out where you go and you give it one example, and then you do the task. And you give it another example, you do the task. And then it just tells you when it has enough data to try it. It was a bit arbitrary because there is actually no specific, "Oh, if you do it 10 times, it's going to work." It's kind of like the more data, the better. And then also, the more diverse the data, the better. And the more clean the data, the better. So it's like a really shitty example that's non-representative could fuck up the distribution for a vision language action model, which is a VLA, which is the traditional approach to training a robot. Since departing from VLAs, that experience actually doesn't make that much sense because a world model, which is now at the core of Neo—it's what drives its autonomous system—is able to just say, like, "Hey, Dar," or, "Hey, I'm Neo. Grab that cup." It'll reach out, grab the cup.

(Dar Sleeper at 00:19:24) And that's something I never was able to do before. You just have to give it examples, what you're just talking about, to grab that cup. But with the world model, you can just say, "Grab the cup." And it'll use a video generation model at its core. So everybody who knows about the latest video gen stuff knows this is kind of well solved by now. But it takes that frame, and it says, "Grab the cup," and it sees the hand's in front of you, and then just generates a video: go grab the cup. And now we're able to use that by doing some extra steps in between, some cool architecture stuff that ground that all in physics and reality. Because a lot of the times, you say "grab the cup," and the video model goes here, and then it goes "boop," and it ports in there because the video model just wants to succeed. But with these additional steps that we use to ground this in reality and ground it in physics and then use it to then extract joint trajectories and put on the robot, we can just do that from scratch now. And kind of one of the next big steps for—I guess this is semi-spoiler alert, but everybody knows this is where it's going—is using that to say "grab the cup." So now I'm you in my house with my robot. I'm saying, "Grab the cup," or, cooler example, "Go get me a beer." Goes and gets the beer, generates its way to grabbing that beer, and then it misses the beer.

(Dar Sleeper at 00:20:33) You say, "Try again." Or more futuristically, it knows it misses the beer, which is—yeah, again, I don't want to spoil too much, but that's very well within reason path forward for these robots—knows it misses the beer, and it keeps trying and then starts to collect its own data. And it can practice and practice and practice, and then it can over a long period of time see, "I grabbed the beer 60% of the time. And these were good examples of me grabbing a beer. These are the bad ones." And if you could separate those two, you can just start back-propagating through the error. So you can just keep practicing until you've reduced the error down to the level where it's nothing, then you've mastered grabbing beer. So that's where robots are headed. And I actually think it's interesting. We didn't just—the world model thought we thought that was, being able to use a world model to generate actions from nothing at all—we were sitting in the middle of last year probably like, "Oh, that's an end of next year thing." We solved it by the end of last year, even a little bit earlier, and then we got the announcement at the top of this year. I think that all of these kind of what seemed like insane, robots teaching themselves, all of these things that felt too unreal to be true are now extremely real. And, you know, if you talk to anybody in all the Frontier Labs around the Valley, they'll be like, "Yeah. That's the path. That's where we're headed." And not to mention, a lot of people, 1X included, having a safe robot to be an agent in the real world, having the world model to generalize against it, needing some extra dexterity to be able to master some of these harder tasks—it's pretty close. So yeah, I'm getting pretty encouraged that not just will your main autonomous—your main experience be autonomous, but it might be autonomous with a level of capability that almost seems insane to us today just in a few months.

(Joel Beasley at 00:22:17) Can you beat one of these up?

(Dar Sleeper at 00:22:19) Yeah. You beat the shit out of it.

(Joel Beasley at 00:22:20) Can you—no. Can you disable it? Like, if it started to get an attitude, could you put it down?

(Dar Sleeper at 00:22:26) So the good news is Neo won't get an attitude. That's the first one. No. No. That's not—

(Joel Beasley at 00:22:31) That's not the hypothetical we're playing with, Dar. Yeah. Yeah. Come on. Let's play.

(Dar Sleeper at 00:22:35) Can you—it's so—so Neo is actually—it's a version of version. It kind of fluctuates up and down. Like, okay, we want to make it more reliable for this version. And then you start off by gaining a little bit of weight, and then you start doing the weight reduction down. So you got the reliability, and now you got the weight back down. So Neo, as it will be shipped to consumers, is almost 60 pounds, with the same strength as, you know, any other robot, if not greater. So it's very capable at that lightweight. So, you know, I just say that because I don't want to make people think our robot's incapable.

(Dar Sleeper at 00:23:08) Hyper-capable robot times being extremely lightweight times these tendon drives. So—it's where it gets super techy and people on the outside don't really get this difference. But if you're in the valley, if you work on humanoids, you really get the difference. It's like every other robot in the world, actually. I don't know if there's any other tendon-driven robots that are in the kind of humanoid space, full dexterous hands, feet, legs, walking, all that. But, you know, it's fully driven by tendons. So instead of having motors in each joint across the body, these big gear systems that are moving at a one to 100 ratio. So a normal harmonic drive robot moving, one-to-one ratio. So when you move this fast, the thing it's here is moving 100 times faster, stiff gears, so it can't back drive. So a human, when you touch me, kind of back drives.

(Dar Sleeper at 00:23:54) And you can try to design that into harmonic drives, but tendon drives are basically much closer to how a human works with having these core motors. So for the forearm—

(Joel Beasley at 00:24:03) It's getting pulled from a pulley. All the tendons—

(Dar Sleeper at 00:24:05) All the tendons from the hands are being pulled in the forearm, and so on and so forth throughout the whole body. Because of that, Neo kind of moves like a human. You know, you come meet Neo one day or, you know, we'll get you Neo to Nashville. You move the limbs, you jiggle them around, and it feels pretty much just like a human. So that adds this compliancy that makes it very safe to get in physical contact with. So if you bump into Neo, he bumps into you, that's not going to be a catastrophic scenario, which it could be very well with another robot. But then furthermore, yeah, you just kick its knees, and it's good. You could just immobilize it.

(Dar Sleeper at 00:24:41) Yeah.

(Joel Beasley at 00:24:41) You're a guy. You're, you know, about six foot, 170 pounds, something like that. You could destroy it.

(Dar Sleeper at 00:24:48) I could body Neo. Easiest thing. You know, as much as it's hard because I love the guy so much that I can't really bring myself to do it yet.

(Joel Beasley at 00:24:57) Yeah. I could. I'll do it for you.

(Dar Sleeper at 00:24:58) Yeah.

(Joel Beasley at 00:24:59) You should make a video. Let's make a—you should get a jujitsu fighter guy to destroy Neo.

(Dar Sleeper at 00:25:06) I mean, I genuinely—I think there's even somebody half my size could body Neo. It's not a very—it's a very capable robot from a strength-to-weight ratio, which is the great part about these tendon drives with these super high-powered motors. But being so lightweight and so back-drivable, it's quite easy to neutralize his ass.

(Joel Beasley at 00:25:25) Well, I mean, I pick up a 60 pound—my eight-year-old, she's 60 pounds. I picked her up and carried her to bed last night. It's not much.

(Dar Sleeper at 00:25:34) I have a terrible joke to make now, but I won't make it.

(Joel Beasley at 00:25:36) Let's do it. No. You should definitely make it.

(Dar Sleeper at 00:25:39) Well, I mean, if you could—yeah. You could definitely neutralize your eight-year-old. Absolutely. Yeah. See, there you go.

(Joel Beasley at 00:25:46) Yeah. That's honestly—are you a parent?

(Dar Sleeper at 00:25:50) No. I would like to be one eventually. But—

(Joel Beasley at 00:25:52) You would like to be one. Yeah. The thought runs through your head. I have an eight-year-old girl, a seven-year-old boy, and a three-year-old boy. And, I mean, yesterday, on a weekly basis, the thought runs through my head: I cannot hit them because I don't know—if you hit them full strength, you can kill them. You can't hit them. But instead, you know, you just wrestle with them. Yeah.

(Dar Sleeper at 00:26:14) Yeah. Yeah. It's the art of the wrestle. The bears do it. You have to do it. It's like—

(Joel Beasley at 00:26:20) Okay. Why the home first? You've got a lot of these guys. You know, I saw one, what is it, Figure, and they're going to IBM. And then you got Tesla. They're going to the factory, and everyone's going industrial. But you guys are like, "No. No. No. We're doing laundry," which is honestly the thing I want to spend money on.

(Dar Sleeper at 00:26:39) Yeah. Yeah. Well, I think it's a lot cool. I think it's fun and cool, and people can connect to it. That's not the right answer. It's actually a lot more strategic than that. So the home really is just a stop along the road. We built Neo to be spec'd to be able to do pretty much any general labor. It's supposed to be just an infinity of helping hands at the end of the day. It kind of takes the world into the new frontier of, once we're not labor constrained, really a lot of new doors open up.

(Dar Sleeper at 00:27:02) It kind of changes the format of society, which I can get really deep into my philosophies there, but at the end of the day, Neo is built to be an agent for taking the world into a whole new world of abundance. But with that said, the home is super important, actually, because not only—there's two main things, and then there's plenty of other things. There's this question of, like, if you can be in the home safely amongst humans, you can really be in any environment. And what that matters for is data diversity. So training big models—if you talk to anybody in the LLM space, you need loads of data, but you also need diverse data.

(Dar Sleeper at 00:27:41) Like, if everything you have is a Winnie the Pooh quote, your LLM will only know Winnie the Pooh quotes. It's very similar for the factory settings. So if your only job is to sort packages, like boxes—box here, grab it here, put it here, grab it here, put it here—

(Joel Beasley at 00:27:56) Uh-huh.

(Dar Sleeper at 00:27:58) It's very narrow intelligence. And what you need is things at scale, because narrow intelligence runs into issues at a few different stages, where it's like, okay, so we're gonna deploy in a factory. We're gonna deploy in a factory because maybe that's the best first use case. I don't agree with that, by the way. I can go to my philosophy there. But we've done this. And so this is why we're doing this: we had robots, EVE, who was a wheel-around robot with claw grippers. And it was doing incredibly cool tasks in logistics and warehousing and security, et cetera. And what we found is you deploy it—A, you run into the whole adoption curve thing where one partner takes so long to get off the ground because there's loads of red tape, et cetera.

(Dar Sleeper at 00:28:44) You know, internal big company bureaucracy stuff. And not to say that that's the end of the world—it's actually needed. But it makes one deployment equal this long sales cycle. And once you deploy them, you have to start with this many, then you move to that many, then you move to that many, and you have to hit a 99.9% accuracy. So there's a lot of things that slow down the deployment process in the meantime.

(Dar Sleeper at 00:29:09) So for the home, by the way, if me getting you your beer has to be 99.9% of the time, we probably wouldn't choose a home. But if it works 98% of the time, I think you'll be pretty happy. I think that's true. And so the critic is a little bit less strong. So there's a bit of this speed-to-adoption stuff.

(Dar Sleeper at 00:29:26) But then once you start actually doing the task, you're like, cool, we got a hundred robots out in the world, and they deployed this cool factory. Let's get it to the next use case. You actually have to do the whole process over again where you go and you master that task, and you don't have a distribution of data that lends itself well to that task. You don't really learn anything from the last task that lets you make this next deployment any faster. Versus if you go into a general environment, you're getting generalized data.

(Dar Sleeper at 00:29:51) You end up in a situation where all the data you're getting is making your model better and better at everything, not just one thing. And so, yeah, I can get into economics and scaling and generalization and why it's important, blah blah blah, but it's kind of that simple. And if your intelligence is growing very narrowly, you are missing the train for generalized intelligence, which is ultimately—if we look to ChatGPT—LLMs didn't really hit their exponential curve until you could just ask it anything and it just did anything. And because there was machine learning models that you could have write a poem for you before ChatGPT, well before. Like, years, years, years.

(Joel Beasley at 00:30:32) 2012, I was playing with those.

(Dar Sleeper at 00:30:35) Yeah. Exactly. And so it's like, this doesn't become something that has real impact or has real scale until you can hit that general. So you want to find the fastest path to being more generalized. And not to mention, it's not just—like I said, the home is important for data diversity.

(Dar Sleeper at 00:30:50) But because you're in the home, you can really be anywhere. It's important for everything. All these other robots end up in that industrial path because it's kind of the only place they can be. Walking around humans and bumping into them and being amongst humans, living and learning amongst humans, as well as executing tasks amongst humans—you can't do that if you're gonna potentially kill somebody while you're doing it.

(Dar Sleeper at 00:31:10) You can't do that if you're gonna hurt somebody every time they come into a collision with you. That just doesn't work. And so the home is kind of—it's like, if you can solve that problem, you can really solve a lot of the problems. And then not to mention—so there's the data diversity, there's the kind of truth of how safety matters and how the home becomes this tight constraint box for, like, if it's safe enough there, it's safe enough for everywhere. So it kind of also plays into adoption.

(Dar Sleeper at 00:31:33) And then more and more and more, when you look backwards at the PC revolution, you look at all these different revolutions of technology. It never really took off until consumers were the ones that made the choice. When you're dealing with consumers, it's basically, how good is the product? If the product's good enough, it will just fly because consumers will adopt it at the rate in which it is fitting their expectations and product. Versus if you go down the enterprise path, you really land yourself in a world where you're only scaling at the speed in which you can go acquire customers, go through these long sales cycles, spin up a new department for each one of these use cases across the board, like this account manager for that one. It's not the right way to entrance a new product into the world.

(Dar Sleeper at 00:32:14) So consumer adoption is also a great unlock for making sure that there's always gonna be demand for these robots. And if you think about why these other companies are doing that path, we're doing the home path. And not to mention, home is really just one thing. There's also other interesting ways to not end up in the kind of, like, we're stuck in a cage in an industrial factory trap. But it really comes down to the fact that Neo is a uniquely capable robot in the sense that it can live and learn amongst humans.

(Joel Beasley at 00:32:43) That is really interesting. I liked when you were talking about the training the models needing a variety of data because that's real important. And it made me think of, like, humans are like that too. Like, if you replicate human DNA within the same family, you get poor results. If the LLM data is too close, you're not getting good—but if it's diverse results, I think it's interesting when there's parallels like that between the different types of technologies and what we are too.

(Dar Sleeper at 00:33:14) If you think about, like, now with world models, how world models generally work, which is like—and I should preface everything with I'm not the engineer here. I tell the stories of the engineers. I design the products, like industrial design and the UX UI, blah blah blah. And I end up so close to it because you can't really design a humanoid without getting deeply technical about it. You can't build an experience around technology you don't understand. But that's the preface necessary because I will butcher a few of these explanations, and I don't want any of the CTOs—this is CTO podcast, and I know CTOs—

(Joel Beasley at 00:33:51) They're very kind, though.

(Dar Sleeper at 00:33:52) Yeah. Yeah. I hope so. The way a world model works is you basically are just visualizing yourself grabbing that. You can talk about it from the more, like, what is actually happening. You take a frame, you generate a path through video, then you take the joint trajectories, whatever it is. And then you do that over and over and over again so you can handle the dynamic way, whatever it is. But what is really happening is you go from a world of thinking and text, which is what robots were doing before, where it takes all the objects, identifies them, says what it is, and writes instructions to try to find its path through the world. It's a super inefficient way of kind of intelligence. It goes to a space where you're visualizing, like, okay.

(Dar Sleeper at 00:34:32) There's that cup there. The first thing a human does is they kind of visualize their path to grabbing that cup, and they use this sea of knowledge from the past and other times that they've actually executed that and kind of tuned—RL'ed their way through—understanding how to do it to then have a visualization of success, and then they do it. And that's really what the world model is doing. And so it's like the analog between human thought and robot thought now in the world where robots are reaching what is probably going to be the way of thinking that is actually most natural to them and how close that is to how humans think. And then you go to, like, okay.

(Dar Sleeper at 00:35:07) The first times it does something, it fails at it. And then, you know, it reaches out to grab the door handle. It bumps its knuckles on the door and it's, "Ow." Robots don't actually go, "Ow," but it bumps it—like a kid. It goes, like, "Ow."

(Dar Sleeper at 00:35:20) And it says, next time I do that, I need to make sure I actually grab the handle because it seems like that was the thing that got me through the door where I pulled down and didn't open enough. And then so that tells me I gotta actually pull it further down and then open. Like, this type of trial and error learning that robots are now about to start kind of more and more and more often participating in. It's very similar to how babies, like kind of—babies and toddlers and kids—start to learn about how to interface with the world. So it is, yeah. It's—I don't try not to get too about it. The other thing is anything I need is just technology, and you shouldn't get too lost in thinking sage wisdom or—I guess, you know, don't smoke the grass too hard—but it's unbelievably cool.

(Joel Beasley at 00:36:04) It is. Absolutely. I think we're all gonna wake up one day and realize that we're just advanced organic robotics. I'll be like, oh, okay. That makes sense.

(Dar Sleeper at 00:36:13) Yeah. I think—yeah. H. sapiens is so odd. It's like, this is how this happens. It is.

(Dar Sleeper at 00:36:19) Yeah.

(Joel Beasley at 00:36:20) But if you look at everything, every technology advance that I've seen in my forty years or whatever has been leading me to that path. Like, they just photographed a neuron making a new memory. You know, they have a video. You can actually see how it forms, and it looks just like a network. And it's just so interesting to me. If the DNA storage, do you know about this? They're like, the best density of storage that we have is actually inside of DNA. Like, we can write data to DNA. I don't know if you know about this.

(Dar Sleeper at 00:36:50) That's very cool.

(Joel Beasley at 00:36:51) Yeah. But we've been doing it for years, but it's just recently gotten pretty efficient. There's a company out of Boston—

(Dar Sleeper at 00:36:58) Like CRISPR you use for this?

(Joel Beasley at 00:36:59) CRISPR is for editing DNA, like in biologics, like making edits to the DNA sequence. But we can actually—like, I could take your hard drive and then convert it into DNA, and then I've got a physical vial of DNA matter that is that data. And you can encode it and decode it through—you can read and write to it. And large companies are doing this right now.

(Dar Sleeper at 00:37:25) I don't know why. Like, my head went to, like, can I store all of Clavicular's streaming history in my own DNA? Like, that's where my head goes first.

(Joel Beasley at 00:37:39) I do wanna talk about the US versus China and the humanoid race because I watch those YouTube videos. I follow the humanoid robot thing for, I think, about seven years now. So I have an annual review I do of looking at them on YouTube videos. And China's always like—it might not be better quality, but they're doing more of it. There's more companies, more variants. Are they crushing us right now?

(Dar Sleeper at 00:38:03) Are they better than us? That's a fun question.

(Joel Beasley at 00:38:05) Are they?

(Dar Sleeper at 00:38:05) Do you ever worry that when you podcast, you're just gonna, like, cascading effect? Like, oh geo—you just caused a geopolitical crisis from you asking me to question it.

(Joel Beasley at 00:38:14) I would probably have more money if my podcast could do that.

(Dar Sleeper at 00:38:18) Yeah. China is—I think—I mean, they're really good at building humanoids. It's undoubtedly—I think Unitree's quarterly—I mean, it's worth looking at the real numbers. I'm gonna butcher the shit out of this. I think they sold like 5,000 humanoids this year and had some billion dollars of revenue.

(Dar Sleeper at 00:38:40) Maybe there's—probably I can get a producer to make me not sound stupid. But that is—those are undoubted—there's like, you can't argue with those results. Like, that's a—shipping is one of the hardest things to do in technology. I hope—especially in hardware. Anybody who's done this—your audience was probably pretty well within the frame of mind. Oh, here we go. Clarification: exceeded 500—pen—okay. They're profitable. Like, it's a profitable humanoid company. It's amazing.

(Joel Beasley at 00:39:08) They are profitable in delivering robots.

(Dar Sleeper at 00:39:12) Yeah. Humanoid company. It's fucking hard. Like, you know, it's one thing to have a platform that's stable enough. And, you know, the truth is they took a great strategy here where the platform is not that stable.

(Dar Sleeper at 00:39:22) People who have a Unitree know this well, like any lab, any human. You have to—you basically become a full-time roboticist with your own Unitree to be able to get your own Unitree going. And then you might as well start a company at that point. But, nonetheless, that's actually really important, and it has started a lot of companies. A lot of people have started their companies around these Unitree robots, and it's great for the startup ecosystem.

(Dar Sleeper at 00:39:46) And I won't go too long about how—I'm not gonna wax about Unitree for too much longer here on this podcast. But it's incredible how they've shipped, and you can't deny those results. And so they're definitely not someone to look past. Like, China as a manufacturing engine is amazing. China as a—I mean, obviously, you know, labor there is a lot cheaper.

(Dar Sleeper at 00:40:14) There's a lot of reasons why they're killing it in the humanoid space. I still don't know—and this is—I've been trying to gather this, you know, I have a bunch of friends that go to China. They go visit all the humanoid companies. I have a lot of friends from conferences, a lot of these—Engine AI. Those guys are the coolest guys in the world, by the way. Super sweet. And I still haven't seen fully a Chinese humanoid company that makes me go, that's gonna be a form factor that we can deliver physical intelligence to the world with. So that's the one thing I, you know, both hyper-bullish 1X, but bullish USA is we have a creativity about us and a genius about us that gives me deep confidence in the kind of trajectory of the US in the humanoid race. But, nonetheless, you cannot discount China. You cannot write them off because they're manufacturing these things in the tens of thousands, and they're shipping thousands of them and making profitable companies out of them.

(Dar Sleeper at 00:41:09) That's just undeniable results. But like I said, I'll go a little deeper on that. A Unitree, for example, it doesn't come with a greatly dexterous hand. It's still planetary actuators. So the energy of that system—they're not the safest thing to be around.

(Dar Sleeper at 00:41:26) Fortunately, the ones that end up all over the place are the little ones that are running around in public. But that's kind of primed for disaster in some sense. You cannot just deploy humanoid robots. It's something that people should be paying attention to. We shouldn't just willy-nilly throw heavy, high gear ratio robots out amongst humans without understanding that, you know, just keep your distance or whatever.

(Dar Sleeper at 00:41:49) But, you know, they're still under-actuation. Their actuation method is maybe not something to me that I believe scales. It doesn't come with hands. It's still, for the most part, their volume version that they're able to sell at a decent price. You know, if you want a G1 that has anything interesting in it, it's still in the order of—anyway, I think it's like $40,000 to $80,000, somewhere in that range.

(Dar Sleeper at 00:42:08) If you get a G1 for like $20,000, it's just a chunk of metal. So the cost is still something that I'm not super worried about based on what we've been able to achieve with our BOM cost here for NEO. So that's the other thing about China—other than the fact that I think they deserve a lot of credit for shipping, a lot of credit for manufacturing, I'm not totally convinced that they have these fully integrated systems that have all of the pieces, that have a level of product vision and a level of—will it end up in a position where, you know, a U.S. company finds the scale path? I believe 1X finds the scale path to really just rip past the competitors across both Western world and China. But yeah, I think the landscape's super interesting. It's not to mention, if you think about it from a geopolitical standpoint, technology is kind of one of the things that determines your standing in the world, both from an economic standpoint but from a better technology.

(Dar Sleeper at 00:43:09) More power is kind of the way that I think the world looks at this. And this isn't really how we think about humanoids. We're just like, we want helpful robots everywhere in the world. But, you know, when you start getting into the realities of geopolitics, it's super important that the U.S.A. is super successful in its humanoid program. Yeah.

(Dar Sleeper at 00:43:27) And the good news is we have a pretty self-sufficient—we have a pretty good system here. Capitalism really works. You have all these startups that are competing against each other. You got all these geniuses running around. Like, I'm gonna choose this startup to apply my genius to, and we're gonna compete here, and we're gonna innovate and innovate and innovate. We're gonna take these crazy asymmetric bets that people in other countries don't necessarily do because U.S.A. But the importance of us—yeah, the importance—

(Joel Beasley at 00:43:55) The U.S. loves you guys. Yeah.

(Dar Sleeper at 00:43:56) Yeah. Yeah. Yeah. I'm—yeah. Yeah. I got a—I have a farm, and so does my family. So yeah. I definitely love the country a lot. It's really important that we are super successful in humanoid robots. But the landscape couldn't be more interesting.

(Joel Beasley at 00:44:14) Yeah. Does the government help at all in the sense that, like, do they give startups loans or grants, or is there any way the U.S. as a country is standing behind, or is it all just private capital, people doing their own investments?

(Dar Sleeper at 00:44:30) I believe Boston Dynamics had a pretty large DARPA injection. It's almost questionable to say—like, don't get me wrong. I'm not saying the U.S. shouldn't be helping a lot in the humanoid race. I actually—I try to make sure to admit where my boundaries of genius are. Like, you know, we haven't talked that much about design. So I'll say everything has been on the boundary of my genius to a degree. Policy and government stuff is way outside of my boundary of genius.

(Joel Beasley at 00:44:48) Yeah. Me too. Yeah.

(Dar Sleeper at 00:44:57) Yeah. Yeah. But I know that there's a lot that the government's doing. If you just look at Stargate and you look at the Department of Commerce and the kind of partnerships with the Japanese and all that—there's a lot of stuff that the government's doing to try to help support robotics, AI, energy, all this critical infrastructure for the future. So I think it's very clear to me. I gotta watch a really cool podcast with this kid named Ty Morris. Are you familiar with him?

(Joel Beasley at 00:45:24) No. I'll check it out.

(Dar Sleeper at 00:45:31) I would look him up. He's a Silicon Valley inner circle guy, kind of young guy, my age. Call me—good to call myself young there. He's just started ripping through the world having the coolest interviews ever with all these different founders. I'm trying to think of some of the recent ones that were super cool. But he had Scott Nolan from General Matter and Secretary Wright of Energy on to talk about what the U.S. is doing about this world. Energy is flatlining the U.S. That's not good. The industrialization of a country is usually directly proportional to its energy creation.

(Dar Sleeper at 00:46:00) And they talked about the next five to ten years and what the U.S. is doing right now and what the U.S. is gonna do over the next five to ten years to really bolster the energy economy here. And what's cool is, I think, actually, the U.S. government's been talking more than ever about what their plans are, actually executing on the plans. Like, here's all this money.

(Joel Beasley at 00:46:26) Did you see the space stuff that they did? Oh my gosh. I missed that news alert. All the space stocks went up like 200%. They announced some initiative, I think, in December, and they put all these—and then all the companies in the sector were part of the parts supply chain for the contracts for—what is it? To go to the moon? I think that's what it is. I think it's to build a base on the moon. That's it.

(Dar Sleeper at 00:46:47) And so—

(Joel Beasley at 00:46:48) I love when that happens. Very sad I missed that press release and didn't buy it. For the space—I think it was December. And then by, like, March, I think they were up like 200% or something.

(Dar Sleeper at 00:46:59) Yeah. Yeah. Because I remember somebody told me to buy a lot of Rocket Lab stock. Like, just—yes. Yeah. But that was well, this was a year ago, and they're like, Rocket Lab, space exposure. And I was like, I don't know shit about space, but I love space. Fuck yeah.

(Joel Beasley at 00:47:12) It's always great before they announce—before they announce the space contracts.

(Dar Sleeper at 00:47:16) Yeah. I know. The way to do it is once they announce and you see it go up 100%, buy option calls based on—you take a mortgage out on your house, option calls, and then put it all down, and then it'll surely grow the next day. This is not financial advice.

(Joel Beasley at 00:47:32) You do have a hard stop. I apologize. It is two minutes after the hour. So I'm just gonna wrap it up. Where can people go to learn more about this robot to make their pre-order? They want this thing helping them out.

(Dar Sleeper at 00:47:44) At Rad Backwards on Twitter, Dar Sleeper. I'm kidding. 1xbottech is the website. I'm working on getting the .com. If the guy that owns the .com would like to answer my LinkedIn message, please do. At 1x.technologies on Instagram to learn about—on Instagram, at 1x_tech on Twitter. And I'm trying to figure out how many of these I have memorized. They should all eventually be at 1X. I'm working on that.

(Joel Beasley at 00:48:12) This is too much. What's the website?

(Dar Sleeper at 00:48:16) www.1x.tech.

(Joel Beasley at 00:48:18) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.