Manufacturing 50,000 Humanoid Robots Next Year | Bernt Børnich, 1X

0:00Today I am sitting down with Burnt Barnek, the founder and CEO of OneX. You want to ship 50,000 humanoid robots next year. So can we start off with you just walking me through how do you actually go from just R&D to actually shipping that number of robots in 12 months? Yeah, so I mean, the factory we're in now, in Hayward is built out to a capacity fully ramped about 10,000 units. So that's really been like the goal this year. We're not gonna do 10,000 because we're gonna be pretty much towards the end of the year when the line is fully ramped. But in parallel, we're building out St. Carlos. So to make this happen, we have to do some parallel building. And that's gonna be 100,000 units at full capacity per year. So 110k per year, but of course, most of that volume is gonna come online pretty late next year. The real gating item here is really, kind of like not just shipping 50,000 units, but I'm sure that you don't get them back, right? So we have to kind of scale our volume at a rational rate. Phototypes are easy, like is it cliché, but it's true.

1:01Prototypes are easy. Production is really hard. We've turned through a lot of issues here to kind of get to where we're confident in ramping volume, but whenever you take like the next big step on the volume, you always find things that you don't find in the previous one. So like, it's just statistics, right? When you do a 100 units, you don't find the same things that you find when you do a thousand or 10,000 because things that are rare just get statistically significant and like your yield really needs to come up. So we hope to find a good balance where we take the time to really iterate on the product and like get the lessons learned and get that back into the fleet and then go into larger and larger volumes as we eat. Kind of cross these gates, quality gates, but the goal is 50,000, so it's gonna be a lot of work. I was thinking about like cars when you were talking and with a car, if you're testing some car just yourself, you're gonna experience most of what it is to own a car just driving around. Whereas with like a humanoid robot, the number of, you know, random, super unusual behaviors

2:03that someone might have this robot in is like multiple orders of magnitude more than a car. And so going from like zero to 50,000 cars is incredibly difficult. Going to 50,000 robots is like, I think probably like one or two orders of magnitude more difficult. How do you kind of plan for that? As I think there's pearls and cons, right? So a humanoid robot is a lot of parts is quite complicated but it is not the same number of parts as a car. And I see a lot of companies kind of compare humanoid robots to cars. And I think if you do that, you should kind of like take a step back and think about what you've made because a car is like 50,000 parts. A well-designed humanoid robot is probably about 1,000 parts and a car is like 4,000 pounds, a well-designed humanoid robot is hopefully less than 70 if it's gonna be safe. So it is actually a simpler system in that sense although like the individual complexity and the removing parts is larger. So you get some on that side but you're of course entirely correct also that like it's a way less constrained environment

3:04that you're deploying into. So you're gonna see a lot of more weird shit. That's just the way it's gonna be. I think a lot of it comes down to expectation management. Like it's gonna be bumpy. It's not gonna be perfect. I really hope we can provide a very good customer service and ensure that whatever happens, we're treating you well and we really figure this out, iterate on this and make sure that you don't have to face the same problem repeatedly. But of course we also need to deploy into some quite structured more like less variance type environments for parts of the volume and not have 50,000 robots into 50,000 different tasks that would not be manageable. Some in place have the robots in their own home and I think you're one of them that's just got this robot walking around this environment and like experiencing it and you're kind of like a few years ahead of everyone else. The other thing too is you probably had multiple versions

4:05of your robot in your home and so I'd like to know what it feels like on just some personal experience level to have the first evolution of that bought in your home environment and also how that's progressed and like how the trust has been built over time. Yeah, it's actually kind of interesting because it's about three years now since I kind of had the first, like, let's call it pilot at home where I actually had our previous robot EVE, the wheeled humanoid, which is a bit like bigger and more heavy but very capable at home and it was a lot of fun and we learned a lot and it was also one of the main inspirations for a while like we had to kind of speedrun getting to like the full biped to be able to do all these tasks around the home and also in like older, diverse environments. Legs are actually very useful so for manipulation. It was a hard lesson, Lord. But I think the evolution over time is surprisingly slow. Like it takes a long time, right? So and this is often because you find something

5:07that doesn't really work and you need some significant data on this then you need to iterate on your hardware and you need your hardware to kind of redeploy and now you get new data and we're not in that many homes so like it takes time to kind of run the loop. But luckily we started very early so we've had a lot of time to think about these things but I think mainly to me the kind of like important barrier that has to be crossed on the trust side is safety and that's something that we're still working really hard on and we really have a big goal to kind of like prove out safety this year in a more formal manner so I'm very excited about that and hopefully we can share some more about that towards the end of the year. How do you kind of decide on there may be a thousand tasks that the robot can do in like a lab environment but then like you said kind of there's this safety bar that you wanna hit before you ship it to, your kid is around it. How fast do you go from creating a new task to actually having that in the environment and like testing it in your own home?

6:08So one of the beautiful things about this system is that since it's so close to a human and how it operates it's actually pretty straightforward to also operate the robot as a human. So this concept of kind of like an expert in place or like an expert operator allows you to really test a lot of tasks way before you're able to automate them which is very useful. Gives you kind of like the kind of customer impression how is it to live with this? How well capable is the hardware to do these tasks? So I think like largely most of the things we've already tested it out and really the big thing now is how do you automate it? And here we follow a very different approach than most of the players in this space. I'm a very big believer and we can solve the general problem. We don't need to go and like automate every single task individually. There's a lot of tasks and there's a lot of complexity and also there's a lot more variability than you kind of like would think it is. So it's kind of like deep rabbit hole, right? To solve every task. And I see humanoids almost more like this bridge

7:11between human data and human intelligence and the machines. We design robots that are so similar to us as possible even in how not just how they move but how they interact with the world, how stiff they are, how compliant they are, all these things so that we can take all this data that exists of us in just video and everything else out there on the internet. And train models that allows our system to be very generally capable across essentially all tasks but less capable than if you specialize right now. But it allows you to solve the full problem way sooner. And I think that's generally what's also needed for the home. So I'm very excited about that. The new one X-World model labs that we stood up, an amazing team. And we're kicking off kind of like the first new big training runs now later this month. And I'm very excited to see how the new model's gonna turn out. It's a bit early yet, but there's some very, very promising promising results there.

8:11And I think we have a pretty good shot at solving the general problem on this training run. No, we're okay. But getting to, it's kind of like a bootstruck problem, right? Because you train a model that is very capable of essentially attempting all these tasks. And it does a good job of attempting them and sometimes succeed, sometimes fails. And this gives you more data. And you kind of fold this into the training and now you get more data. So now the model is even more capable. And then you kind of hill climb like this. But the really important thing here is to have models that are so capable that you don't need the human in the loop to generate the data. You just need a large fleet of the point robots. And then you will get a very capable data of both successes and failures. And you close the loop in a real world essentially to learn, right? And then you need the capable model. You need a safe robot. Because if it's not safe, then it will harm itself for the environment when it's trying to do things. It doesn't quite know how to do. And then you need a lot of deployed robots. So that's one more reason to see if we can get 50,000 of them out by the next.

9:14On the kind of scaling the manufacturing side, what is the process been like going from the first one to like the first 10 to the first 100? And what have been the biggest bottlenecks along that journey? So to some extent, we're a bit lucky. And I like the system has been the sign from day one for manufacturability. So even if you think about like this is a tech tree. And how do we choose to build robots? It's a bit different. So while you normally would have like normal motors, which aren't that strong. So now you need very complex gear systems. These gear systems introduces a lot of like energy and overall artificial dynamics into the system that nature doesn't have. So now you need a lot of sensors and high speed control and all those things to kind of mass. Yes, you're kind of like fighting physics. We've taken a very different approach. So we have some very capable motors we develop in house and also manufacture. They'll also do these tendon drives. And they're drastically simpler actually. So not just more performance, but drastically simpler, like less parts, less complexity, less tolerance, less sensors and all the things that can go wrong. So the system is quite manufacturable in that sense.

10:17Now the downside of course is that since this is a new type of system, there's no supply chain for this. So we have to manufacture all of it ourselves. And that's taken us a decade, right? Like the company is 11 years old now. So we really had to kind of like climb the maturity ladder of these production processes. But now that that is getting quite mature, it's really a superpower because we do it all in house here. So whenever anything happens, we can just go back and iterate very fast and make sure we fix it up three. So I think the most important metric we have right now is we use about four weeks from we do like major changes on the full system in CAD until a new robot walks off the line. And it took us a decade to kind of build that momentum where we can iterate so fast and harder. But that allows us to take the feedback from the line, you know, and feed that back into engineering and iterate in about four weeks. So then you can really, really, really kind of tune in your manufacturing process. So you pick off any mistakes that's only like assembly or anything that's hard to do

11:19with respect to quality yield, calibration, all these things down the line. But there's no like secret sauce. You just need to like grind through all these problems. And in the end, you read the crosser fingers that you're going to run out of problems. If you just fix enough of them, you're going to run out of problems. You still believe. Yeah, I still believe. But like since that's kind of the only way to do it, then you just need to be able to really benchmark yourself on like how quickly can I iterate because that's how quickly I can kind of like turn through these problems. And we see less and less problems. So there's seems to be going in the right direction. So sooner or later, we'll run out of problems. When you're like starting to put these in people's homes that aren't inside of the company, I think the bot at the beginning of next year is going to look significantly different than the bot at the end of next year. And the feedback that you get from January customers are going to be very different. You know, you can incorporate that for the December people. How do you kind of as quickly as possible incorporate new information and new feedback from the robots in the field and then basically spin up new manufacturing lines in order to scale

12:22that? There's a lot you can do now. It just like fleet monitoring and data analytics. That was very hot, very hard to do before. So I think that's a very important part of it. But there's nothing really that replaces the need to just like stay close to your customer and really listen to your customer and get that feedback. And then also like when we talk about iteration here, right? You iterate on the entire fleet. So we do have a big goal in the beginning at least for like the first units that we will do a lot of re-manufacturing also. So like if you find things that we need to fix, then we fix it not just for new robots, we should also fix it for the existing robots. Would that be like literally like quote unquote recall, recall back in? Well, but not necessarily recall, right? Most of it is like the product is pretty small and light. So like generally you can service it in the field. And like if you if we figure out that like we. New left feed on the robot, right? We replaced place to left feed in the field.

13:23We don't necessarily recall the robot, but technically you could call it a record. I was looking at the hand video and I was thinking to myself, this reminds me of this idea from Steve Jobs, where if you look at the inside of like an Apple computer, it's gorgeous. And people were kind of wondering to themselves like, why is this? And it's like because Steve wanted to design something that was both beautiful on the inside of the outside and also functional. And I think when I look at your robot, it's that has that same feel where it's like beautiful to look at even when it doesn't have its like clothes on or a skin on. What's your kind of like philosophy on designing something where even the internals look good? I'm I'm I'm probably like blunt at her in like a special way because I think sufficient the advanced engineering is art. And this is clearly advanced engineering. But I think the beauty of the machine kind of comes from the simplicity in not the complexity. So you take something that's extremely complex and you manage to boil it down to something is like it is the simplest form, the simplest implementation that can fully solve the problem.

14:24And there's a real beauty to that. And I think often you can see this in systems. You can kind of see that like, oh, first principles wise. This is just like the minimum implementation of the system. And there's a real beauty to that simplicity. And that's really what I hope we can deliver. And it kind of comes from itself right? You don't need to design for it to be beautiful inside. If you've done your job on like, how do you design the system? Then this just automatically happens because it should always be the minimum complexity implementation. That solves the problem. That's how you achieve scale. Do you ever have a feeling where you're looking at something and you say to yourself, like it is clearly not there yet because it doesn't look simple. It doesn't look beautiful. Oh, yeah. And like every day, every time I look at the robot, all I see is everything that's wrong. So I'm not that lucky in that sense. Everyone else sees something as like, this is beautiful. I'm like, no, no, no, no, no, no. Like this is wrong. This is wrong. And I think that's probably going to be like that forever. So I'm not the right person to ask that. But I do think, like I find beauty in every time we can do simplifications.

15:28And I think the biggest simplification that's happened over the last year is, for example, wire harnessing and just ensuring the complexity of the cabling really going down tremendously. Like minimum number of connectors, minimum number of wires. And this is just something we drive top down. Like we literally have like a cable and connector budget. For example, there's a certain number of connectors that you have to have for like a first principles point of view, but there's not that many. And every connector that you have more than that, that's an exception that really needs to be defended and needs to be approved because it shouldn't be there. And we should have a plan for how to get rid of it. And this is the same for moving parts, right? Like you want the number of parts in your robot to be the number of moving parts. Two parts don't move with respect to each other like, why are there two parts? Like they should be one part. You can't always achieve this. But you should absolutely know what is the first principles minimum complexity and how far

16:31away are you and how are you driving it down to or set? Because the ultimate goal should be to reach that. Because there been any specific example where there was like a very obvious, this should just be one part that you can think of. This is very niche. My favorite is actually how the rotational part and shaft on the motor is now one part less than what it's always been. Because some very smart people figured out how to do the bearing assembly of the shaft without needing two parts, which usually it's always two parts because you kind of like need to sandwich it in, but it's actually one part. To me, that's very beautiful. It's half the number of parts for that. If you took burnt from five years ago and you teleported them to today or teleported yourself back, would you make any decisions drastically differently than you have? I'd keep the team way smaller and leaner way longer. I think a lesson you learn again and again and again as YouTube development is that the beautiful products are the simple products and the beautiful simple products are made by very small, extremely talented teams.

17:32The more people you have, the more complexity you will get in your product. This type of engineering just does not scale with a number of people. You can scale your manufacturing with having lots of people doing manufacturing processes, testing, automation, all these things. But like the core product design, once you have multiple people, you need interfaces. Once you get interfaces, you kind of introduce artificial complexity. There's a real very material advantage to how much of the system can you have in your head and understand so that you as a single person can make the decisions across the full system. Usually I would tear to like across the subsystem, but ideally there are no subsystems. It's just the system. Right. Now this is not doable for this complex of a product, but you want to get as close to that as possible. When you are trying to keep the entire idea of like this thousand part product in your head, how do you kind of come up to speed as fast as possible with ahead and have the most accurate representation of what this thing is in your head at any given time?

18:32I think it just comes down to like, have you understood the fundamental principles of how the machine operates and why? And if you really understand the fundamentals, then everything else is just logical, like if you have a good design. So then the complexity kind of goes, logical, it's away. And then you look at it and then you're like, actually how do we spend this much time? Like this isn't that complicated. And like that should really be the goal, right? And maybe other people look at it and are like, oh my God, this is complex. But once you really know how it works, then the magic goes away, right? And it's just like this pretty simple machine. I think that takes a lot of time. So you also need to create an environment where kind of people stick around and get that intuition and understand the understanding of the system. But you can also do a lot on the tooling side, which is quite important. Like how do you build the design tools on the Electromechanical side that allows you to iterate on the product in a way that's like, we're finally correct.

19:32And this is increasingly becoming more and more productive because you have so good tools, especially on the coding side. But that's always been a challenge. Like finding people who are very deep on Electromechanical, but are still general enough that they can kind of develop their own tools to solve the problem. Because most people that are very deep on Electromechanical side, they're not necessarily software people. But by far the most valuable people are the people who are very broad across all of these domains, preferably even down to like material science and everything else, right? Like how many domains have you actually mastered? And can you start to see kind of like the correlation between them? So kind of like the boundaries to spare. The best people generally end up building a lot of tools for how to do your job. With long term millions of people are going to have these things in their homes. Ideally, this is something that humans like form a connection with and they start to love. One of the videos that I saw you were like hugging me out. And I was thinking to myself, how do you go through the process of trying to design something that is fundamentally something that humans can learn to love or decide to love?

20:38Yeah. So there's the obvious answer, which is like, oh, you do the right things with respect to like warms and materials. And it's easy to design things that are scary and then they look kind of science fiction and modern because they're scary. Like that's kind of like the simple cheat code. We don't do that. We try to design things with warms like technology that just blends into society. And you shouldn't even think about it as technology like that will be the success. So that's kind of like on the hardware side. And that's hard to do, but it's pretty straightforward. I think on the intelligent side, it's incredibly interesting. One of the things I'm so bullish on with the world models and what we're doing there is the robot actually learns social behaviors also through all this data. And a good example is just like if you get the robot to like, let me know some bags out there and we have the robot to go and like give you a, but in goodie bag of swag, it hands it to you and it doesn't let go, right? And then you grab the handle and then the robot let's go.

21:38And there's there's no data in the training data. Right? The robot doesn't have data of the specific tasks. It's just these kind of like social behaviors and how to interact with people just falls out of the model body language falls out of the model. Like everything kind of collapses into one big old model that understands everything from how to do the task, but also how to like, how do we social? How to simulate older agents like older people? Like if you want to move around and do things in an efficient manner, then you need to understand and predict what all other people do, right? And we do this all the time. And I think that part of the intelligence is very, very interesting, but also very, very hard to get right. I think ultimately that will matter a lot, right? Because that's how you create the trust. That's how you kind of like create that warmth. And I think it's also going to be quite individual. Like people don't, like it's personalizable. Like people don't necessarily want the exact same thing. But I do think that these machines will, they will learn the last for a long time.

22:43And they will become a part of the family and a part of society. And it's going to be very interesting to see how that involves. I think it's going to be quite beautiful to have like this companion throughout your life. That kind of like is always on your side. Remember is everything that's happened. And it's always there for you. And I think I think we could all need that. I think there's very few products that most people would describe as delightful to have in their lives. And I think this is going to be one of them where when people think of the experience of interacting with humanoid robots and neos, it's going to feel like delight. Like things just are happening. You didn't really tell it to do something. Just like magically did some task that you weren't expecting. And you have this little surprise and joy. How do you think about designing something that is just going to be like a delightful product? I think ultimately it's all in the data, right? So it becomes a data problem of how do you kind of like create the desired behaviors

23:43through how you tune your data? I think there's still a lot of unsolved research there. Just like a lot of work to do. I don't have the solution. Like everyone else were working on it. But I think it's something that will really, really, really greatly improve the enjoyability of using the models. And it's been a lot of discussion lately, right? When you look at the best LLMs now, they're kind of becoming worse personality-wise because you have to be so careful to make sure they don't do anything wrong. And we're clearly going to have the same problem here. Like we want to make sure these robots are caretakers of humanity and like they help us. And finding the balance there I think is going to be very hard. So yeah, that's something we're going to work a lot on. And I do hope that through kind of exposing the robot to the right kind of like behaviors that we would like to see in other people.

24:45This is emergent behaviors. It actually just occurred to me if you have a robot in a home environment, it doesn't not all homes are actually happy. There might be learned behaviors that are not good. But remember, most of the LLMs are mainly trained on discord forums ready. Like worse forums than that. Like it's all over the place. So it's like the dark side of humanity. And then you take that and you know there's this great meme, right? Where you have the big Gutulum monster and then someone slaps the smiley on it and says like, now it's ready. And that's kind of like what a modern model is. You have this like beast and then you're just like, no, no, you're going to behave well. And I think there's better ways to do that long term as you get more data. Right now everyone is extremely data constrained. So are we right? But as you get this incredible tool to learn about how to operate in our world, maybe you can be more selective on your data.

25:46I don't know, but maybe you just encourage everyone to like be kind to our robots and your robots will be kind to you. But there is a balance here because the willingness to kind of like break rules and explore is a large part of how we learn, right? It's how you get new ideas and creativity. And maybe that's one of the reasons you don't see that much creativity in models today. You don't really kind of encourage and expose that kind of behavior. How do you expose more of it? Like if you were to try to drive the robots to do creative things, how do you like see that? I think play is just incredibly important. I dream of a future where robots kind of like curious and playful. And if they don't have anything else to do, they don't want to figure out how something works. I don't know, go shuffle your feet in the sand and figure out how the dynamics of sand works. It gets, of course, it's pure speculation. But I think the ability to kind of like verify your hypothesis because you literally have the real world to your grounding things.

26:48Hopefully we'll give you some more ability to explore because you can actually figure out where you're not your hypothesis for a Sultan. What you thought I would. When we were on the drive over here, I was watching this kid just randomly. We were stopped at a stoplight. And this kid was just feeling around and like touching the world. I think it was like some handrail and they were touching it and they just feeling it around. And then it just occurred to me, there's all the tasks that you want your robot to do in your home. Maybe you wanted to do the laundry, make you coffee, all this stuff. But then probably most of the time, if you're at work for eight hours, this thing is just their sitting idle. But you could potentially like have it collecting data by taking actions and just literally like a child experiencing the world, you know, experiencing the home doing things. And I think also like, we haven't talked about this yet, but you know, there's the concept of the Neo platform where we open this up to developers and everyone to like build on it, which I think is going to be incredibly important because it's such a big problem. And I don't think we can solve it alone.

27:49We really need to enable as many people as possible to do the development and like partake in the journey. And these robots will get exposed to very different environments and very different data. And I think that's also very important because you want to like, as far as the distribution as possible. If you think about the home, I think you're partially right. But I think also actually as these systems become generally very capable, our standard will just go up. Right. So like, everything is going to be like perfectly full. All of your clothes will be ironed. Like, even your sheets will be ironed and they will be like perfect on your bed. And like all the glasses will be on the line in the cabinet. And like, there will be no dust anywhere. And like, you can create a lot of work if you want to. Like if there's like, if there's no cost to the labor, you can come up with a lot of things that are like marginally useful. But I'm going to do it anyway because like, why not? So I'm not sure there will be that much idle time. We'll see. We'll see. You know, when you were talking about

28:51perfectly arranging the glasses, it was I was thinking like most humans, like my laundry is not perfectly folded. So if you will emulate it me like it's it's folded, but it doesn't like, you see creases, you know, if most people's environments are imperfect and the way that humans actually operate in the real world is just constantly imperfections. How do you collect like the best real world perfection data? I don't think you do. I think you just learn how the world works. And like we learn that like, there will be wrinkles in your t-shirt if you do like you do. And sometimes you will have the like random example where like it's actually straight, right? And both them are equal important because how you learn how to get from one transition from one state to another. And it more becomes like how do you ensure like that you have the intelligence emerge over this so that if I say actually I want my laundry to be like perfectly straight, then the robot

29:57understands how to do that. It's a pretty good path there. It's a pretty well explored problem, especially self-driving. It's a good example of this like most drivers are terrible drivers. For driving, it's a very it's a very thing you want the robot to do exactly, you know, perfect job every time. Yeah, but like if you just clone human behavior in traffic, you're not going to get that out, not at all. So you can't actually just do that, right? You learn from all the data about how people drive. It's very useful data. You learn what happens if you do something. You don't learn how to drive. You learn the consequence of driving, right? Then when you have an understanding of how to operate a car and how traffic works and how typically like all the cars move and like how pedestrians like might move and like predicting all these things, now you can learn how to operate in a manner that is desirable, given that someone tells you what's the desired behavior. But I don't think models necessarily should explicitly learn to a clothing human behavior. That's kind of

30:58like a very old, that's how we usually do things like years ago, but like that's not how you do these things anymore. You learn how the world works and then you can essentially search in that space for like whatever optimal solution that you would want to your problem. If you had to think for a few years and you guys are not developing all the like models or whatever for your robots and there's kind of like this almost mod store, where people can create things and create like new actions or tasks that a robot can do. And then someone can just like download it onto their Neo and suddenly a new skill is unlocked. How do you like do that? I think it's going to be essentially exactly the same as you see in large models today. So there will be a few dominant models just because they have the best data and e-walls. There's like this not so well kept secret in AI that there are no secrets in AI. It's all just data and e-walls. Like why are the best models the best models is because these companies have the best data. They were early, had huge, huge

32:00their basis. They got enormous amount of still get like an enormous amount of data. And since they have so much user data, they are able to train better models. And when you have better models, you get better user data because you get more users and it's kind of like the self enforcing flywheel. There are other models where you can kind of like specialize. You can download an open source model. You could try to like do some fine tuning on it, do like get it better on your specific task. But it has to be a very hard, very niche task before this makes more sense than just using their general model. And I don't think this is going to be different. We see the same trends, the same scaling, the same patterns. It's just robotics has until now been like kind of like a toy problem. Everyone's been working on extremely small sets of data and they've been working mostly on robotics data where it's like, oh, I gathered 4,000 hours of like these specific behaviors with my robot. And like this is a very big data set. No, it's not like 4,000 hours is nothing. Even like few hundred thousand hours is nothing. And if you see the models that people train now of like

33:02equipping people with sensors, right? And then like having that happen at scale. And then they're saying like, oh, we train this foundation model at 200,000 hours of data, which is huge compared to what it used to be in robotics. It's actually still tiny. Like general intelligence emerges at like hundreds of millions of hours a day. So you're not going to go gather that data. That's why I deeply believe that you need to use these machines as kind of like the bridge between human data and machines. And I do hope that through doing this, the models that we train can operate a lot of robots, not just Neil. And you can generally go in that direction, right? You can take a large, complicated, like very capable model, a strain on this enormous diversity of data. You can distill it down on a smaller problem. And then that model will be very good at the smaller problem. You can't go the old way around. You can't take this model from like this smaller problem that only saw that data and distill it up to solve a general problem. That doesn't work. It's a one way thing. So you kind of have to have the most general system that has the most diverse data and like has experienced the

34:04most things. And then you could take that intelligence and distill it down to like more specific applications. So the question is when that happens, right? Because we're not there yet. And in the meantime, there will be very good applications for a lot more kind of specialized models. And that's what we're seeing today, right? If you want to fall laundry, then actually if you go and just gather a lot of data on folding laundry with our robot and train the VLA, that will work better than our role model, but it won't be good at doing anything else. It doesn't have a general understanding of how the world works or like it's essentially intelligent. It won't exhibit kind of like social behavior and like understanding how it interacts with people. It'll just fold the laundry. And that's kind of like the regime we are in today, but it's quite quickly evolving. So I think 2027 will look very different. Maybe not 2026. I think 2026 to specialize models will still be better. But I think 2027, you're going to see a shift towards more like general intelligence. So initially you kind of have to use like teleoperation in some form or another to like, I don't know,

35:06bridge edge cases and stuff like that. You need all the data. That's a simple answer. So like you you use the web data, which is going to be 99% of the data. You use simulation synthetic data. You use sensors that you put people with to get some of that data. You use the eccentric video data. You're going to use robot data with teleoperation. You're going to use robot data where the robot is just learning by doing. You use all of it. I think the specific data of like, oh, I have an intervention in the sense of like the robot failed. And like now I correct the robot. Not that important that scale. It's extremely important that small scale. Like if you want to like solve a specific problem with a very small model data, this data is essential. And the technical term would be that your distribution is very small. So like you know how to be within this kind of realm here, then the robot knows what to do. If it gets outside, you need something that tells us how to get back on distribution. So it knows what to do. But if you have a general model that is very

36:09capable and diverse, then actually it's very hard to get out of distribution. And then you don't need this. You can always kind of like find your way back to where you want because you're still in distribution. It has to do with a robustness of model. First you try on the internet data, then you get a little bit of your own data. But eventually in order to really get the data that you need to train a super genius robot, you just have to have a whole bunch of these in operation. Like if you think about if you have 50,000 next year, suddenly and they're operating, let's say 16 hours a day, you're getting 16 times 50,000 every single day of new data. I assume. When you're thinking about like your bottlenecks for scaling your data is like one of the biggest bottlenecks, just we need more robots in the field. Yeah. And they need to be in very diverse environments and very diverse applications. In a reality, you're almost never data bound, your diversity bound. So when people say like, oh, you need very large amounts of data, that's true. But you don't actually

37:11need just large amounts of data. You need an extreme diversity of different experiences and tasks and things that you haven't seen before. And if you have enough of that, that becomes a lot of data because you have so many different things. But lots of data of the same thing doesn't really help you. Right. Let's say you shipped 50,000 robots next year. What is the breakdown of where those different robots are going to go? Because like maybe, you know, initially I was thinking like 50,000 homes, but maybe you don't want that. Maybe you want like 10,000 homes and you know, 2000 that'll just go into the woods and build a cabin or something. I don't know. I think you just just crossed the stack. That's like a couple years more. But I, one of my dreams is to be able to leave my house and say like, hey, Neo, we're having a part part of this weekend. And I would like my garden to be Japanese. And then just like it'll like go and reconfigure it into our garden. Like build a nice pavilion. And like you come back like a week later and it's like, oh, it's ready. That's nice. But that's going to be a couple years out. So, so this year there will be

38:16a couple of enterprise applications where we have very large customers so we can deploy a lot of robots into specific articles. And then there's going to be some to the whole. And it's going to be a lot to the platform. It's going to be a good mix of all of them. I think it's important to get like information from all of them and like get a list and learn from all of them. But I also think it's important to have some applications where solving the task to actually be truly useful in like a return investment for your customer point of view is very important. Because that's how you drive adoption, right? And you need scale to get the cost down and to get the reliability of everything else. So a good mix of these. We're not quite ready to announce yet what the specific enterprise applications and customers are, but that's in the works. Do you kind of think about it also in the same way that maybe you need a bunch of like folding data? And so one way to get that in an enterprise setting might be like, hailed maybe, you know,

39:17here's 5,000 robots and they just spend all day folding clothes and then you get, you know, 5,000 times 16 hours a day of folding. So that would be one application. We're not going to do folding as our main application. So I can debunk that myth. Everyone else is doing folding. So there's already a lot of folding data. Okay. So everyone else is doing folding. What is the one thing that no one is doing that seems fucking obvious to you? Is that you could you say? No, Someday I'll reach out to you again. Someday being just a few months and tell you like, here's what I thought about. Okay. Next year we can. Okay. We think we've figured this out. We think we have a couple things we're going to do that. It's going to be very material. I remember this one situation where one like oil company bought a fertilizer company and then suddenly like all these other oil companies started buying fertilizer companies. Like the first guy it was just like, I just bought it, you know, for shits and gigs and then a bunch of other people started doing it for no reason. And is it like full? Some guys started on folding and then everyone else is like, well, they're working on folding. Maybe we should do. No, it's actually folding is kind of we also did folding. Like you can go on our YouTube channel. Like you can see us

40:21using eve actually a few years ago to like fold shirts. So I guess we were early in folding. It's been this incredibly hard problem because it's it's soft goods, right? Like so it's you never know how it's going to look like you crumpled and looks completely different. Like a cup, the same cup looks like the same cup where every time I might be slightly dirty, but it's the same shape. Clothing is a different shape every time you look at it. So as incredibly hard to simulate because it's like soft and deformable. So it was just this large open problem in robotics, how do you do folding? Extremely hard to do in a classical sense. And it just happens to be one of those things that are very easy to do with AI. Very little data and it just works. I don't know, maybe someone's working on like a very nice academic paper that can explain to you exactly why that is the case. I don't actually know. But it is one of the simplest things to do. But it's something that used to be very complex and now is very simple. So it's like the most overused

41:24demo because it looks like you're extremely capable because you can fold these clothes, but it's actually not that hard. It's actually quite simple. It's one of the simplest problems. So I love this line. Starting to start up as like chewing glass and staring at the abyss and being a founder is being someone that enjoys that process. When was the last time that you experienced utter disappointment? I mean, this happens every week. I mean, there's always things that don't go well. So what I've learned is you have to enjoy the journey. Like that's the secret. Like there's because there is no end, right? Like there's it's not like, oh, if I just fix these things that are always wrong, then everything's going to be good. No, no, no, no. Then you get to play on the next level where more stuff is wrong. And if you fix that, you get to the next level where even more shit is wrong. And it's just worse every time, right? Because things become bigger and more complex and like, so you're never going to get to the end of it. So if you don't think that's fun, then you're not going to last like you have to really enjoy the journey. And I've always just been a fan of like, I'd rather

42:26have the highs and then endure the lows, then kind of like just being the middle. So I feel very fortunate that I get to have the highs and the lows every week. It's never boring. At least I can say it's never been boring. Sometimes I'll admit I walk it up and thought why the heck didn't I just design an app? Like why don't I decide to build a robot? Like there's easier ways to make money. But you weren't really in it for the money. But no, I wasn't. And that wouldn't have been as fun.