This 22-Year-Old Built TikTok for Mobile Games, and It’s Growing Fast | E2276

28:54What we're building here is actually true general intelligence where you can ask the robot to essentially do anything and it will do a pretty good job at it. Let's go back. We're going to get into world models in a second, but the point you made about Neo, the robot for your home, designed for that, having similar physical attributes to humans, you mentioned like finger skin tension, for example, that allows you to map how humans interact to the world to the robot itself so that way it can learn more cleanly accurately quickly from how the world operates so essentially as you make it more human it learns faster it does right it's just like if if i pick something up i do that in a specific manner with my fingers like i rotate if like if i rotate that one like you just rotated that in your hand right The way you use your fingers to do that, it's very hard to transfer that to a three-finger gripper or a claw or whatever.

29:50It doesn't transfer. Then you would need to go out and gather all that data. That's one part of it that's just so important. It learns faster. If you've got an engineering team at your company, I'm betting there's a solid chance they're spending far too much time on infrastructure. You need your team building your product to delight your customers, not configuring your virtual network. Render is the all-in-one cloud platform for developers that allows you to deploy, scale, and secure your apps and agents with zero ops. Most cloud platforms ask you to split your focus between product and infrastructure, or they force you into platform constraints that you know you'll outgrow in six months.

30:30But just connect your GitHub repo to Render and you are live. L-I-V-E, web services, cron jobs, manage Postgres, the whole stack in one platform. It's time to find out why 5 million developers are already using Render. Go to render.com slash twist and apply. For the Render startup program, you'll get anywhere from$500 to$100 ,000 in free credits, depending on your stage and who your backers are. That's render.com slash twist. Now, the second one, of course, is that the world is made for us. Yeah. If you want to get around the home and do all these things that you don't want to do yourself, you kind of have to be very close to a human.

31:11And I also think there's something quite magical about creating something that the embodiment kind of connects to us emotionally. And the companion part of the product is also so big part of the product that doesn't get talked enough about. It's not just about doing the labor. it's about doing the labor and being your companion throughout life and how this can kind of like help you interact better with technology and I think we're very quickly trending at least for me now that's using this every day in the direction of like I don't talk to my computer I talk to my robot it gets you away from screens and gets you more present in like your everyday life and I think there's a beautiful story to be told there someday about how this can get us away from screens Well, I think also there's an enormous elder care component to this.

31:59I mean, we've all heard about how society is graying, how our elderly parents and friends are lonely. And so to me, like, sure, we have pets. That's great. But pets can't help you with dishes. And you know what? What if you're older, need help, and you're a little bit lonely? Well, we can kind of fulfill a lot of that with one device in the case of Neo. And I don't view that as dystopian at all. I view it as a way that humans are taking care of humans via technology. And to me, that's a good future. 100 % agree. And this is actually one of the main things that gets me up in the morning. I think we are at a point in history where this is not optional.

32:33We have to apply technology to solve this because we have to be able to give everyone essentially dignity in the way you get treated as you age. And people sometimes mistake this for being a replacement. It is not. And I like to say, as someone who has one at home, it doesn't replace my dog. it doesn't replace my kids it doesn't replace my wife it is something new and just like a dog is also something new that you add to your family this beautiful companion throughout life that is always on your side it's helping you with everything it remembers what's going on and I kind of almost see it more like my Hobbes in Calvin and Hobbes yeah if you don't know that reference Calvin and Hobbes is an American child focused comic strip and a young boy has a stuffed tiger that he thinks is real and they play games and such.

33:30So that's, that's the context of it. So essentially it becomes, if you don't get that context, go read it. Yeah. Go Bill Watterson. It's amazing. It's a great story. I grew up on it. So that's, that's very near and dear to my heart. So here's the thing though, you're talking about, you know, YouTube demos and we're not going to name out any names. We're not going to say figure. We're not going to name any of the companies out there who are famous for doing a lot of demos online versus the ability to make something that works in your home. So you have Neo at home right now, and I know pre-orders are up.

33:59It's going to start shipping later this year. We'll get to that in a second, 20K or 500 bucks a month. So how good now is Neo at general household tasks and how quickly is it improving against that problem set? And the context here, Bernd, is that I'm curious about the 1x world model. And I'm also curious about how tele-op plays into this because I love the idea of being able to call an expert into my Neo to help it learn a new task. But at the same time, to me, with my unlettered perspective here is that certainly after enough usage, we'll have closed those gaps because enough people will have needed that bit of help.

34:37So I'm curious about compared to user demos, how we're doing today and how quickly we're improving Neo. I want to back up one step. I'll answer that from like a bird's eye view. So first of all, you asked a question earlier which i kind of covered but all of it which is why the home first right because this is about general labor so it's not just the home but it has to start in the home because we want our machines to live and learn among people and we wanted to like hold the door open for grandma and we wanted to understand the social context of work and this enormous diversity is actually what gives intelligence so if you're in this narrow environment where you're seeing the same thing every day you don't actually learn it's the same for people like we know it's So that's why kind of the home has to happen first.

35:20And that's also why I think like this, let's call it the early adopter program. Like what we're shipping here, the 20K robots literally was called the early adopter program. And it's going to be rough. I'm just going to like tell everyone up front, it's going to be rough because it's the first time in history anyone does this, but it's going to be one heck of a journey. And right now I would say in my home, the robot is doing a reasonably good job it's not doing everything but it's like it's doing my laundry it's doing a lot of the tidying and cleaning it's doing a lot of fun stuff just around like companionship and just being around like opening the door if like a guest is coming over or all these things that's a normal party trick opening a door when the door dash guy comes over is actually a lot of fun just to see what you package so now the way this works right now is there are two modes.

36:14And we're going to listen to customer feedback also. So like, no, this is locked in stone. But like the way it works is there's two modes. There's best effort autonomy mode where you're running the role model and it's doing like whatever I ask it to do. And it'll do best effort. And you can ask it to do things with just speech. You can talk to it. You just talk to it. And it's extremely general, right? The thing that blew my mind the most this other day was like, I was just asked the robot, like, hey, can you pick the post-it note on the board over there and read it to me? And the robot could do that.

36:43That's not the training data. That's pretty magical to me. That's actual true general intelligence. Now, next time I ask it about the same thing, because it was so fun, hey, look at this. Neo, go read the post-it node. It didn't get the post-it node. So it doesn't always work. That's my probabilistic AI. Dang it. And this is actually the magic of the world model. Because what it gives you is, it gives you this incredible general base layer of intelligence, where you can have the robot sensibly, just based on voice, it'll have a sensible approach to essentially any task. Now all you need to do is to have it try a bit and then it learns.

37:21And that's a lot of fun. There are tasks where this is extremely fun and just works. Like it's very good at laundry because if it screws up folding my shirt, I just ask you to do it again. Right. Not a lot of risk. If it's taking grandma's like ancient vase out of the cabinet to put it on the table, I would probably not do it that way. Okay. So there are tasks where are easily repeatable and where failure is not a problem. Then the robot is actually very good. Opening a door is not a good example because we make the robots just like us, right? To be soft, compliant, low energy, lightweight. So it won't hurt itself.

37:55It won't hurt the door. It can just try it until it gets it. So for these tasks, it's starting to work extremely well. For tasks where you don't want it to fail, then we're still relying largely on scaling our fleet also internally to work on these tasks to get them to where there's an extremely high probability of success and also that you have a very good way of ensuring that if the task is going to fail, you identify it early and you stop. This is another thing that's incredibly cool about the world model because it works very similar to how we do when we're thinking, right? So you said you have kids.

38:31So if you're going to go around and pick up the coffee cup with hot coffee, then you will immediately kind of like simulate forward what can go wrong like through your mind immediately shows up like you pouring coffee on the kid or like all these things yeah yeah yeah and then you select the safest kind of like trajectory to achieve your task based on all these constraints and this is literally what the world model is doing it is looking forward simulating like if i take these actions what will happen and that's kind of like a search and then like hey here's the best way to do this yeah and we're working very heavily on the safety side of that current and like how do you ensure that the robot always takes the safest possible paths to do things and how do you ensure that if there's um unduly risk that the robot does not do the task now this is still working progress this is still working progress but it's incredibly important progress because safety has two aspects it's the aspect where like the robot just is physically not able to harm you because it's light enough and low energy and soft and all these things that we've been working on for a decade.

39:34And that's kind of like, that's a separate thing because that's just, you prove out that this is safe. Standard certifications. And then you have the AI part of things where right now for customers, we limit to some extent what they're always allowed to do because we - It's not allowed to cook, for example. It's not allowed to cook, et cetera, right? No hot liquids, no like dangerous items, these kinds of things. But of course, long-term, we want to be able to do this with a very good safety profile. and that's really like I think now the bleeding edge of like what's getting worked on on the AI side to ensure these things can be done safely so in my oh sorry please so in my house that best effort AI mode is what I use when I'm home and it's a lot of fun when I leave for work I just take off my Neo app and I say that I'm leaving do all the daily chores when I come home it's all done and there is sometimes tell-op involved in that to make sure that everything gets done super well.

40:32Yeah. And I don't care. I'm not there. I don't care either. I just want the laundry to be done. I want there to be clean socks for every child when they're needed. You know, like, I mean, that's why people are like, oh, 500 bucks a month. I'm like, yeah, but do you know how much this would save my life? Like, it gets worth it pretty quickly. Now, I want to talk about the world model for a second. You guys have written a lot about how VLMs are not a sufficient answer to general intelligence for physical AI, aka robots. For folks out there who are listening who are not as deep into AI as you are, can you explain what a world model is and what it took to build the one you guys have released and apparently are going to be updating very soon?

41:14VLMs are essentially take a language model and then bolt on some actions that a robot can do. And how does this work then in practice, right? So, okay, you take a screenshot, essentially, of the world. So, yeah, and you have some text of what you want to achieve, and you have a picture of how the world works, and then you make a plan. And then you start executing this plan, and you take a new screenshot of the world and do this. And it doesn't really capture the dynamics of the world. and the beautiful thing about human intelligence is that we understand how the world works down to like we can visualize what will happen when we do things and to do this you need to capture like the spatial and temporal dynamics essentially that's a complicated word but what it essentially just means is we live in 3D and we care about time we see how the world evolves in 3D over time language models is like a 2D screenshot.

42:16It doesn't have time. It doesn't have 3D. And it works incredibly well, by the way. Language models are incredible. I freaking love them. It's just it's not the full solution for general intelligence. It's a very narrow type intelligence that works really well on a subset of problems. And it's very exciting to see that when you actually train these world models which then essentially what we're training them on is, can you predict what will happen? So a very canonical example would be, if I take this and I drop it, what happens? That doesn't sound too complicated. You can do that. And that's like, it's not magical.

42:54Like it learns how physics works essentially. But when you start doing this at scale, there's a lot of magical things that happen. So for example, back to why the home and why among people. To navigate a social situation, to get you a Coke in the fridge, the robot needs to understand how people will behave. Oh, and the world model includes humans. We can see now in the world model that people appear in the world model, right? Because the robot is thinking about what people, how they will look, what they will do, etc. People appear in the world model and behave like people. This is kind of like an Inception-type argument here, but it's kind of like an AGI a complete problem because to be able to fully understand how to interact with the world, you need to be able to fully simulate how people work.

43:41And it is just, in my opinion, it's the natural next step in intelligence, right? So it's not that we're not using language. Clearly, language is a part of our intelligence. But language is not the base one of our intelligence. That is our sensory visual body and how we interact with the world. and you can see this with kids, right? They learn how the world works and then they start to express this through language. Slowly. Yeah. So it's really exciting and we are starting to see, like I said, the first kind of breadcrumbs of things with robot data like this working better than the pure digital data.

44:24Getting us increasingly confident that the future of AI will be models that actually train on embodiment, not just the web. Now we still train on the web. It's not like instead of, it's in addition to, but it's currently a missing component that I'm very excited about. So when you get the early adopter NEOs out into the market this year, and I do want to ask you about how many, when and so forth in a second, but putting that aside for now, as you get more of them out there, that increases essentially the NEO footprint in the real world to test its vision model against reality and I presume greatly increase the flywheel of learning that you guys can then bring to bear on future iterations to its intelligence, right?

45:08100%. Like this is one of the most important things we do, right? And why we need to get it out there as early as possible. That's why I'm excited actually that you're saying it's going to be a little rough for the early adopter because that means you're getting it out probably as soon as you can to get the learning started to make it better. So how many NEOs do you need in homes to have the right influx of data to learn and improve the world model and embodied intelligence as quickly as you like? Is it 100 NEOs? Is it 10 ,000? I don't have a good sense of scale. The honest answer here is that no one knows because it's not been done before.

45:46True. I can give you some first principle numbers. Yeah. We know that pre-training on a data set, roughly the size of YouTube, gets you very far with respect to general intelligence. Now, of course, that data doesn't actually have what we call agentic behavior, and it doesn't have physical interactions with forces. The agentic behavior is actually extremely important, so let me spend 30 seconds on it. When we train on video, you only know what's going to happen next. You don't know what was the action that the agent took. So like, you know, if I'm going to pick up my phone over there, I think, first of all, I have a goal I want to achieve.

46:30I want to pick up my phone. So I know the goal. And then I decide to take an action to go do it. And then you see the result. So these are like the three different types of data. Video only has the last one. Robot data has all three. It has the internal state of a robot. What was it thinking? What was it trying to achieve? What was the actions it decided to attempt? Here you see the result. So this data is way richer. So we hope we can get away with way less data. But if you look at YouTube as an example, about 10 ,000 robots, you will be having about the same influx of data as YouTube has. So that's a good baseline.

47:05That's not a small number, but not an insurmountably large number because you could get half of YouTube with 5 ,000, I presume, it kind of scales up and down. And the YouTube corpus is so big. Like that's actually a pretty impressive amount of data inflow from just 10 ,000 robots compared to all of YouTube. And just to tell the correct story, it's not all of existing YouTube, but the upload rate to YouTube, which is growing, is about the same as the influx from 10 ,000 robots. Oh, okay. I see. I see. I see. Okay. Well, that's still doable. That's very doable. It's very doable. But it doesn't stop there, of course.

47:41Like, the goal here is to create something that is so intelligent that it can really accelerate the progress of humanity. You will have robots helping us build out all our infrastructure, make sure we have enough manufacturing, enough compute, data centers, power infrastructure. Everything is done sustainably because we don't have to cut corners on cost and labor. Why not do everything sustainably? Progress science, like these AI models will not really help us solve all the outstanding problems in science or especially medicine. and without doing lab work, right? Your AI has to design the experiment, run the experiments, qualify and make sure the data was done correctly.

48:27Iterate on this. That's how you do research. And today it's extremely bottlenecked in that models are getting good at coming up with suggestions, but they can't close the loop and check whether it worked. Well, we have to go into the real world for that. And if we're going to do that, we're going to ape human institutions like laboratories and so forth. So it makes sense to have a humanoid robot doing that. It's just, it's funny because when we're talking about Neo today, I'm thinking about having a helper in the house because that's what I need. And I'm fixated on that because it's what I want.

48:57But in your example of improving the intelligence of these general purpose robots over time to do more and more things, to me, they're moving in some sense away from the home and into a dark factory or a dark lab where there's not humans per se. It's just robots doing what we need them to do. So it's interesting that we end up training them at home to be more generally useful out on their own. It's almost like we're raising children as a class of robots and then sending them out into the world. What's the time gap between, you know, Neo getting good enough at the home stuff that you're like, this is ready for everyone to buy one.

49:33And when we can have humanoid robots helping us automate laboratory work to accelerate science. Is that a short time period or is that relatively long? It's quite short. The home is actually the most complex. Really? It's such a good place to start because you have to solve. Then you just like you jump into the water and now you need to learn to swim. Right. So like I do want to make sure that we really strive to satisfy all of our customers and the demand. And that's kind of like what's constraining it now. So exactly when we roll out in some of these other markets really depends on how quickly we can scale in manufacturing.

50:14Okay, so let's talk about that. Before we jumped on, you guys just opened up a new facility in San Carlos. You're going to have, I think you said design and even manufacturing under one roof? Absolutely everything under one roof here. So research, development, AI, production, service. And in production, we also mean manufacturing, development, and R &D. right? We build the machines to build the machine. Everything under one roof. And this, you know, actually, this is one of the things that I'm very proud of with One X and it's quite unique. We literally go raw materials, even further than raw materials.

50:52We develop new alloys and materials all the way off to the foundation walls on the AI side under one roof. And for For zero to one technology, this is incredibly important because the development speed is just going to be a function of how quickly you can iterate across the entire stack. And if you're reliant on suppliers, it's going to be extremely hard. But even more importantly, almost all of the great discoveries that we have made on how to make robots safer, more affordable, and just in general better, they are between the lines of these disciplines of science, right? so it's when someone on the hardware side says like hey i see you're struggling with that if i do this i think your model will learn better uh or like oh you need these tolerances in assembly that's really hard you don't actually need that i can just use a neural net and we can learn how to calibrate this and you can like loosen up your tolerances and we can produce cheaper and it's like this incredible kind of almost like bell labs types think yeah of science and I think it's the most exciting place in the world to work, actually.

51:57It's one of the few, I think, really good examples of when in-person work is not just better, it's much better. And I like that you're having that effect. But at this new facility in San Carlos, when you're at early stages of productivity, how many NEOs can you make per quarter, per year? Is it 550, 500? The factory that we have in Hayward, that is now up and running fully. And that will be the factory that delivers home robots this year. Okay. That can do tens of thousands per year. The one that we're building in Hayward now can do hundreds of thousands per year. Tens of thousands per year is a lot, given our earlier point about YouTube.

52:39But I'm curious, is there enough early commercial demand from early adopters to keep your factory running at that pace? Or are you guys more looking to sell like a few thousand to get them out there to start the faster learning process? there's demand but of course the product quality needs to be there so right now we're being very cautious and we're essentially doing very fast ramp of like a batch then we evaluate in the market because we do we're not at end customers yet but we are in homes and things are under NDA and stuff but of course we're testing this so it goes into homes and some older industrial applications.

53:24And then we get data. We figure out how well this works. We figure out what's not optimal. We rev the design. We do a quick batch again. So we're kind of like running our factory at like full steam, retool, full steam, retool. And we're doing this until the product is rock stable. You only get to do this once. This is the first time in history that anyone does this and ships robots like this. And we want to make it right. That's also why we haven't really given a specific data on shipment. We've said we're shipping 2026, and we will. We're good on track for that. But it is when it's ready, because you'll do this once.

53:59That makes perfect sense to me. I can wait six months. It's fine. I will not wait six years, but I will wait six months. I'm a financial nerd, Bernd, so I'm really curious about the $20 ,000 price point. Everyone knows economies of scale bring down the price of things and so forth. You're doing smaller batches, but does$20 ,000 covered the bill of materials? Does it cover bill of materials plus labor? Is it profitable? Is it super unprofitable? I just don't know where to peg that number in terms of economics for humanoid robots. So I have to take it back all the way to the beginning then. Because if you don't think about the scale and affordability from day one, how are you going to manufacture this at scale?

54:44Then you're not going to get there. Because you essentially get locked into a corner because you're making specific decisions about the design and actually even the technical direction. So if I'm going to make a parallel, I'd say, do you want to make an electric car or a gasoline car? You're not going to make a gasoline car in small volume. And then when you're like, now we're going to ramp, let's make an electric car. No, like you spent the last decade making the wrong technology. It doesn't work. So we really thought about this from day one. There's going to be billions of humanoids on the planet, meaning you need to think about, is there any special raw materials in there that are rare?

55:20Even further, like, how much raw material is in there? Neo just weighs 66 pounds, right, or 30 kilos. Almost a third of most of our competitors. Well, that's a third of the raw materials. That really matters when you're going to make a billion. Yeah, because it dramatically reduces the number of ships that have to bring gallium or whatever, you know. Yeah, mining refinement. At a billion, you're starting to even think about, like, Like, oh, that's a significant percentage of the world's aluminum and magnesium. Like, how are you going to refine that, right? My first thought was when you said billions was, do we have enough metal available to us?

55:54And then I think about how many cars we make per year. But it's not an insubstantial demand on raw materials, resources, and commodities. That's a lot of new product to bring to market. Oh, clearly it is. But you have to think about that. And then you have to think about tolerances, right? How accurate does this need to be? because it sets the cost of refining. Essentially, at these volumes, you take raw material cost and you refine it into finished product. And that refinement step can actually be very cheap if you have the right output. If what you need is something that doesn't need to be that accurate.

56:34So that has to be a core part of your design. And what we're doing here with these unique motors that we have developed that allows us to pull on these tendons instead of the classical gears. One of the earliest discoveries of the company, even before Eve went to the industrial applications. Yeah, yeah. So like Neo is just the nth generation of that technology, right? That is extremely lenient with respect to manufacturing. Like we can get away with very few parts, very loose tolerances, no special materials. And if you then do all this and you start also working really early on how to automate assembly.

57:12And of course, we are using robots in our factory, robots building robots. NEOs are building NEOs. Text me a picture of that when we're done, thanks. Then we can get very low on cost. That being said, it's an incredibly complex system. It's probably one of the most complex systems on the planet. So it's going to take us some volume to get the cost down to where we would make money on that and have a margin. But it is a sustainable business, which is the most important to me. I'm just really glad to hear that because$20 ,000 is the point at which I can probably just buy one, you know, without having to like do math.

57:49But if it was 50, then I'm going to say, I'd rather have, you know, a year of my college paid or something. So if it works at 20, hot damn, that I can see, you know, the real volume coming out and just being tremendous. I'm really excited about this. And I take it, given the work you've done on supply and components and stuff, you're relatively de-risked from China in terms of supply chain issues? There are certain key components that it's very hard to get. Magnets being the obvious one that everyone's talking about. Now we're deep into magnets, right? Because like even the first year of 1X, what I was working on was like designing new processes for magnets because of the way we do our own motors.

58:32And we have some great companies there that we've been working with in in china on like co-developing these specific new ways of doing magnetics um we're working towards also ensuring that this is something we can do in the u.s yeah but and there are amazing programs around this that are getting spun up so i'm very hopeful that this will be something that we can handle in the future but it's not there yet it's one of the many things that need to happen but i mean outside of that most of it is raw materials that is generally accessible in the open market, right? So copper, aluminium, et cetera, steel.

59:11It's cheaper in some of the countries in Asia, but it's also available here. And you don't have to worry about boats, shipping, times, tariffs, and other things. All right. Last question before I let you go, Baren. You guys initially raised from, as far as I can tell, some Norwegian investors. You went ahead and expanded your capital base with investors from around the world uh how much capital do you need to get neo through its launch this year and then into uh its next generation um is that another couple hundred million is that billions of dollars i'm curious the short answer is we have enough money to ship neo okay uh the longer answer is i think the impact that this is going to have on humanity is not yet fully understood it's probably going to be one of the most impactful technologies ever.

1:00:00Everything we consume is physical in the end. And intelligence becoming general and physical is going to unblock essentially a new way of life, right?

1:00:19I am very excited to push that to happen as fast as possible because I think it can build a very, very beautiful world where we all have what we need and where we can all focus on what makes us human instead of doing physical labor so that we can get paid. And that's a long way of saying we will probably raise money because we can accelerate the path. Yeah. But we don't need it to ship to consumers. That's the reason. I'm just trying to figure out how much you need. like uh because if you were a sas company i could literally just like sit down think about your number of employees and your current cac and like i can do the math robots are so different so well let me say like this like if you want to go max max here it's we need as much compute as everyone else right we're going to train the best models now we're going to train way better models than everyone else because we're going to have we're going to have better data i'm not saying we're going to be smarter than them.

1:01:23There's a lot of smart people there, but we're just going to have better data. And then we need to also build out all the manufacturing. So it's going to be a very capital intensive race. That's the short way of saying that. We do have a very potentially unfair advantage here though, which is that we can labor arbitrage our own scale. Because you're making your own general purpose robots. No, I mean, have you thought about just calling up Jensen and saying, look, I'm in a hurry. I'm going to buy so much compute. You should definitely give me another pile of money so I can just go faster now. I guess I'm happily impatient, if that makes sense.

1:02:02I would like this future to become even faster. So I just hope that the investors out there get it and are knocking on your door constantly trying to give you more money because I don't think we can get to this future fast enough. I'm just excited. I love the positivity. It's going to take a future. Sure. So come back on the show when Neo starts to go out to non-NDA people because I'm really curious to see how the reviews come. And I'm going to start asking my spouse if I can throw my name down. But I think she's going to have something more practical in mind. But Bernd, thank you so much for coming on and explaining this to me and walking me through it.

1:02:37If people want to learn more, what's the website? Where should they go? Wonex.tech. Clean and simple. Clean and simple. This is a lot of fun. Thank you for the show. I appreciate you.