見せるだけで仕事を覚えるロボット：Standard Bots CEO が語る物理AIの現在地

I personally think factory, field, and warehouse, legs are not the right form factor. And the reason is because it's a lot of unique parts for the legs. The wheelbase is both fewer parts, it's way simpler. You're getting ultimately a better ROI. You're getting better performance for less cost. Now, that's when you have clear aisles, which you happen to have in a warehouse or a factory or agriculture. Only 6% of factories have ever even tried a robot in the U.S. today. You can just show a robot what you want it to do. So, like, what this enables is you can perform the job some number of times, and the robot learns how to do it and can generalize to some extent. And today it's within the training distribution. I think we're going from a world where we almost deploy no robots, like in manufacturing, 30-something thousand a year in the U.S., to millions per year, tens of millions. And so I think what you see in the next couple of years is serious progress on that. The chat TV moment will become, like, undeniable and very obvious. And so I think that's what's coming. Welcome, humans, to the Neuron AI Explained. Today, we're going to talk about robots and about getting AI out of the chat window and doing physical work.

StandardBots builds industrial robots designed to learn tasks through demonstration, which is super cool. And we're joined today by its co-founder and CEO, Evan Beard. Before we get started, please take just a quick second to like today's video and subscribe to the channel so you never miss an interview or one of our live streams. Evan, welcome to Neuron. Thank you for having me. Oh, we really appreciate it. We're excited. We love to talk about robots. Me too. So for someone, let's say, who's never stepped onto a factory floor, what does a standard bots robot actually do? And, like, what's a job it can handle today that might surprise people? Yeah, so what we do, well, today we're America's largest industrial robot manufacturer. So we've designed the robot, the hardware, the software, the AI stack, all from the ground up. And really what we set out to do is make it so it's easier to automate things in a factory. And I think you covered this in a previous podcast, but only 6% of factories have ever even tried a robot in the U.S. today.

So I think it's a lot lower than people would expect. And part of the reason for that is you get these products, and we're so used to iPhone apps and simple, easy interfaces, and you unbox them, and I've done this, and they don't even have a plug on the end of the cord. And the menus, they've got single letters in them, and it's like you've got to memorize everything in there. So it's like a whole other level of unusableness than I've ever seen. Yeah, so that's where you want to compete as an entrepreneur, right, as you find one of these areas. So anyway, so that's kind of the status quo. And the other part of the status quo is all of the robots are just replaying motions. So it's like the crane game at Denny's where you have a joystick often just like that, and you're kind of going inch by inch, and you hit save. And then you jog it a little more, and you hit save. And, you know, very slowly you set up trajectory and then you can replay it. And that's how we paint and weld today in the 30-something thousand robots that are deployed every year in our country.

But everything has to be in the same position every time in that world. You know, like the thing you're welding or painting, it can't ever be within a millimeter or whatever the tolerance you care about. And so that's the backdrop under which StandardBots was started. And so we set out to do two things. And one is make a robot that's more like an iPhone experience, like designed for mere mortals. Yeah. And so that's a really big deal of people because now you can have your normal team who's not – they're not technical experts. They can unbox it. Most of the robots we ship, we're just shipping a box. Customer can unbox it and install it. And that's like – this is where a lot of robot companies, I think, fall down because they don't hit that level of usability where you can actually scale the business. And so now we're deployed in hundreds of companies across oil and gas, automotive data centers, aerospace, doing – we're actually in, like, I think 350 different use cases now. So, like, insane amount of – and, like, totally different stuff. This is not – so – and so I think – you asked, like, what would surprise people.

I think what would surprise – what surprises me is every week we get requests for these jobs where I'm like, that – people do that? People do that? Yeah. Like niche stuff. Like you'd think it was niche, but it's like need a thousand robots for it. And so that's what surprises me. And I think it speaks to the vastness of the TAM and just like how many physical jobs there are. Many of them repetitive, boring, and dangerous. Many of them really hard to hire for or like average retention for some of these jobs. Like I just saw a job the other week. You have to do it in a refrigerator. So people literally, like, they're working in, like, 50-something, you know, like, literally their whole job is in the refrigerator. Oh. And they told me the average people stay on that job for, like, eight days. It was, like, insane. And I'm like, eight days? So you're hiring new people at that, you know what I mean? Every day. Yeah, like, almost every day. Like, that's, like, the overhead on hiring these people to work eight days. Oh, my God.

Because you're wearing, like, a coat to work and to go. And can you imagine working in a refrigerator? I mean, I wonder if they monitor, like, the steel. Oh, my gosh. They think it'll be fine is what happens. I think they're like, I can do that. How bad could that be? And then they get there, and they're like, no, I can't do that. Yeah, exactly. Oh, my gosh. So you came from building software companies, right? Yeah, that's right. Well, I graduated college. I went to computer science at Duke, and I did Y Combinator right out of school. And one of the partners was building humanoid robots in, like, 2011. and they were these huge, I mean, it was like eight feet tall. This was before AI was the thing anyone was scared of, and so these things looked scary. But this team would come in every week, and I was building a CRM at the time, which we later sold to Salesforce. But this team was coming in, and they were soldering, and they had like a whole machine shop in the back of the Y Combinator office back then. This was like one of the early years of YC. Our partner was Paul Graham at the time, and it was like, you know, very small relatives to now.

And, um, but this team would come in and they were, it was like one of the partners had their company there and they were just having fun and like soldering and tinkering and assembling things. And I'm like, holy cow, you know, you know, when you like didn't know something was on the menu for life. Yeah. And that was like, when I saw that was on the menu and I'm like, okay, I never even realized I could do that. That looks so much more fun than building. So I'm like, okay, how about we sell this to Salesforce? And then I figured out how to go do that. But here I am. That's so awesome. I love that you saw it and thought this looks so much more, I guess, fulfilling. Is that the word? Like fun? Like a thing you would enjoy doing during the day? Yeah. Yeah, because the other thing you learn in Y Combinator is if you don't enjoy what you're doing and you're doing it for 14 hours a day, that's not the best. No. You got to love it to work those long days. What do they say about startups? You get to work on whatever you want as long as it's 18 hours a day or something like that.

No joke. No joke. Well, we've had industrial robots now for quite a long time, and I'm wondering can you kind of pick a task and sort of explain what AI changes about how the robot learns or performs it? Yeah. So the big thing that's been happening over the last few years is all the advancements that are – a lot of the ones that are in LLMs, transformer-based architectures, and other things are now coming to the physical world. And so this happens in a couple of ways. One is you effectively can have like a cloud code for your robot program, and we've rolled that out. So now that's in our summer software release. So we've got customers previously – like we've tried to make it like an iPhone experience, but when you have cloud code doing that, it's even easier. You know, it's like a whole other – You're bi-coding your robot. Exactly. No, that's what it is. You tell it what you want it to do, and it builds out the whole thing for you. And so that's like one way. But then there's this other way, which is, and I think the other thing is even more exciting,

which is you can just show a robot what you want it to do. And there's a lot of ways to show a robot. And, like, I think language is actually one of the worst ways. Like if I were trying to teach you how to assemble an iPhone, I don't know how to do it, but if I were trying to teach someone that. like you would you you words are not the best is like i'm going to show you a video or i'll let you do it and i'll provide correction to you as you're doing it or all of those yeah or all yeah or all those but but probably like reading a reading text is not the best way so no so it so it's you want to provide um tell the robot how to do something either by demonstrating through teleoperation or we've got gloves that you can do to demonstrate um or like a handheld device so it really depends on the job and so what we built so like what this enables is you can perform the job some number of times and the robot learns how to do it and can generalize to some extent and today it's within the training distribution you're not really you know it's not going to go and like all these models are good at what's in their pre-training mix but it can generalize within that pre-training mix

pretty well so if you have like yeah you know you're pouring soda and you've got different cans like you can have like soda it's never seen before the can geometry can be slightly different than anything it's ever seen. And it's able to generalize within that training set pretty well, and enough for this to be like magic. Because you can automate jobs you couldn't automate before. You can't automate better than I do. You know, that's so wild. What kind of, if you don't mind me asking, what types of models are you using, like, I guess, architectural-wise? Are we dealing with world models, LLMs? Yeah, well, okay, so, you know, you've got your diffusion models and you've got your auto-aggressive models like LLMs, the diffusion models, of course, that's how we generate images, right? And so it's an iterative denoising process, and then some people do it in one step. And these type of models have been the trigger. That's kind of what first caught on in the physical world. And there's some intuition behind why that might be the better route, which is, you know, you're outputting, like, where the gripper should go, for example.

Yeah. And with an iterative denoising process, like, I can, you know, I can make a mistake and then kind of correct because I'm continuing to denoise. With an auto-aggressive process, you know, it's like one token at a time. If you go the wrong direction, you've kind of committed that way, and now you've got to get back. As opposed to, like, being able to zero in on a finer and finer point with diffusion, I guess. Yeah, exactly. That's right. So, you know, some intuition why you wouldn't necessarily want it to be auto-aggressive. But what I now believe and what I think the research shows is ultimately it doesn't really matter. You can make it work rather aggressively. You can make it work with diffusion. And so I think model architecture ultimately doesn't matter. You brought up world models. That is where a lot of this is going in that regard. What really is happening there, and when we say world models, it's kind of ambiguous, but what I know you to mean and where the space is going is training on Internet scale video. So it's like every YouTube video, every TV show and Simpsons episode. And, like, you get some – and these models can obviously – they get an understanding of physics.

And you can, like, give it the front of an object and it can generate what the back of it is. And they do that because the model has some understanding of, you know, those kind of objects and, like, how they're shaped. And you can generate video of something and it can accurately show you, like, if you drop that on a table, like, what happens? Does it bounce? Does it shatter? You know, so the models have this sense of how different objects interact with the world and the physics of it. And so if you use that as the base of your robot foundation model, it can be quite powerful. And so that's the idea behind the world models in robotics. And we're really excited about that line of work. And so we you know working there and starting to see it early innings in this NVIDIA obviously has Cosmos which is an open world model Yeah And and then there you know there people training them from scratch and doing the pre themselves from scratch um but it is i think it an exciting route here and what what clear is you can get a big lift by not having to relearn all that stuff yeah how can you tell uh or better

yet we're going to teach a robot a new job who who demonstrates it what kind of specifically does the robot record and how do you know it's it's good to go and ready so uh in the ideal world customers demonstrating it and that's what we've built is it basically a post-training toolkit so customers can provide their own demonstration the reason that's important is because you want it to be scalable and so if you have to go send a team out with every deployment that's not very scalable so we focus very much on that and making the tools now that's a hard problem because you're not just solving like the model isn't just good enough to do it but you actually have to have a lot of UI and, you know, and tutorials, good, good ways to explain, you know, if you want that iPhone experience, you know. Yeah, exactly. And of course, like the, the models and the tutorials and the software and the hardware, they all have to work really well together. So yeah, there's a lot to that, to making it easy enough for like someone who maybe a lot of times didn't, you know, didn't graduate from high school and you had to make it simple enough to be able to pick this thing up and

train an AI model. So that's our bar for usability. That's awesome. Yeah. Yeah, because people aren't hiring, like, I mean, if you're an Amazon, you're hiring robotics engineers. But if you are, you know, a small factory, is it really? I think so, yeah. Oh, wow. It's, but, you know, and I can see where if you were a small factory, small manufacturer of sorts, and you were looking to find ways in, Even medium-sized, like, you may not have enough to keep a full-time robotics engineer busy. Yeah, with that, you know, the other piece of it is you don't necessarily know. If you started a manufacturing business and you've been running that for, it's like trying to hire a software engineer when you don't know how to write software. Yeah. It's like, you know, you don't know how to hire them. You don't know how to interview them. You don't really know if they're good. You know what I mean? They all sound smart, but they might not know anything about the particular software you have. So it's quite a high bar to bootstrap this, and I think it's quite daunting.

And so 6% of factories have tried a robot, but actually, even just replaying motions, you could automate a lot more stuff. In other words, the things you could automate that would have a positive ROI are a lot higher than 6%. You know what I mean? So there's actually the usability gap and the fact that these things are so arcane and you have to hire effectively power users who have memorized their user interfaces. to come and install them for you. Like that is quite a barrier. So like we kind of knew like solve that first. And then, and meanwhile, we're going to make it so you can also, so you asked me earlier, you know, like walk through example, like what? So there's a lot of jobs where the example is, I don't know, take loading, unloading machines. That's a job where you can do it today. You need to, you know, you generally hire an integrator to come and install it. Like I was saying, like your team isn't going to do it. They're just hit start, stop, reset. Now with AI, you can make it. So you can vibe code. You can make it so it can handle things kind of scattered on the table without any structure to them rather than like –

because otherwise you would put them in a perfect grid. That's how most robots work today. So now you have to print the grid, and you have to like put them in the right positions on the grid. And what if there's a hole in the grid because the guy that programmed it probably didn't account for that. You know what I mean? Or what if they loaded it in the wrong corner of the grid, you know? So now there's a lot of just like memorizing this stuff, and it makes it very brittle. And so AI can make it so, like, the barrier to get this deployed, the barrier to, let's say the part changes and now your grid dimensions change, so you've got to go re-mill the grid or print the grid. So it just makes it so you can just, like, plop the stuff on the table and anybody can get it going. So that's, like, a whole other. So that's, like, one example of, you know, ROI can be totally different and, like, the fear about I'm not going to see the ROI can be alleviated. And then there's this whole other category of job was not automatable before. And we're doing stuff like that. Examples like cloning plants. You have to cut up the plant and plant a little piece of it. And we're able to train a model for that, and you can do it in a couple hours. That's delicate, meticulous work, too.

Yeah, we have a lot of repeatability. A lot of people say, oh, with AI, you don't need accuracy in your robot anymore. It's kind of like, well, does your brain surgeon? He can visually servo. He's got eyes. Now, would you want him to be shaking when he does your surgery? I keep saying that one of these days there will be a generation of kids who are amazed that we once let humans just cut us open. Absolutely. Or drive. Or drive a car. Or drive a car with a one-lane road where you're going like inches away. If you still think that you driving is safer, you really need to look at some studies and maybe watch less news because the numbers are staggering. Yeah. No, I think it's probably going to be banned on the highway, right? I mean, at some point. I think you're probably right. If an agent caused an incident at your company tomorrow, what evidence could you actually produce? Most teams can see the prompt and the final answer, but the commands it ran, files that change, or systems it reached in between, that can disappear the moment a session ends.

Origin gives companies endpoint AI observability by recording an agent's work as a trace from who started it and what they asked to what the agent touched and changed. See what that looks like at originhq.com slash AI explained. And now back to today's video. So here's a question. We mentioned a minute ago that like, you know, one of the brittle aspects. I love that word. That's a good choice. Was about, you know, if things move, if they're not perfectly positioned. So what happens with your robots who have been trained in this different way? When a part moves, lighting in a room changes. Say a bulb goes out or maybe something arrives to it upside down. Yeah. Well, I'll give you the answer today and then where we want to go with it. So today, if that's not in your pre-training mix, it may or may not work. Yeah. So ideally, if you know lights are going to go off at night and on during the day, you collect it in both situations. situations. We want to go with that is really having the base model be so smart that it can

generalize on its own. Like if I taught you how to assemble an iPhone, I turn the lights off, of course you can still do it. So that's where things are obviously going. That's exactly what you want. You just demonstrate once. Yeah. Wow. So if the models get smarter, you won't have to cover the whole. It can generalize much better. That's very cool. So where, I keep thinking of this, you know, in the way of language models, but I'm wondering, like, what needs a different kind of model to turn, like, what the robot sees into physical movement? Is it seeing those motions and copying them? Is it kind of looking at it in terms of, like, you know, I don't know, a grid or a coordinate type system? Yeah. Well, I mean, the most basic inputs and outputs of the model would be in a different place, you know, different people do different things.

But the input ultimately is what's your current position? What's your current gripper state? Because you could have a two finger grip or three finger or a hand or a weld torch. So, like, what's the state of that? Is it on or off? Or what's the width of the fingers? And obviously the camera data. So sometimes a wrist camera, overhead camera, both. And so that's kind of your input, and the output is, all right, give me the next 20 places we're going to go. Imagine like a little line coming out of it, trajectory coming out of it. You don't plan that far in the future, but you're not just planning one point. And interestingly, that's actually one of the little innovations that has like really improved performance, is not just playing one point in the future. And so that's basically the inputs and outputs today. Now you can process those inputs even more, of course. An example would be you can give it depth data explicitly, right, like from a depth camera. Or you could say I want the model to learn depth kind of implicitly. So I'm just going to give it like two or three, you know, camera views that are RGB,

and I want the model to learn depth representation inside the model. So there's different views on all sides of the aisle on these things and how, you know. I was just at IROS earlier this week, academic conference for robotics, and there's a whole workshop on neurosymbolic programming of robots. You've got, you know, symbolic programming and the classically, and you've got the neural ways, which is entirely neural. And this consensus of the researchers is you really need both and combine these things. So it's some kind of processing of the inputs, and it's, you know, and you've got the model takes that in and decides what to do. Dang. So are there tasks where traditional programming might still give a better result? Or are there tasks where what you're doing is significantly better than that? Or is it a combination of the two sometimes? Is there? Absolutely, yeah. If you're doing the same thing all day and there's no variability, then traditionally, you know, replaying emotion,

I don't think you want to use AI where you don't need it, right? Like we're not getting these LOMs to do basic calculations probably. We're still going to put that into the CPU and let that run. So I think the same is true in robotics. And I think that like the second part of your question, are the things better at? It's just the jobs you couldn't automate today. And there's so many of those. Or someone wants to automate something and there's like the part changes or the process changes or, you know, or there's emotions too complicated. and so those are the things where, and by the way, that's most of what's out there. If you look at $9 trillion a year in repetitive physical tasks just in manufacturing every year, most of that is too variable for, like, replaying motions. Wow. Do you think there's a point where robots will be able to learn on their own? Like, maybe this is a terrible job, but I'm thinking, like, say a mechanic who has a variety of different jobs every day but understands basic things

and has to do problem solving as they go to figure out the next step. Like I realize right now we're looking at things that are identical, near identical, or at least in this same realm. But I'm wondering, like, do you believe there's a future where it's less like that and like there's, I don't know, some kind of real-time learning as it goes? Yeah, I mean, we have that built in. It's actually a really important aspect because things drift out of distribution. That's just the reality of the real world. Like you're never, you know, the boxes change color because it's a new season. So you have to be able to get on the side. Yeah, exactly. And there's a lot of long tail events. Like you might be palletizing boxes, taking a box off a conveyor belt, putting it on another conveyor belt or a pallet, and then there's a shoe that comes down one day, and it's not in any box even though, you know, the box ripped open. And so you have to – that can't keep going or it's going to jam things downstream. And so you've got to grab a shoe. So you want to be constantly recording data and providing – if the robot didn't know what to do, provide a human correction or a label for that.

So we've built out systems for that, and that's a – it's a – you know. Wow. I think that that's where some advantage comes because when we're, you know, we're deployed in hundreds of applications, you start to see a lot of long tail and be able to collect data on that, and then you have a better model. and like who wants to like I think you going to start to see models take different use cases where they just dominate because the pre mix is so good and it has all the variability in it and the ability to handle the long events And then, you know, like maybe someone else tries to charge less, but do you want it to not be able to handle, you know, do you want to have to be intervening that much? Yeah. That's the – How hands-off do you want to be, you know? Yeah, exactly. That's such a cool idea because, yeah, I mean, the fact is, Our world is not the perfect grid. Our world is a little messy. It's rough around the edges. There are, you know, even with factory equipment, belts break, gears bust, things happen that, you know, boxes get, you know, flipped, dented, banged, opened, torn apart.

You know, there's a lot of edge cases there. So it sounds like, theoretically, as these things are out and working, they're going to get better. with time. Like, if you buy a robot today, it's going to improve with time. Yeah, I think that's true of this new world of robots powered by physical AI. That's so sick. But it's not true of most of the robots out there today. Yeah. That's such a big step and makes a lot of sense. And I could see where companies would start to see intrigue in that if they if they all knew if they all understood what was possible i mean i assume there's a communication struggle there in making that clear to people what is and isn't possible i think it's becoming more clear over time i think if we zoom out a second we're in like 1973 equivalent to the computer industry yeah and i think i think it's not quite dawning on

everyone yet that we're there you know like it's like the personal computer just started to come out and like the products are it's amazing in some ways and another it's like early versions but um it was very clear to some people like in but it was not clear to like the head of ibm in the early 70s said why would anyone want a personal computer you know and so that's sort of where we are and hearing from investors they're like what use cases do you after and i'm like well it's like a computer what use cases do you use a computer for like everybody's usage pattern all of them Yeah, all of them. Like everyone uses it differently. Like it's a tool for being more productive. It's like the power tool of the 21st century. This is not a product hardware category. This is the category that makes all the other categories, you know. And so it's much more of a fundamental thing. That's what draws me to this is this is the thing I can work on for like 20 more years or more, you know. Absolutely. And this is something that you can build an NVIDIA or Apple-sized company. And so I think this is what sometimes gets lost on people. when they're like, tell me the use case.

It's like, well, the world's changing. Yeah, everything's a use case. Yeah. So your agents can talk to each other, but they still can't think together. That's the gap Outshift by Cisco is closing. Outshift is Cisco's incubation engine, building the Internet of Cognition, shared intent, shared memory, and guardrails that allow agents to work together from anywhere. Open source protocols and code, a system for intelligence that scales. Read the paper and experience the demo at outshift.com. That's O-U-T-S-H-I-F-T dot com. Check out the links in the description. Now back to our video. That's so amazing, and I love that kind of, like, forward thinking. And that's a thing that's difficult for people. Like, it's often a struggle for people to think, like, we do this thing this way today. What's that going to look like in five years, 10 years, 50 years? And, like, you know, in the world of the old IBM CEO, like, he's living in a world where a data center held a one megabyte hard drive, you know.

And it absolutely seems insane at that scale. But, like, I assume that, you know, that's 50 years ago, that 50 years from now, the amount of compute we're watching go up in gigawatt data centers is what will operate in some little device we carry with us. which is nuts. That's an interesting thought. That's obviously what happened with CPUs and memory and everything. That would be wild if we could. Yeah, I hadn't thought of that. We went from literal factories holding less memory than the webcam I'm talking to you through. Yeah. No, it's so true, but the way we're scaling GPUs right now is you just make them more power at a time. I mean, it takes more power, and obviously they are getting more efficient, too, but it's also like our cooling systems are getting better. Like water's cooled down. So it's a lot of like just bigger at the moment. But that is fascinating that most likely in decades it will be much smaller too. Yeah. Yeah, and I mean I assume, you know, if you look at the like the technology curves on that stuff,

like it doesn't look like it's slowing down. Yeah, that's true, yeah. You know, even thinking as far back as like, you know, an OG iPhone or something, there's such a crazy difference. This is super interesting. I'm really enjoying this, and what you're doing is fascinating, man. Thank you. I'm looking through my list here to see what else there was. I wanted to make sure I didn't forget. So a robot, getting something right in a video is different from, like, doing something for an entire eight-hour shift of work. what measures tell you when this is production ready or even how often someone might need to intervene? Like it feels like maybe there's a threshold you cross or is that something that like you teach it and it just it's ready?

Yeah, it's a good question. The answer is so simple and it's really – it does depend on each use case by use case, but it just depends what the customer is willing to – like what they're okay with. Like it may be okay for someone to have to walk over once an hour. That may be a tremendous ROI because like instead of someone standing there all the time, they're like walking between 10 or 20 stations, you know? So like failing once an hour could be totally okay for some things. But if you're going with something where if you fail, it like ruins a part that costs $2,000. Making GPUs. Yeah. Then like maybe that one, that isn't okay. So it really depends on the use case. And so it just comes down to what the customer needs. And, you know, there's a lot of things where we have to have many nines, and there's other things where it can be surprisingly loose. And so maybe an unsatisfying answer, but that's the obvious answer, too, when you think about it. It's the same way you think about, like, should we use AI to automate a task. It's like you're looking at a difference in, like, some tasks, a little variance, doesn't matter.

Like, if it's meeting notes, it doesn't matter if it's a dash off. But if it's, you know, filing my taxes, precision matters. And I think it's interesting that that same paradigm carries over to robotics and learning that way. It's really cool. Are there any jobs that customers ask for that, like, you still have to turn down? Yeah, there are. Yeah, we definitely can't do everything. you know if there's a lot of latent state like things happening behind the scenes that's hard soft goods of course are still challenging in many ways I think if there's super long horizon like actually no one solved assembling an iPhone we've talked about that but that's like still is six hours of human labor doing that many different steps and obviously these humans are extremely good at it for six hours of, like, very good humans doing it.

And we'd all love an iPhone to be cheaper. Yeah, yeah, yeah. I loved when the Chinese tariffs came on and it was all the Wall Street guys in the vest. You see that meme? It's like, the boys get ready to assemble iPhones. I remember that. That was good. Oh, my gosh. You emphasized kind of learning from real deployments. What data is, like, the hardest to collect? And how much of what one customer teaches a robot can maybe help another robot in the future? Is there an amount of, like, training data sharing? Yeah, we're training base models that have a lot of different things in them, and I think that's also an advantage we have is we know what's useful to people so we can get that pre-training mix right. But ultimately, right now, no one's really creating super general models. And, like, even LLMs, to some extent, people say, like, they interpret within the training set.

Like, if you left out the biology training data, they wouldn't be able to solve anything, you know, in the biology domain. So it's kind of like it's all just kind of interpolating. And even humans, right? And so, like, the answer is if it's a similar task, even if it's a very different industry, like, it can be very helpful if you have semantically similar tasks in the training set. Interestingly, language is a good measure for, like, the semantic distance of words is very similar to the semantic distance of physical tasks. That was a very interesting thing I've come to know. Wow. But, yeah, so if it's semantically similar, then it is, you know, it can be quite useful. But if it's like totally different, then we're not – you don't see much of the left. That's fair. That makes sense. What – okay. Suppose – here's the shifted direction. Suppose I'm a small manufacturer and I want to buy my first robot. Yeah. What is like a deployment cost? What does it feel like?

I mean, what am I doing? Am I going to a website? Am I contacting someone and figuring out what my needs even are? Because I probably don't know. It makes me wonder, like, what makes it worthwhile and allows me to understand what I need to do? Yeah, so generally people will come. They'll fill out a form on our website. We'll talk to them. We have a team, apps engineers, that will look at a video of your task. We'll tell you, okay, we can automate that or we can't, or here's what we recommend. So we'll suggest, like, hey, we've got a couple different sizes of arms. We'll say, like, we recommend, like, one of this arm or two of these arms on this base, or we've got mobile bases coming out. So, you know. Nice. We'll suggest. And then there's the grippers or end effectors. There's hundreds of those that we support. So we kind of, like, we'll suggest a configuration. And then from that point, we're generally mailing a box. Some customers, they want more of a, you know, more hand-holding. So then we have integration partners we work with who can come in and, like, install it for you.

and do all that kind of stuff. We also do direct integration for bigger companies if they really need that. But our goal is make you independent so the next robot your team is doing on your own. Amazing. Amazing. I can see where to a small business owner there might be some trepidation around it's just going to arrive in a box and I'm just going to turn it on and we're going to figure it out, right? Yeah. I think where the initial reaction might be, whoa, depending on who you're dealing with. Yeah. I mean, honestly, this is where reputation comes in. I mean, it's so important in business in general to build a good reputation. But our belief is if you not going to be successful we not going to sell you the robot We got a huge backlog of customers We don So we like the reason we take a quick look is because we assessing the customer too Like is this person going to be capable of this Yeah.

And so it's because, you know, it's still like some people can't even set up their iPhone is the reality. They've got to go to the Apple store. So we're assessing that too. And ultimately, everyone is successful at the end of the day. Like, you can't build a big business not having your customer be successful. So we care deeply about that and care about doing it in a scalable way, but we're making sure that people are. And then folks come to know that. They know that if we're selling you something, you're going to be successful or not. And we offer, you know, phone support, and we also now have cloud code built in. So if you have any question on, you know, I say cloud code, but it's like a Gentic help, you know. So you'd be like, if you have any question in the world, it's got our manual. It has thousands of examples of how to program a robot in all these different environments, and it can pretty much answer whatever you want at any time of day within seconds. And you'll be like, hey, I need to actually, like, change it in this way, and it'll just edit your program and answer, you know. So we've got features nobody else has that really make sure you're successful.

That's really – that's an awesome thing. How do you decide whether, say, your job needs a fixed arm, a mobile two-arm, or a walking robot, and where do legs earn their keep? Like, where does that matter? That is still human neural networks. So, you know, looking at the video and talking to them and, you know, signing. I think legs could be great in this is a whole other debate, but I personally think factory, field, and warehouse, legs are not the right form factor. I agree. And the reason is because you can – it's a lot of unique parts for the legs. Yeah. Even if you take Elon's numbers and you deconstruct them, like it's still a lot of money per set of legs. And so if you're selling 1,000 to somebody and they – the wheeled base is both fewer parts. It's way simpler. It can – it's more efficient per, you know – Probably more durable. Yeah, more durable. It's much safer because it's stable.

You can make it, you know, very stable. Yeah. So you're getting ultimately a better ROI. You're getting better performance for less cost. So now that's when you have clear aisles, which you happen to have in a warehouse or a factory or agriculture. Now, if we're talking about cleaning your house, you might not clear aisles and you got stairs. So I do think legs will start to come online there. I think like a military application in a forest, like obviously legs were evolved. Yeah, because that's the best form factor for going through. Yeah. Like that kind of cluttered environment. But I think in a home, I wrote an article with Paki McCormick, and we coined this like home alone eval, which is the humanoid is going up the stairs, and you've got an eight-year-old or a team of them, and they're trying to trip it and flip it and whatever. And it has to be resilient to all that and not hurt them, not hurt the kid. And so if you've got fishing line across the stairs and the humanoid's going up, and that is a really hard thing to train in the simulator because you have to simulate everything that someone could come up with. to get this thing to fall on you down the stairs or fall on you.

Yeah. You know what I mean? And even an adult. Oh, my gosh. You get an eight-year-old jump on your shoulders as an adult going up the stairs, and you, like, almost flip, and you're, like, grabbing the walls, you know? And so this isn't easy even for human intelligence. What a fun thing to sit and think about, though, if you've spent all day trying to figure out. That would be fun, right? That sounds like a good bachelor party. Yeah, that's a great job. Yeah, that'd be a good bachelor party. This is going to try to go up the stairs, and you guys have 10 tries to stop it. You go to Home Depot and get whatever you want. That's right. That's right. Oh, my gosh. That's awesome. You, I believe, announced a $200 million round in June. Is that correct? Yeah, that's right. That was our last round in Seriously. Excellent. What does that help you solve and get you closer to, if you don't mind me asking? Yeah, it's really about a couple things. One is, I mean, I think it's an endorsement of what we're doing, that it's working, that we're growing revenue so quickly. And with startups, you're convincing the world, like, hey, we should put a little more into this one.

Let's get a couple more people working on this. And these customers, the customers vote and the investors vote and the press votes and podcasts. So I think it's just a sign. We've got some votes right now. And it's really worked. The customers are loving the product. and the employees are really enjoying working at the company. And I think what it lets us do is just more of what we're doing, building amazing products, automating tasks for customers and really caring that they're successful. And, you know, it's like that, we're putting that care into automation that you feel from like a really, really great brand or company. And that's like the thing that I think the customers get from us is we are going to make you successful and you're going to, you know, you're going to be able to, it's hard to hire. you're going to be able to shift the role you couldn't hire for. You can go, you know, instead of having to have somebody work in that refrigerator and you're spending all this time trying to hire for that and they turn over so fast. So it'll allow us to do more of all that. It'll be go-to-market, sales, more of that, producing more robots and building more products and some really great products on the horizon.

Hey there, humans. Quick heads up. We're throwing a little neuron get-together in San Francisco on November 18th, starting at 4.30 p.m., and you're invited. Come hang out with us for drinks, bites, and a live recording of the Neuron podcast. Plus, special guests and a room full of people who spend probably way too much time thinking about AI. Huge thanks to Slack, Alumni Ventures, and Sonar for helping us bring the whole thing to life. Now, it's completely free, but space is limited. So scan the QR code on screen to RSVP and save your spot. Hope to see you there. That's awesome. How do you, in doing that and trying to get them further around, how do you balance, like, domestic capability with also being able to keep automation affordable enough that small manufacturers are able to, you know, wade their way into the robotics world as well? Because I assume we're probably just starting to see more of that. Yeah, that's a great question. Well, our view is we're not going to make it more expensive to make it made in America,

but we are making it made in America. So those are the set of constraints. This is already constraints. I like it. And so what do you have to do then? You have to automate a lot of things. We'll still have a massive team in the factory, but we're going to automate a lot of the things that other robots have not automated. And ultimately, if you get down to it, it should be a small amount of labor. like we should be able to make them even in the New York area price competitive with anywhere in the world and that's because ultimately like if we have you know cost of if you look at the raw like the raw materials are a couple hundred dollars to go into the robots if you can automate a lot of the process the index is super high right now like the ratio of final cost to the raw materials so it should be quite possible to compete and so that's that's what we're doing and we're being really smart about how do we cast and how do we mill and how do we, and a lot of times our robot is instrumental in doing those things. I was about to say, you have robots. That's a real bonus. Yeah, and that's also a benefit is this becomes a showroom, and we really want this to be

the showroom in the country for if you want to go and learn how to best-in-class manufacture, we've got all those processes there. We're using it to build more of them. Yeah. No, exactly, yeah. So I would love to have you come visit. I would love to. That sounds so, so incredibly cool. We'll play D&D when you're – There we go. We'll make that happen. We'll make that happen. So one last question before I let you go. What's your – 12 months from now, even a couple of years from now, what's your current, like, pie in the sky? What's the milestone you're zeroed in on for physical AI and robotics? You're saying a couple of years from now? Yeah, within the next couple of years. Yeah. Like shorter term milestone, I guess, like that. I think that we're going from a world where we almost deploy no robots, like in manufacturing. Yeah. 30-something thousand a year in the U.S., to millions per year.

Wow. Tens of millions. And so I think what you see in the next couple of years is serious progress on that trajectory. I think it's still quite early where people, even investors, like there's some, obviously robotics has never been hotter, but there's still like, you know, a lot of people invested in robots. There haven't been a ton of IPOs, et cetera. I think that you're going to see that completely shift in the next few years. Yeah. And the chatty moment will become like undeniable and very obvious. And so I think that's what's coming. I think so too. I think you hit the nail on the head, man. It's interesting. It's been interesting just to watch the prices go down, even in just basic humanoids, to go from hundreds of thousands of dollars in the last decade or so and work their way down to both more useful and more attainable. And now, like, $20 will get you something that's fun. $20,000, not $20.

Let me speak that. And there's even all of this, like, entry-level stuff that I think is really cool for, like, bringing up the next generation of people who will be building these things. Like, I look at, like, the micro duck from Hugging Face. Like, it's a little thing. And Ricci, Ricci Mini is another cool one. Like, it just screams, kids should want one of these and learn to build these while they're young. Yeah, 100%, yeah. It could be more fun than that. Right? What an exciting time. I was like, I remember, you know, I always wanted stuff like that. For me, I guess my initial robot, and I've got him right up here somewhere, was Johnny Five from Short Circuit. It was like my childhood robot. And I always thought he was one of the coolest. And the Rocky robot was another cool one. I think kids finding that interest will really, they're just interesting to be around. They're interesting to watch. If you go to a trade show, it doesn't matter if it's humanoid.

It doesn't matter if it's industrial. There is a crowd of people at every robotics booth, it seems like, and they just want to watch it happen. There's something about them that draws people in. I think so, too, yeah. I think there's a long train of people who will be working on these in the future. I think, yeah, you see it at a trade show or you see it, it's just like me, and then you realize it's on the menu, And then that can be your whole career in math and science or robotics. Exactly. Well, Evan, thank you so much for joining us. Thank you for having me. This was so much fun. Yeah, this was great. We'd love to do it again. We will. We'll absolutely do that. And at some point I'm going to make it up and check out your place. Working with us go to learn more about standard bots, the work you're doing. Maybe look into getting some robots in their own factory. Yeah. Happy to help people. Excellent. Excellent. Well, if you like today's video, please take just a moment to like and subscribe. Don't forget to check out the Neuron's other projects like our daily newsletter read by 700,000 or so people just like you,

as well as our new sister newsletter, Robotics Insider, that we've just launched in the last couple of months. We'd love for you to be a reader. Thanks for joining us. We hope you'll be back next time. And on that note, that's all we have for today. So, farewell for now, humans. Thank you.