00:27:33 · Bernt Øivind Børnich
[00:27:33] S1: And I 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 gonna be incredibly important because it's such a big problem, and I don't think we can solve it alone. We 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, like, as far 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. Alright? So, like, everything is gonna be, like, perfectly full. All of your clothes will be ironed. Like, even you... Even your sheets will be ironed, and they will be, like, perfect on your bed. And, like, all the glasses will be online in the cabinet. And, like, there will be no dust anywhere. And, like, can create a lot of work if you want to. Like, if if if there's, like, no... 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 gonna do it anyway because, like, why not? So I'm not I'm not sure there will be that much idle time. We'll see. We'll see.
00:28:50 · Ty Morse, Bernt Øivind Børnich
[00:28:50] S0: You know, when you were talking about perfectly arranging the glasses, it was... I was thinking, like, most humans, like, my laundry is not perfectly folded. So if you emulated me... Like, it's it's it's folded, but it doesn't... Like, you see creases, you know? Yeah. 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? [00:29:17] S1: I don't think you do. I think you just learn how the world works. And and, like, we learned 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 and and both of them are equally important because how you... You learn how to get from one... Transition from one stage 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 understands how to do that. It's a pretty good path there. It's it's a pretty well explored problem, especially self driving is a good example of this. Like, most drivers are terrible drivers. [00:30:06] S0: For driving, it's a very it's a very thing. You you want the robot to do exactly, you know, perfect job every time.
00:30:06 · Ty Morse, Bernt Øivind Børnich
[00:30:06] S0: For driving, it's a very it's a very thing. You you want the robot to do exactly, you know, perfect job every time. [00:30:13] S1: Yeah. But, like, if you just clone human behavior in traffic, you're not gonna get that. No. Not at all. So you can't actually just do that. Right? You 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 that 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 through cloning human behavior. That's that's kinda like a very... That... That's how we used to do things, like, years ago, but, like, that... 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
00:30:13 · Bernt Øivind Børnich, Ty Morse
[00:30:13] S1: Yeah. But, like, if you just clone human behavior in traffic, you're not gonna get that. No. Not at all. So you can't actually just do that. Right? You 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 that 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 through cloning human behavior. That's that's kinda like a very... That... That's how we used to do things, like, years ago, but, like, that... 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 [00:31:14] S0: think forward 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. Mhmm. 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.
00:31:32 · Bernt Øivind Børnich
[00:31:32] S1: How do you, like, do that? I think it's gonna 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 evolves. There's, like, this not so well kept secret in AI that there are no secrets in AI. It's all just data and evals. Like, why are the best models the best models? It's because these companies have the best data. They were early, had huge user bases. They got enormous amount and 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, it's kinda like the self enforcing flywheel. There are other models where you can kinda, like, specialize. You can download an open source model. You could try to, like, do some fine tuning on it to, 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 gonna be different. We we see the same trends, the same scaling laws, the same patterns. It's just robotics has until now been, like, kinda like a toy problem. Everyone's been working on extremely small sets of robotics data where it's like, oh, I gathered four thousand hours of, like, these specific behaviors with my robot, and, like, this is a very big dataset. No. It's not. Like, four thousand hours is nothing. Even, like, few hundred thousand hours is nothing. And if you see the models that people train now of, like, equipping 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 two hundred thousand 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 of data. So you're not gonna go gather that data. That's why I deeply believe that you need to use these machines as kinda 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 NEO. And you can generally go in that direction. Right?
00:33:33 · Bernt Øivind Børnich
You can take a large, complicated, like, very capable model that's trained on this enormous diversity of data, and 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 all the way around. You can't take this model from, like, this smaller problem that only solve that data and distill it up to solve a general problem. That doesn't work. It's a one way thing. So you kinda have to have the most general system that has the most diverse data and, like, has experienced the most 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 kinda, like, specialized models, and that's what we're seeing today. Right? If you wanna fold 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 it doesn't have a general understanding of how the world works or, like, it's essentially intelligent. It won't it won't exhibit kinda, like, social behavior and, like, understanding how to interact with people. It'll just fold the laundry. And that's kinda 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, the specialized models will still be better. But I think 2027, you're gonna see a shift towards more like general intelligence.