Digitized | Spotlight
00:03:19 · Winnie, Michael Cho
[00:03:19] S1: almost like how it works though, you have to be quite naive to kind of build something like this, you know? [00:03:24] S2: Yeah. I mean, was a hunch. I mean, I I was keeping track of what's happening in the research world. I mean, even back then I used to read a lot of research paper, especially in computer vision and all that. So I kind of knew things were heading that way, but it's still, I mean, me and my brother, we are not both this training, So we just... But we were just very curious and we all like, we like to hack stuff. Mean, it's a Wiki that way. So it started like that. [00:03:56] S1: I love that. In the early days, in the early days, where were the robots that you guys were building over on the Saturdays and nights and weekends? What were, like, the first iterations of what you and Sam were doing?
00:03:56 · Winnie, Michael Cho
[00:03:56] S1: I love that. In the early days, in the early days, where were the robots that you guys were building over on the Saturdays and nights and weekends? What were, like, the first iterations of what you and Sam were doing? [00:04:08] S2: Oh, you mean where... Did you say where were the robots? Oh, what were the... I think it's all somewhat robots, but they they have, like wires dangling out all the time, they barely worked as you can imagine, they are overweight, very you know a lot more expensive than what we have now. So I think I would say we've come a long way, yeah but it was all just learning by doing kind of thing. I mean on one hand I think it's bad in that we, to be honest, we didn't know what we're doing at the beginning but on other hand it was also very good in that me and Sam, we had no preconceived notion of what should or shouldn't work. It was all just a blank piece of paper, we just assume we're idiots, we don't know what we're doing. But our philosophy was always just to test it in the real world, possibly because we don't have a proper lab anyway, so we just put the robots out there and see how they break And then... But we always made a conscious... Well, from day one, we made a conscious effort to gather the data and that has worked wonderful for us down the road when we started having more partnership or that. Yeah, it started that way. It's just a weekend project. Yeah.
00:04:08 · Michael Cho, Winnie
[00:04:08] S2: Oh, you mean where... Did you say where were the robots? Oh, what were the... I think it's all somewhat robots, but they they have, like wires dangling out all the time, they barely worked as you can imagine, they are overweight, very you know a lot more expensive than what we have now. So I think I would say we've come a long way, yeah but it was all just learning by doing kind of thing. I mean on one hand I think it's bad in that we, to be honest, we didn't know what we're doing at the beginning but on other hand it was also very good in that me and Sam, we had no preconceived notion of what should or shouldn't work. It was all just a blank piece of paper, we just assume we're idiots, we don't know what we're doing. But our philosophy was always just to test it in the real world, possibly because we don't have a proper lab anyway, so we just put the robots out there and see how they break And then... But we always made a conscious... Well, from day one, we made a conscious effort to gather the data and that has worked wonderful for us down the road when we started having more partnership or that. Yeah, it started that way. It's just a weekend project. Yeah. [00:05:17] S1: Cool. I love that story. I guess, like, in getting started, what were some key challenges or bottlenecks that you faced in the early days to kind of getting some of these general purpose physical AI and how does, like, BitRobot help solve them?
00:05:32 · Michael Cho
[00:05:32] S2: Yeah. So like I say, I think a big thesis of mine even like, I would say nearly four years ago when we decided to really do it full time was data is gonna... Like real world data could be a big big part of of solving Embody AI. And so you asked about, like, how we got into crypto, like, how has crypto involved? And so to be honest, when we... Me and Sam first started, I I thought, like, most skeptical crypto is scam. Like, I was very skeptical about crypto to be very honest. But then I discovered Helium. Like, this is... I remember... Yeah. This was, two months after starting the company. At the time, was only me and Sam. Right? And so yeah, and then Helium just totally blew my mind. I thought it was crazy that you can crowdsource both the hardware cost as well as the actual deployment across the world. There they were selling you know, hundreds of thousands of units, I think, by by that point, you know, at at about $500. And we were also trying to get our robots down to $500. And I wanted that that that hidden map that hidden had, and and so I thought I I need to learn about that. And I would say fast forward to now, I definitely 100% believe in VPN if it's done well. I think some things you can scale a lot faster using, you know, this kind of permissionless incentivization scheme. And I think especially for robotics in... Especially in in the research world, they call this like in the wild data, which is, you know, the the data is collected literally where the robots are supposed to be, right, instead of being in labs and warehouses and and things like that or a a controlled environment. So... And this is very critical because unlike the older days when you do, like, what they call classical robotics, meaning the robots are... These are robots that you'll probably find in factories, you know, like, they are doing the same thing day in, day out, like, you know, thousands of time every, like, every month. So you can use very deterministic kind of, like, coding to get that... Get those work, and those work very, very well. But those don't generalize. Right? But if you use the new paradigm, you know, the most extreme is you train thing what they call end to end. Ultimately, these AI models is just a reflection of the data that you collect from.
00:07:54 · Michael Cho, Winnie
So it makes a big difference if the AI models are trained using data that's collected from various environments. And so in our case, the SAWO robot, we now have robots in more than 50 cities because people are starting to buy them. These are cheap, like like toys. And it turns out that diversity in environments, it's... It it makes a huge difference on building a model that it's already quite generalizable even though you don't have that huge amount of quantitative data, but they are already quite generalizable because of where the data is collected from. And I think crypto is the best way to diversify and and get people around the world to contribute. Yeah. [00:08:36] S1: Yeah. No. I think crypto and kind of like a lot of deepened products have proved one of the early use cases around dramatically reducing upfront capital costs, incentivizing the right players. I'm curious to kinda understand how you guys think about kind of that incentive design. Like, how do people earn? What are the data that you guys are collecting? Walk me through what uniquely suits kind of robotics r and d and how... What the data looks like in comparison to some other projects that maybe listeners are more familiar with, like Helium that you mentioned?
00:09:08 · Amanda
[00:09:08] S0: With BitRobot, we're building a platform for distributed open robotics research and development. We're doing this through what we call a network of subnets. You can think of a subnet as a market that's focused on a particular type of output that's pushing embodied AI forward. So creating a specific type of dataset, facilitating, evaluation of models in the real world. And then the types of data that we're collecting, and across that include teleoperation, video, and simulation. So these are the three types of data that are needed for robotics. And then the benefit of having a network of subnets is there's many different approaches right now to how to move robotics forward. And so we can basically pursue all of them in parallel. So like pursuing teleoperation, but also egocentric video. And then another trend is cross embodiment or co training. And so this is like, okay, let's make sure that the model works on different types of embodiments, different types of tasks. And so we're able to kind of, through our different subnets, showcase, different types of data and then pull that together in something that can be a more general purpose AI model. So yeah, would say those are some of the strengths that we have. I don't know if, Michael, you want to add?