Sources with Alex Heath
00:40:36 · Ellis Hamburger, Tony Zhao, Alex Heath
[00:40:36] S1: It just gave us the finger. Thanks so much, Tony. [00:40:40] S0: So it's... We have the hand actually here and the gloves. So the goal part here is that you actually don't have to use a whole robot to be generating data to train the robot. All you need is to wear these gloves and do the task, and we'll be able to, like, leverage that to train the robot. So what that allows us to do is we don't have to get data solely from our customers. Right now, we are actually having, like, more than 500 people in The US as contractors, and it's part of their work to be, like, just wearing the glove and do the chores, and we'll pay them for that. And that's how we train the robot without, like, needing to deploy it in the first place. [00:41:23] S2: How much are we paying for this? I mean, can you hook a hook a brother up? I mean, I'm an entrepreneur now. I need some side income. [00:41:29] S0: I think they're they're actually paid very well. In terms of wall clock time, every hour they spend, they get paid, like, $40 ish, which is very good. And the additional nice thing... I I think it's actually a very good job as in, like, you don't need to leave your home. You can just stay there, and you can do it anytime. You can do it for, like, ten minute in the morning, ten minute in the afternoon, and ten minute at night. Right? And we don't control for that. All we need is to get, like, this amount of well instructed, like, you know, clean data from your home, which brings us the diversity and the dexterity that's needed.
00:41:29 · Tony Zhao, Ellis Hamburger
[00:41:29] S0: I think they're they're actually paid very well. In terms of wall clock time, every hour they spend, they get paid, like, $40 ish, which is very good. And the additional nice thing... I I think it's actually a very good job as in, like, you don't need to leave your home. You can just stay there, and you can do it anytime. You can do it for, like, ten minute in the morning, ten minute in the afternoon, and ten minute at night. Right? And we don't control for that. All we need is to get, like, this amount of well instructed, like, you know, clean data from your home, which brings us the diversity and the dexterity that's needed. [00:42:06] S1: Is that like an ongoing project in terms of the training or do you feel like at some point it's gonna magically be done?
00:42:06 · Ellis Hamburger, Tony Zhao, Alex Heath
[00:42:06] S1: Is that like an ongoing project in terms of the training or do you feel like at some point it's gonna magically be done? [00:42:13] S0: That's a very good question. I think, I think we'll be... I think the gloves will be here to say we're going to be using it for a very long period of time just because if you think about places where AI works the best, It is when it has almost access to unlimited amount of data. Right? We have the whole Internet to train the large language models, and these days, people are still paying, like,, like, scale, like, crazy amount of money to label their data. I don't think robotics will be a a kind of exception to that. And it's probably even in the other direction of, like, in robotics, we do not have a whole Internet scale robot data to start with. So we have even more work that we need to do. So I think we need to, like... Yeah. Like, the data collectors wear the gloves. We're expanding them actually quite quite a bit this year so that we can Yeah. [00:43:09] S1: How many people are wearing gloves, doing dishes right at this very moment, do you think? [00:43:13] S2: And how do their wives feel about it? [00:43:16] S0: Sometimes both them and their wife do it together, which is [00:43:20] S1: That's romance. [00:43:22] S2: That's $80 an hour. [00:43:26] S0: Yep. So at this point, we have around, like, slightly above 500 people on The US collecting data for us, and we're going into the thousands this year.
00:43:26 · Tony Zhao, Ellis Hamburger, Alex Heath
[00:43:26] S0: Yep. So at this point, we have around, like, slightly above 500 people on The US collecting data for us, and we're going into the thousands this year. [00:43:36] S1: Yeah. And I know there is the design of the product, the brand, some stuff on the software side, I'm sure, and how you interact with it. But, like, what percent of your moat would you say is this data that you're collecting with the Glove? Because I'm not sure that anybody else is is doing that. Are they? [00:43:53] S0: Yeah. So I think our approach of building a glove that is co designed with a hand, I actually... I don't think I'm aware of any company that is doing something. I think there are similar approaches, but not exactly the same. [00:44:06] S2: Everyone in robotics talks about the hand as a problem. Right? Like, it's one of the top problems, but you're right. Like, to my knowledge, no one is doing this as a data source like you are. [00:44:16] S1: How can that be?
00:44:16 · Ellis Hamburger, Tony Zhao
[00:44:16] S1: How can that be? [00:44:18] S0: Yeah. And this is actually kind of reflective of a a broader thing of, like, how we think about systems in robotics. So people tend to think about robotics as, like, singular things like the hands, now the data, and maybe the the legs and, something like that. But it's actually a lot more interconnected, than it look like. So for example, if, like, it could be true that, let's say, the bottleneck is the hand, and you can build a very, very dexterous hand. But the problem is, like, how can you get a lot of data for it? Like, what... If without the data, the whole thing doesn't work no matter how good the hand is. Right? So what we do is [00:44:55] S1: You can make a really nice fake video if you want. I feel like many have been accused of that in the last year.
00:44:55 · Ellis Hamburger, Tony Zhao
[00:44:55] S1: You can make a really nice fake video if you want. I feel like many have been accused of that in the last year. [00:45:03] S0: Yeah. I will not comment on that. Our videos But I I think it's actually more about, like, instead of building the most dexterous hand and call it a day and that's the optimization they want to do, we're like, hey. Can we build a hand that is so cape... Is capable, but we can also get a lot of data to unblock the AI part. So when we design the hand, it's actually not about like, hey, let's get the most amount of fingers out of it. Let's get five fingers. Let's get like 22, 27, whatever degree of freedoms and pack all the motors in. But can we design it like together with the glove? And the glove is something that people can wear and efficiently collect data. So I think when it comes to like for us in terms of the moat as a company, I actually think it's less about whatever data that we have because we're going to have so much more than that. But always approaching these problems with a global knowledge of, like, AI, of software, of hardware, and, like, the whole thing together, like controls. And that kind of always steers us to making decisions that is slightly different from the rest of the the field. And I think that will be the kind of the core competency that will continue to to own and that will continue to help us make bets that are somewhat unpopular in the beginning, but are post op obvious.