00:18:38 · Priya, Anika
[00:18:38] S1: Yeah. Okay. Like, different maid services. [00:18:41] S2: Yeah. Like, basically, everyone in India that's middle class or above has a maid. And they were like, okay, let's make like the Uber for maids in India. And so that's what it is. And then they realized they're sitting on a gold mine of possible data where like they can just... Because it's a marketplace, like, can just have the right to like strap cameras onto all of these people. And it's probably like a really low margin business already. And so I think, like, that data... Like, the data collection gives a little bit of a buffer. And I could totally see it being like a more core situation where you start by being about recruiting and then just become a data labeling company. And then there's like pure play data labeling companies like Human Archive is a big y c one. They're also doing a lot of work in developing countries and they're doing multimodal stuff. So like video across like homes, factories, retail. I know we were talking about, like, on one end of it, China is, like, optimizing for this one factory. I think on the other end of it, most of the American startups are actually trying to get as diverse of settings as possible. A lot of these contracts actually explicitly say you can't set us footage from the same place multiple times. Like, this has to change. That has to change. So that makes sense. Like, our friend Mickey recently launched one too called North Star. There's just like a... Like, the long tail end is like really long. [00:19:53] S1: It's really long.
00:19:53 · Priya, Anika, Maya
[00:19:53] S1: It's really long. [00:19:53] S2: Because there's even like non VC backed ones that none of us even know exist that are just like using their... I mean, you really just need connections in the in the developing countries. [00:20:01] S0: Could just [00:20:01] S1: go to a robotics reading circle and then find some connections. Well, I [00:20:05] S0: was gonna say it's really like an execution based game here. For sure. Like, surge for example, and when we were just talking about, like, text labeled data, they were bootstrapped. [00:20:15] S1: Yeah. And the reason people do gig economies and developing economies like India, for example, is because the price arbitrage is just, like, insane. Like, when you look at how much you pay these people in India, it's a dollar per hour maybe in US dollars versus in The United States, it's an average of 30 to $50. If you're in the Bay Area, it's a lot more than that. And so wage arbitrage in this case is a lot of the business model. And I think that was something that was interesting about Shift that I was looking at because they were marketing New York City. And I was like, why would you do the most expensive place in The United States? Yeah. [00:20:47] S2: But I understand that that I saw it was like, pet PPE dropouts. [00:20:51] S1: And I was like, oh, this makes sense. [00:20:54] S2: That's no hate. I went jobs. I went to a worse school, and I had the same major. Disclaimer alert. So... Yeah. What? That's true.
00:20:54 · Anika, Priya, Maya
[00:20:54] S2: That's no hate. I went jobs. I went to a worse school, and I had the same major. Disclaimer alert. So... Yeah. What? That's true. [00:21:02] S1: That's so funny. Yeah. But when I was looking into it, apparently, they're actually an offshoot for, like, a German data labeling company that's been operating for a while. And they're operating in, like, 15 different countries. They have 10,000 different operators, and it's primarily in Turkey right now. So it seems like they're, like, The US American facing Mhmm. Arm of this. [00:21:20] S2: Yeah. But still, it seems like they're trying to, like, mass do The US because then they said they're launching in Seattle, SF. I think... Like, they said a bunch of cities. I don't understand the math on that. And the key reason I'll say I don't understand the math on that is it's not that there's no one else doing US based data work that's like a big player. Like, for example, Tesla is doing it for Optimus, but, like, that makes sense to me because it's like you're willing to burn paying USD prices because you own the full stack of this. Right? So, like, that can be a cost center. And, like, if Optimus works out, like, obviously, you'll make your money back and then some. But where I think it doesn't work for some... Someone like a Slash is no one has even come out and said, like, EO centric data works. [00:22:02] S0: Oh. But... No. Wait. But what do you mean it does... Like, that's how they trained the figure in physical intelligence models.
00:22:02 · Maya, Anika, Priya
[00:22:02] S0: Oh. But... No. Wait. But what do you mean it does... Like, that's how they trained the figure in physical intelligence models. [00:22:09] S2: But they had a lot of teleoperated data. And, like, do they say they're breakdown anywhere? Because, like, all the research I read, like, example, Stanford has a paper on it and there's, like, just like general consensus. Mhmm. My understanding of it is like people are thinking like, oh, probably there's like some marginal improvement if you have a ton of egocentric data, but it's like a bet you're making. [00:22:30] S1: Yeah. [00:22:30] S0: Oh, you're saying like, you think most of the figure model and physical intelligence model, like most of the data is actually teleoperated. I thought a lot of figures data is from YouTube videos of like the third party view. And then a good amount is ecocentric data. The other thing is, like,
00:22:30 · Maya, Priya, Anika
[00:22:30] S0: Oh, you're saying like, you think most of the figure model and physical intelligence model, like most of the data is actually teleoperated. I thought a lot of figures data is from YouTube videos of like the third party view. And then a good amount is ecocentric data. The other thing is, like, [00:22:49] S1: in this scenario, low signal data actually affects the model, like, way more negatively than it does for LLMs in my opinion. Because, like No. I didn't know that. No. I'm just thinking I'm just thinking about it. Because, like, I think about, like, our, like, cybersecurity models. And, like, the quality of data is, like, so so important Yeah. In cybersecurity that, like, you end up just, like, tossing a ton of data. Where it's like, we run all of these experiments on our ML models, and, like, at the end of the day, we we take, like, huge chunks of data. It's, like, billions of messages. And we're, okay. We, like, actually can't use any of this because it's, like, so low signal. Yeah. And I feel like in LLMs, like, obviously, they do it to, like, some extent. But I feel like for a robot. [00:23:26] S0: Like, LLMs, actually, they don't toss a lot [00:23:29] S2: of the rules. [00:23:29] S0: Like, I met a girl the other day who did her thesis. Like, she recently did her PhD thesis on this, which was like, if you just do a rejects match against like the word... Like, how do you say this? Like, can I say that? [00:23:41] S1: What's the word? What's the word? Like, porn. Oh, yeah. [00:23:44] S2: You can say... I was I was gonna say...
00:23:44 · Anika, Priya, Maya
[00:23:44] S2: You can say... I was I was gonna say... [00:23:46] S1: You can say that. [00:23:47] S0: No. Like, I can't say that. [00:23:49] S2: No. No. Wait. You know, I I had this problem this episode too earlier. [00:23:52] S0: Yeah. Yeah. No. Because I saw you. Yeah. Yeah. Okay. The the p o r n word. [00:23:58] S2: That's a word. [00:23:59] S1: She's gonna say you somehow make it worse. Okay. Let's just move on. Anyways, her thesis [00:24:04] S0: was like on... Like, if you literally just like use simple search filters, like rejects matching filters to filter out this data. Mhmm. You come up with a really... With like a lot better of a model. So imagine if you like, spent more time doing like LLM filtered data. [00:24:19] S2: I see. And you [00:24:20] S0: could still come with like, And then you pass this to the model. So that made [00:24:24] S2: me think, oh, wow. A lot of [00:24:25] S0: the data that's going into these LLMs are not actually [00:24:27] S1: crash data. [00:24:28] S0: Yeah. They're not actually filtered [00:24:29] S1: that much. Wait. Would you guys have this in your house? [00:24:31] S2: I was gonna ask. [00:24:32] S0: Yeah. For sure. [00:24:33] S1: Oh, for sure. I was for sure. You think I'm worried about my data? [00:24:38] S0: Yeah. Like, oh, is it, like... Because it's gonna see me, like, negative or something? [00:24:42] S1: No. Priya just, [00:24:43] S2: like, loves being like, and I would let this early stage y c company invade my home. [00:24:52] S0: You know, like, I I guess I just, like, don't care about my privacy. Is that bad?