00:02:59 · Maya, Priya, Anika
[00:02:59] S0: Yeah. [00:02:59] S1: In the last five months, which is insane. We'll talk about that. [00:03:02] S0: Cool. So should we get into our first story? [00:03:05] S1: Yes. We can start with robotics data labeling and ecocentric data. It's been coming up a lot because there's a number of companies that have been top of mind specifically on x, I think. So first is Shift. They are basically giving away free house cleanings if you basically strap something to your head. And anyone... It's for anyone in New York, so that's like a really expensive market to be doing this. So I feel like a lot of people were talking about it on x. The other one was Human Archive, and so they're doing the same thing, but in India, which is the more common approach that we've seen is them basically tapping into the gig economy in India and doing it that way. But, yeah, in the last few months, I think we all knew physical intelligence was, the hot new topic, but we're seeing all of the data providers pop up now. [00:03:46] S0: Yeah. And this makes sense. I mean, DoorDash had launched their like tasks app, which is like also for egocentric data. And there's just a general need for all of this data in a POV point of view because these are not things that are just like on the Internet easily to like scrape and use. But yeah. [00:04:02] S2: Which is what they were able to do the last time we created LLMs is they just scraped like Reddit and yeah.
00:04:02 · Anika, Priya
[00:04:02] S2: Which is what they were able to do the last time we created LLMs is they just scraped like Reddit and yeah. [00:04:08] S1: The entire Internet. Literally, the entire Internet. Yeah. And so the different types of data that are needed for foundational models, world models, humanoids, all of those different things fall into, like, best to worst quality of data, I would say. So, like, at the top, you have actual real robotics data. So that's like from the sensors and from like an actual robot taking action on something. Because then you can see, did it fail? Did it succeed? Like, what do I need to change to make sure that it does succeed next time? That's why that's so helpful. Then after that, you have robotics human operated data. So you you might have, like, seen those videos of people in the different suits, like, moving around. And so that's, like, simulating what a robot would do. And because you actually have, like, the depth perception and you have, like, the three d world actually around you, that's why that's valuable.
00:04:08 · Priya, Anika
[00:04:08] S1: The entire Internet. Literally, the entire Internet. Yeah. And so the different types of data that are needed for foundational models, world models, humanoids, all of those different things fall into, like, best to worst quality of data, I would say. So, like, at the top, you have actual real robotics data. So that's like from the sensors and from like an actual robot taking action on something. Because then you can see, did it fail? Did it succeed? Like, what do I need to change to make sure that it does succeed next time? That's why that's so helpful. Then after that, you have robotics human operated data. So you you might have, like, seen those videos of people in the different suits, like, moving around. And so that's, like, simulating what a robot would do. And because you actually have, like, the depth perception and you have, like, the three d world actually around you, that's why that's valuable. [00:04:51] S2: And that was a whole controversy when one x launched and said they're gonna be the first consumer humanoid robot in people's houses where people were like, oh, it's all teleoperated. Do you want someone in India, like, peering into your house? The reason it's teleoperated, it's because anything the robot can't do, and right now the robot can't do most things on its own, ends up being teleoperated. But then that data that gets collected is, like, the second highest quality data. But it's not like immediate feedback loop. It's like you collect all of it and then you feed it back to the model and then Yeah.
00:04:51 · Anika, Priya
[00:04:51] S2: And that was a whole controversy when one x launched and said they're gonna be the first consumer humanoid robot in people's houses where people were like, oh, it's all teleoperated. Do you want someone in India, like, peering into your house? The reason it's teleoperated, it's because anything the robot can't do, and right now the robot can't do most things on its own, ends up being teleoperated. But then that data that gets collected is, like, the second highest quality data. But it's not like immediate feedback loop. It's like you collect all of it and then you feed it back to the model and then Yeah. [00:05:20] S1: Yeah. Which is like the same thing that happens with Waymo. It's like Waymo's get stuck all the time, and I call support all the time. There's this one specific turn right next to my office that I get stuck on every single time. Oh, really? It's happened three times now, but every time I call them, they count my ride and give me $10. So I'm like, five minutes of Yeah. Additional time, but I get a free ride. Yeah. But like that, someone comes in and actually tell her operates it. So it's like the same approach that's taken there.
00:05:20 · Priya, Maya, Anika
[00:05:20] S1: Yeah. Which is like the same thing that happens with Waymo. It's like Waymo's get stuck all the time, and I call support all the time. There's this one specific turn right next to my office that I get stuck on every single time. Oh, really? It's happened three times now, but every time I call them, they count my ride and give me $10. So I'm like, five minutes of Yeah. Additional time, but I get a free ride. Yeah. But like that, someone comes in and actually tell her operates it. So it's like the same approach that's taken there. [00:05:45] S0: Yeah. And then the the thing I was gonna say actually is in that robotic data, you get force torque, which is like how hard you're gripping the things or how how hard you're like pushing things. And you also get like tactile, like different tactile data, like pressure distribution across your different fingers and stuff. These are just things that you cannot get from a human. So there are, there is so much need for a human operated robot specific data. [00:06:08] S2: That makes sense. [00:06:10] S0: But but... Keep going.
00:06:10 · Maya, Priya
[00:06:10] S0: But but... Keep going. [00:06:11] S1: Yeah. And then after that, it's simulated data. So there's a lot... That's a whole category of its own startups that are working on being able to simulate data. NVIDIA Cosmos is like one of the big platforms that does that. And it's basically just synthetic data. The reason that people use it is because you can actually get it at scale. It's super cheap. Like the first two that we discussed are super expensive and difficult to get. And so at scale, it's just like the best thing that you can get. And then the last one is egocentric data. So that's what, like, all of these startups that have been in the news recently are focused on, and that's when you have a camera either strapped on your head or like metaclasses or something like that to record what's going on. But then like Maya was saying, you don't get the force... Was it forced work? Yeah. You don't get forced work. You don't have depth perception. You don't have all access to all of that. And so it's not ideal data. It's like the worst quality. But when you have millions of people in India with cameras on them, it's scalable.