00:03:52 · Sebastian Johnson, Eric Landau
[00:03:52] S0: Mhmm. [00:03:53] S1: So one thing that I don't think people realize is that AI has actually barely touched our lives. It's, like, very minimal impact. You might use it as a chat box, or maybe, you want to edit your script a bit or send it to your questions. So things that are, like, minor within your day to day routine. But in the limit, when this stuff is actually working, it will be touching every single aspect of your day to day life. You're gonna wake up to an AI system. An AI robot will be something that has cleaned your house over the night. You're gonna get, your vitamins delivered by a drone that comes in, at at the day of... You're gonna be driven to work in a self driving car. You're gonna get a a health scan from an autonomous system. All the infrastructure will be checked by autonomous systems. So AI will just be in every single aspect of our lives. And to make that work, to make these systems actually work and, be reliable in the real world, the key ingredient is data. And so we've gotten to a point where a lot of these companies are getting funding for demo and POC projects where you can work off of open source datasets or maybe just scraping the Internet, and so it's sufficient to get to the stage that we're at. But data is like the, it's the bit that kind of does the last mile in getting these, systems from, like, you know, 90% accuracy to 99.999% accuracy, which you need for real world reliability. So we're gonna get there, and it'll be there, but, we're we're, you know, we're still kind of getting to that stage.
00:05:13 · Sebastian Johnson, Eric Landau
[00:05:13] S0: And what does that data look like for the physical world? Like, how do you capture and store? What what is that data? [00:05:18] S1: Yeah. I mean, it's all the data that you take in as a person. Right? You're you're seeing me, like, you're hearing me, and you're getting sensory data from, the air around you, from the temperature. So it's sensor data. It's audio data. It's text data. It's visual data. It's multimodal data. And, one of the bets that we made a couple years ago was really investing in, multimodality. This is a time when, there was a lot of focus on just text only, use cases, and we thought that, ultimately, AI systems would kinda move towards more, multimodal native, functionality, which happens a lot within the physical, world data. And so that put us in a strong position to cap... Capture a lot of the, the tailwind in physical AI in particular. So just capturing all of the different types of data that, a person takes in on a day to day basis. [00:06:05] S0: And so you... I saw that you've 10x revenue for, the physical AI work that you do. Can you talk about maybe, like, a client or a customer or two that you're working with in the physical AI space and what that really means for them?
00:06:05 · Sebastian Johnson, Eric Landau
[00:06:05] S0: And so you... I saw that you've 10x revenue for, the physical AI work that you do. Can you talk about maybe, like, a client or a customer or two that you're working with in the physical AI space and what that really means for them? [00:06:16] S1: So maybe a few different examples. One is, woven by Toyota, so the self driving unit of Toyota, where, they have obviously quite a lot of data of, self driving cars in in the real world, and it's very difficult to manage and curate that data at scale. So they use our our platform to kind of manage and curate for those types of, self driving applications. Another example is, Pickle Robot, which is, a very exciting robotics company. They started in, logistics doing, loading and unloading of trucks, and they use us for, like, their, data engine. So, you know, where the model doesn't do well, kinda find the next bit of data, send it over to an automated, pathway to do a mixture of, like, kind of agentic annotation and human loop review, and then send it back to the model to to do retraining. So those are two kind of, like, physical examples in, the robotics and self driving world. One thing about a lot of the, physical AI companies is that they're extremely secretive, because, they're at the very forefront, and they don't wanna know... They they don't want other people to know what they're doing and what their stack is, so I I probably can't share too much about, some of the other customers that we're... But we're working with, like, a lot of the major, robotics companies in the space.
00:06:16 · Eric Landau, Sebastian Johnson
[00:06:16] S1: So maybe a few different examples. One is, woven by Toyota, so the self driving unit of Toyota, where, they have obviously quite a lot of data of, self driving cars in in the real world, and it's very difficult to manage and curate that data at scale. So they use our our platform to kind of manage and curate for those types of, self driving applications. Another example is, Pickle Robot, which is, a very exciting robotics company. They started in, logistics doing, loading and unloading of trucks, and they use us for, like, their, data engine. So, you know, where the model doesn't do well, kinda find the next bit of data, send it over to an automated, pathway to do a mixture of, like, kind of agentic annotation and human loop review, and then send it back to the model to to do retraining. So those are two kind of, like, physical examples in, the robotics and self driving world. One thing about a lot of the, physical AI companies is that they're extremely secretive, because, they're at the very forefront, and they don't wanna know... They they don't want other people to know what they're doing and what their stack is, so I I probably can't share too much about, some of the other customers that we're... But we're working with, like, a lot of the major, robotics companies in the space. [00:07:31] S0: Do you think physical AI companies in this space or maybe physical AI data companies like yourself, can you get to the size and the scale of some of the LLM companies that we've got?
00:07:31 · Sebastian Johnson, Eric Landau
[00:07:31] S0: Do you think physical AI companies in this space or maybe physical AI data companies like yourself, can you get to the size and the scale of some of the LLM companies that we've got? [00:07:40] S1: Yes. The short answer is yes, and the, the data demand for physical AI will be as large, if not larger, than, for LMs, where LMs, we have the luxury of being able to scrape the Internet as the kind of precursor for a lot of the data. So, you can just go in and, have a smart web scraper taken and, take all of human written words or publicly available written words and use that to seed the model, but there's no equivalent of that for the physical world. So taking your entire life of all the experiences that you have and consolidate that across, you know, thousands of people and then giving that to a model, that dataset just doesn't exist. So a lot of these companies will be extremely data hungry to, get... Just get the models to the baseline level, which is equivalent to to LLMs, but they're still at an earlier stage than, like, what a GPT probably, like, two is. So we're still at the very, very early days of how these physical models will realize in in the real world. [00:08:38] S0: And they're collecting their own data for their own use cases. Right? We're not seeing this sort of, like, centralization of physical data yet, are we? We're not seeing, like, the open AIs, the data scraping on a huge multinational level. We're seeing it with individual companies scraping data for their own use cases. Right?
00:08:38 · Sebastian Johnson, Eric Landau
[00:08:38] S0: And they're collecting their own data for their own use cases. Right? We're not seeing this sort of, like, centralization of physical data yet, are we? We're not seeing, like, the open AIs, the data scraping on a huge multinational level. We're seeing it with individual companies scraping data for their own use cases. Right? [00:08:53] S1: Yeah. They're, they're using a number of strategies to to get the data that they need, but it's very Wild West and ad hoc now. The LM labs, they all had kind of a similar recipe of just getting the initial dataset that they needed to, bootstrap the model. Yeah. But a lot of the physical AI companies are, doing their own thing to get to the same place. And probably the ones that will be the most likely to win are the ones with the smartest data strategies because the model there, as of now at least, is is pretty commoditized on on that front. [00:09:24] S0: Interesting. And I wanna talk about this transition that you've gone from being, data annotation to this physical AI, much larger kind of infrastructure play. When we spoke that this was something you were betting on Yeah. Was this a bet that you made about the direction of AI, or was this being close to your customers being pulled in this direction?