· Peter Ludwig, S0
[00:00:00] S2: Also in that process, you're doing model training and evaluation, you're doing data collection and post-processing of that data. These are very complex systems, but that exact analogy, you can apply that really to any kind of machine or any kind of robot, and those same things apply. [00:00:00] S0: Your team, while we were prepping, actually described this very beautiful loop where you can take information and data that you're getting from a tractor or an underground mine, and then that's relevant to, let's say, a drone or autonomous vehicle that's also relevant to an autonomous sea vehicle, and they all feed into this broader platform and inform how the platform gets made. I would love for you to describe that loop and how that helps build a general-purpose, broader platform.
· Qasar Younis
[00:00:00] S1: An AI system really is always two big components. One is the actual platform that you develop the intelligence, and the other is the actual intelligence. So it's almost like the world model and then the actual intelligence that you'll deploy on the machine. On that second half, on the intelligence that you're deploying on the machine, the way to think about it is this data engine. It's a feedback loop. So as the car is in the real world, consumes data, as in it consumes the world around it and creates data packages, it also is not successfully navigating specific scenarios. So you can almost mark, hey, the car had difficulty in doing this. So then how do you help the brain on the car navigate that scenario that it wasn't able to navigate last time more successfully next time? Well, you can do it a couple of ways. You can expose it to lots of scenarios that humans maybe have already literally human-driven. So it imitates how humans handle that scenario. You can create a synthetic environment where you show it, this is how you would navigate this type. And there's a bunch of techniques, but the macro point is there's a data loop. It's just feedback. So the machine interacts with the scenario, it can and cannot navigate that. Now step back, don't just make that a car, make that any type of machine. It can be a drone, it can be a mining dirt mover, it can be a combine. And the same thing happens. The interesting thing we've learned in our development of intelligence and models is as we take scenarios from, let's say, a drone, or we run autonomous trucks right now, L4 trucks in Japan, we take data from those trucks, it actually makes the performance of those models in fairly different environments better. So what's really happening is the model is getting a sense of physics in the real world. This should elicit some corollaries in the chatbot universe. Chatbots used to be very, very specific. Transformers happen, and general chatbots now actually can perform really, really well. The same thing has happened in self-driving. We benefit a lot from that because we see this diversity of data in all these different use cases. And so it feeds that data loop and the data engine.
· Peter Ludwig, S0
[00:00:00] S2: It also gets to a bit more of the distinction between digital AI and physical AI. Because in digital AI, if you're thinking about these general-purpose models, those are oftentimes trained on the internet. Plus maybe some extra data that the model company has built, and that's usually text data. But in physical AI, almost all of the data is actually proprietary. It's like data that we ourselves are collecting through our own vehicles and partnerships that we have with our customers collecting that data, because you're ultimately building these models on data that's not available on the internet. [00:00:00] S0: While we're on the topic, Peter, maybe you could talk about the quantum of data that you're able to collect — it's almost unimaginable across all the different vehicles and machines and industries and applications. I'm thinking of this mega brain or giga brain in a way. Maybe just talk about the data that you're able to collect and then the actions that you're able to take on top of that data that maybe no one else is able to do.
· Peter Ludwig
[00:00:00] S2: We do have an enormous amount of data, and that is a fact. It's very meaningful because, first off, it's expensive to do, but it's also very difficult. The actual tech stack required to do reliable, high-quality data collection is surprisingly deep and complex, and there's not that many companies around the world that really have a very high-quality tech stack for doing data collection for physical AI. That's a pretty fundamental moat and long-term advantage that we have. There's also all of these other very deep things that are unlocked based on that data — this combination of imitation learning with reinforcement learning, which we're very deep in, which I think is really the critical unlock to scale physical AI. So a lot of the talk right now in autonomy is this topic of end-to-end models where you can take data that's been collected, you can train a model off of that data using something called imitation learning, and that allows then a machine to effectively mimic what was being done in that training data. That's great, and it's been proven that it's very effective, but oftentimes it doesn't actually get you to a fully productionizable solution. What we've now added and really innovated on in a big way and done a lot of research and actually published a lot on as well, is reinforcement learning. And so you take that base imitation learning model and you complement that with a really powerful simulation environment using very highly performant reinforcement learning. And that can actually smoothen out a lot of the problem cases that you'd get with pure imitation learning. And we see this as the technical path towards large-scale, widely deployed physical AI. And I think we have some pretty unique advantages across the spectrum right now.
· S0
[00:00:00] S0: Are there other rate limiters that we should discuss? Obviously, Dana is a huge unlock — Peter, you were mentioning how the operating system is a rate limiter, in terms of it's just very hard to design, develop, build, test, and analyze all these different systems. And Dana is now going to be able to go ahead and do that. Are there other rate limiters — whether it's anything from the chips to the sensors, to the actuators, to materials, power is a big issue now — or anything that's around the actual technology that you're building that worries you?
· Qasar Younis
[00:00:00] S1: The diffusion of this technology will be at very different rates and very different ways. Fable comes out, it can work on your phone and your laptop because those environments are quite standardized because of the browser and the operating system and a bunch of other things — app stores, payment methods, and stuff like that. So it's very easy to consume that intelligence as an end user. There's just impedances in order to diffuse this intelligence into physical machines — they could be manufacturers, it could be the operators of the farm and the mine. So there's a lot of other things that kind of get in the way, and maybe just as importantly, the actual dollars. When people talk about self-driving cars, I always like to use passenger vehicles because everyone can kind of understand it — it's maybe a little harder to grok ports. In terms of your personal vehicle, let's say you just bought a Honda Accord yesterday and then tomorrow self-driving is available for free — well, you still own that Honda Accord. Over half of Americans live on a fairly small savings account. When they buy a car, it's a big deal and it's a big purchase, and they're going to use that car for ten, maybe 15 years, regardless of what the other product that's available in the market is, just because of the nature of economics and how much money that they have. And so the diffusion of this intelligence into machines has a lot of different complexities that you won't see on a desktop or on a phone. But I think those are also moats. Once you figure out how to make a mine autonomous or a farm autonomous, we as a technology provider in that ecosystem, I think, are really advantaged because we're really in there — again, just the same way silicon is so sticky. Once you're a chipmaker and your chips are in a bunch of machines, that's a real deep moat. And we are both at disadvantages and advantages of those realities.