Speaker unverified: belief

Transcript context

…or using to something that's essentially expected. Could you just have cars going around in the parking lot? Well, not your Gen 1, but all the Gen 2 vehicles. They're gathering training data for this. So the obvious ones are if you're in self-driving mode, if you're on the highway, for example, in our vehicle, and it disengages, that's a piece of information, we train off that. More interesting is when you're driving, you as the human, the model's running in parallel and when you do something different than the model predicts, that's interesting. It's like if you change lanes and the model predicted you would not change lanes. Or vice versa, if the model predicts we should change lanes now and you don't change lanes. On a singular basis it’s sort of interesting, but you multiply that across thousands and thousands of vehicles, it becomes a really powerful method for teaching a digital driver how to drive. Yes. I feel like the big debate that plays out in the public realm on this stuff is basically the Waymo versus Tesla approaches of more sensors on Waymo's versus fewer sensors on Tesla. And Elon's being quite dismissive of LiDAR. Where do you stand on this sensor debate? I think first is I'd say there's alignment amongst Rivian, Tesla, Waymo around the use of AI to train. And so, the view that you're building a large scale foundation model, multi-billion parameter model, you're building that with the data flywheel of the vehicles that are deployed. There's consistency there. I think, for some reason this has been a lightning rod issue in ways that I wouldn't have expected. If you look at sensor theory, and even if you have noisy sensors, it just mathematically proves that having more sensors is a better approach. It’s always seemed nonsensical to me. Yeah, more sensors obviously better and we know how to— So that's the reason you have more than one camera. So even moving beyond more than one camera, we have many cameras around the vehicle, but then saying we want additional modalities that have a non-overlapping set of strengths and weaknesses with a camera. And it's, there's no arbitration of the camera versus a radar or camera versus a LiDAR, because it's just not how the models are built anymore. They have a worldview and then the addition of different modalities. It actually helps balance out some of the noise and the signals. Because every signal's going to have noise, every camera has noise, every LiDAR has noise, every radar. So you just want as many sensors as you can afford in the car, then combined— e and the signals. Because every signal's going to have noise, every camera has noise, every LiDAR has noise, every radar. So you just want as many sensors as you can afford in the car, then combined— And then it becomes as many as you can afford and as many as you can process, which ties to how much you can afford. And so, a few things have happened in the last five years that have changed things. One, as we move to a neural net based approach, that's what we talked about already. The second is that the cost of sensors is six, seven years ago a LiDAR was $20,000. 15 years ago LiDAR was $75,000. Today a LiDAR is like 200 bucks. So it's a very low cost sensor and it solves certain things like very bright light, very low light, extremely well. And similarly, a radar is a hundred bucks, depending on which one, you can buy the crappy radar for 25 bucks. You can buy a great imaging radar for a hundred bucks, 125 bucks, and it gives you rain, snow, fog in ways that cameras obviously don't perform very well. And if you've ever driven in really thick fog, you've probably thought to yourself, boy, wouldn't it be great if I had a radar in my forehead? That's what planes have. Which I think all the time. But you would navigate those situations with a much higher degree of safety. But then the last piece I'd call out is when you have multiple modalities, it allows you to train your perception stack better. For example, if you're driving in the fog, there are subtle little things your eyes will spot and you're hunting for, like you're looking for the little red dots of the car in front of you from their taillights, but you have to hunt for them. Imagine your brain had the assistance in training its neural net to immediately know where those objects are, vis-a-vis radar. So your brain can more quickly identify the visual characteristics and you build this much more robust neural net to be able to see in hard situations, bright lights. And you're like, let's say coming down the 101, you've got that sun right in your eyes. It's really helpful to have something that doesn't affect. So I think it's strangely become a debate topic. I think it's not really— Everyone's actually in agreement on the high level points. I think everybody's actually in a more aligned— Everyone agrees on the overall AI approach and then… Yeah, actually there isn't that much disagreement on the sensor stuff or maybe less than it would appear. You're describing some of the progressing components. So you're describing how LiDAR and radar have really changed in becoming a lot cheaper. What else is happening in the land of components that you work with? Are tires getting better, batteries, tires, windshields, anything like that? I think on the electric side, we're seeing the motors become lower cost. Okay. So a lot of components that are pretty important for electric vehicles, is there like a meaningful cost curve of improvement happening on them? c side, we're seeing the motors become lower cost. Okay. So a lot of components that are pretty important for electric vehicles, is there like a meaningful cost curve of improvement happening on them? Yeah. And then I think if you take the complexities of designing real-time operating systems and a compute platform across the vehicle, these are things that car companies don't historically do very well. We've got, as an industry, there's this massive supplier base of tier ones that provide functions and computers alongside of them. So you buy a seat that historically has come with a little computer, you buy a car with power windows, it has a power window computer. And so with the exception of two companies, Rivian and Tesla, every car on the road has, depending on the amount of content, anywhere from 50 to maybe 150 little computers that run the car. This is why you, in a normal car, could not adjust the side mirrors from the center console, regardless of whether— It's two or three different companies. It's different systems— Companies and different software. It's islands of software written by little suppliers, on little ECUs that go back to suppliers. And so that actually underpins, we did a 5.8 billion deal with Volkswagen where we're providing them this—the second largest car company in the world— we're providing them a zonal architecture. So a platform that allows us to run the whole car with a very small number of computers. Depends on the size of the vehicle. I think people don't realize how much with traditional car manufacturing, the brands are in the car assembly business. And you're going to “car Costco” and you're filling your cart with, “We need a backup proximity sensor and we need a rear view mirror adjuster. We're spending a lot of time on our side mirror adjustment system and things like that,” but they're assembling all these components from sub-manufacturers that are not integrated. I think your point is, in a traditional car, the backup beeping sensor, oh, it can only beep. That's the only thing it does. And so you can't integrate it with the rest of the car and the systems because of that approach. Whereas, you think the integrated approach is just definitely going to win.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.