“Somehow, it didn’t leak.”
When I caught up with Rivian founder and CEO RJ Scaringe after the company’s “AI & Autonomy Day” keynote on December 11 at its Palo Alto headquarters, he marveled that the company had managed to keep the event’s news under wraps until it was ready for its big reveal.
It did—and there was a lot to discuss. At the keynote, Rivian unveiled its Gen 3 platform, which will turn the maker of EV trucks, SUVs, and vans into an autonomy company, a focus he says will subsume “the whole business” of transportation.
Debuting late next year in a version of the upcoming R2 SUV, the Rivian Autonomy Computer platform is powered by a chip the company designed itself, the RAP1 (Rivian Autonomy Processor). The R2’s self-driving features will also draw on data from a lidar unit that sits inconspicuously at the top of the windshield—a far cry from the spinning lidar towers atop vehicles such as Waymos. (Controversially, Tesla’s cars don’t use lidar sensors.)
Rivian also showed off a new voice-controlled user interface called the Rivian Assistant that will be available as an update for its current vehicles as well as for the R3. A bet on the future of car interfaces shifting toward talking rather than tapping on screens, it features integration with Google Calendar—hinting at the kind of productivity-related features that might become more useful as cars take over more of the work of driving themselves.
I spoke with Scaringe about all these topics and more. Our conversation has been edited for length and clarity.
At least in broad strokes, how much of what you announced today was part of the original Rivian vision and road map?
Well, that was 20 years ago. But something like today is the result of many thousands of people working on it for the last few years. Development on the platforms we showed today started in 2022.
One of the threads connecting a lot of your news is you doing stuff yourself rather than being dependent on other parties. Was there a period where you weren’t sure which way to go?
Maybe the way to answer that is that when we launched in 2021, we had what we call our Gen 1 architecture. And it was a very different approach than what went into our Gen 2 and, of course, what’s going into our Gen 3.
We had a perception platform—some of which was our own, much of which was not our own—that fed into a planner, which was our own, and made a set of rules-based decisions around how to drive the vehicle. It was very basic driver assistance: low level two features. And by virtue of that being the architecture, we had a natural limit. We realized that as we approached the launch—that the world was going to shift away from these more deterministic and classical systems to a true AI-based system.
It’s sort of ironic. When we say “self-driving,” we might think historically it’s always been AI. But in the beginning, actually, there was no AI. Some of it was very sophisticated if-then statements with good machine vision.
What it’s shifted to now is true AI, and that happened in the early 2020s. As that was happening, we came to the view that we needed to completely shift our approach. And when we made that decision in early 2022, we approached it as a clean sheet.
With that clean-sheet approach, it was, “Let’s design our own perception platform. Let’s design our own compute platform with Gen 2, leveraging Nvidia as a supplier of the chips themselves, the inference platforms themselves, and go build a data flywheel that will allow us to build a neural net-based approach.” The vehicles ultimately launched in the middle of 2024, a little less than a year and a half ago.
And then, in parallel to that, we also kicked off some big hardware efforts, the biggest of which is an in-house chip. To go from zero—no chip design team, no chip in-house, no chip IP—to launching a chip takes time, it takes many hundreds of millions of dollars, it takes a very, very large organization. But we made the decision in ’22 and we’ve been working towards it. Somehow, it didn’t leak. But it’s now nice that we can talk about it publicly and it’s going to be in the vehicles next year.
An AI-centric approach required this vertical integration of perception. It doesn’t necessitate owning compute, but owning compute allows you to deliver it at a lower cost level and, in our case, a higher performance level.
We just have such a conviction that [autonomy] isn’t just a part of the auto industry. If you look out a little bit, this is the whole business. And so we built this view that where we deploy most of our R&D should be this category.
Was it completely obvious you needed lidar?
There’s a thinking around lidar that needs to be shed, which is that they’re expensive and mechanically complex. The old Velodyne lidars, even what you see on the roads today, they were really complex sensors. But they’re now very low cost, extremely reliable, and solid-state based.
Ten years ago, the best-performing lidar you could buy was maybe $70,000. Five years ago, it was maybe $5,000. Today it’s in the low hundreds of dollars. And so it’s become so cost-effective at turning the entire fleet into a ground truth fleet. It’s really helpful for training your cameras, especially in adverse conditions.
When you peel back the onion and you look at the cost trajectory of the sensor, it’s become this sort of strange debate, because Tesla’s taken such a stance on it. But it wasn’t really a debate. If it was a $10,000 sensor, it would’ve been a different story. But when it’s a few-hundred-dollars decision, it’s much easier to make.
Your new assistant’s Google Calendar integration made me realize that if I don’t have to spend quite as much time thinking about driving, there’s a lot of opportunity to be productive in the car. To what degree are you trying to build a richly powerful assistant?
Google Calendar is just one of what will become many instances of integrations. Today there’s a limited set that have been set up to go agent-to-agent. But any platform that’s going to truly survive, and not just get gobbled up by an AI platform, will need to become very, very capable in terms of enabling agent-to-agent. Effectively like SDKs [software development kits] that just make it very easy to plug in. The goal is that essentially any app that you might want to use, we’ll be able to plug into our agents and it’ll be seamless.
So you can reach across apps like you saw today. You talked to the car, the car was able to reach into Google and find the calendar. We were able to tell it to move something, and it reached back to Google through an agent and moved everything around. It’s the tip of the iceberg.
All these things will start to become so natural where the car becomes like a personal assistant. If you want to move your schedule or order food or schedule someone to come to your house to fix the plumbing, all this stuff just becomes very, very easy to do. And if you are no longer driving the car, you may want to be able to use the car to help you with more of these things.
Do you have any sense as to what the future looks like in terms of robotaxis and autonomous private vehicles coexisting?
I definitely think they’ll coexist. They’re anything but mutually exclusive. The existence of level four [autonomy] is what enables both of them, and the technology from a level four point of view is the same. We’re focused on the tech, and the initial instance planned is a personally owned vehicle. But it doesn’t preclude us from doing robotaxis or rideshare.
Rideshare today is such a small percentage of miles. I used to be of the view that we’d go from 99% of the world’s miles being in personally owned vehicles to 50% being in personally owned ones and the other half of the world’s miles being in shared. Maybe that happens, but I think it’s probably more likely to go from 99% to 90%. Maybe that’s because I have kids now, and the complexities of car seats and soccer balls and soccer outfits.
I think it will be different country to country. When you look at the wealth level in the United States, if you can afford a car today, a lot of people would still rather own one and have the simplicity of it always being available for them and their family. But I actually don’t need to have a strong conviction on this either way. If there’s a heavy shift in the model of consumption, we’re equally ready for that.
I’m surprised how much attention the business model gets. If one end of the spectrum is traditional ownership as we know it today, and the other end of the spectrum is pay as you go, we’re not being very imaginative. There are going to be a lot of things in the middle. Maybe I own the vehicle during the daytime and somebody else uses the vehicle at night. Maybe the vehicle’s mine during the week and another family’s during the weekend. Maybe the vehicle’s shared among five or six people as opposed to infinitely shared.
There’s just going to be a spectrum of new ways to consume mobility, the moment the vehicle can drive itself. And our view is we’re going to exist across that entire spectrum. But the only thing that’s absolutely certain that’s necessary for any point on that spectrum is level four. Robotaxis don’t work with level three. Personal level four obviously doesn’t work with level three. You need level four. So that’s what we’re focused on.