You know, I think a year ago, there was a point where if you asked Gemini how to hold the cheese on your pizza, it would say use glue. Technically accurate, but would not result in a good meal, right? And the good news is that there's a person between what Gemini said to do and actually producing the pizza, and they go, "Haha, that's silly. Try again." When the trucks move down the freeway at 70 mph, we just don't have the ability to let that happen.
This is Unsupervised. I'm Dave Ferguson, co-founder of Nuro. Join me as we turn over the big questions in autonomy and physical AI. Chris and I have known each other, I was figuring this out, since 2002, when I came to visit Carnegie Mellon and I crashed on your couch. I was just thinking about that. Yeah, it's been a while. You had to burn it afterwards, probably.
But that weekend literally changed the trajectory of my life. You know, I decided to join CMU. The two of us then worked together on a bunch of different projects: the DARPA Urban Challenge that you led, and then later you pulled me to Google when you led the Google self-driving car project before it became Waymo. And now for the last nine years, you've founded and have been leading Aurora working on AV trucking. So Chris, it's a huge privilege and treat to have you here today. Thank you so much.
Thanks for having me on. It's always great to hang out with you, even, you know, if we've got microphones instead of beers, but it'll work. It's been a hell of a ride. I remember us out in Atwater testing for the Urban Challenge.
Yeah, we're in that old aircraft hangar. And I remember in the side there was like an old rusty pull-up bar, which was where we were getting our exercise. I felt like you were getting your exercise. I felt like it was the guy from Footloose in the background.
Trying to avoid catching tetanus.
And winning the DARPA Grand—
Winning every challenge together, so yeah, which was amazing. Tell us a little bit about trucking. Why AV trucking? You've worked so long in this field. Why is this the right application to be pursuing?
We've been building Aurora for going on 10 years now, and we had this vision. We kind of understood that trucking would be a really interesting application. And why? Because it's really hard. Anyone that you know is a truck driver, you should thank them. But it's absolutely essential that goods move on trucks. If you look around the room we're in, look around any room, literally everything in that room at some point moved on a truck. So, that combination of competence, technical features and capability, market need, economic opportunity, we're like, "Okay, this is where we should go put our energy." So, where are you guys?
We are focused on building the core Nuro driver, the foundational technology that can be a universal autonomy platform. And so, for us, we used to work on these delivery vehicles. We thought it was very exciting—this entire vertical from manufacturing them, owning and operating them. And then we realized, well, we knew for a long time that it required a lot of capital. I think what happened is the market changed, and getting that capital became very challenging. So, we decided, "Look, we should find a path that is more capital efficient for us to get to market and scale and realize a huge impact." And so, what that meant for us was licensing the technology, working with partners to get it on as many different platforms as possible. So, we had this big robotaxi push with Uber. And we have some other stuff in the works on the logistics side that we're very excited about. And really, it's about building that core AI AV technology, and then trying to universally apply it to a ton of different platforms. And for us, it's just made a huge amount of sense that we work with partners and have them do what they are literally best in the world at doing: building cars, running rideshare services. And we get to focus on what we hope and believe that we are sort of best in class at being able to do, which is the AV side.
Yeah, that's like you said, that's been our approach from day one. It's like we're good at this, we can be proud of that, but we should be humble about the things that we aren't, right? And make sure we go find great people to work with.
There's a lot of buzz around and a lot of folks that like to talk about the nomenclature of AV 1.0, AV 2.0, end-to-end models, all of the foundational model technology we're seeing across the OpenAI, Anthropic, Gemini's of the world. How are you or not leveraging all of this technology? What does that look like?
There are places where algorithmic solutions work, right? I take for example, processing a GPS signal. We probably don't need to learn how to do correlation across the signal coming from a bunch of satellites. We can just write down the math and do it. And so we should do that where we can. You can't deny the impact that machine learning and AI has on the technology we're developing and building. The approach we've taken is what we call verifiable AI. And this is really about, yes, we absolutely need to be using modern techniques, but we have to make sure that we are able to kind of understand how they perform and put constraints around them.
I think about where the enthusiasm around large models comes from, and it's obvious. If you'd asked me 5 years ago, would we be where we are today? The answer would have been I couldn't have seen it. And so you absolutely have to be paying attention to that, but you also have to understand how to use the technology in the domain it's applied to. A year ago, there was a point where if you asked Gemini how to hold the cheese on your pizza, it would say use glue.
Technically accurate, but would not result in a good meal, right? And the good news is that there's a person between what Gemini said to do and actually producing the pizza. And they go, "Haha, that's silly. Try again." When the trucks move down the freeway at 70 mph, we just don't have the ability to let that happen, right? Cuz that's the equivalent of it deciding to jerk the wheel and turn off the freeway. We've thought about, okay, how can we take this kind of technique and apply it in our space? It's important to have defined interfaces. Think of it broadly as two systems: seeing the world and then reacting to the world. Each of these is kind of an exponentially hard problem.
Very difficult, very complicated. But to try to make sure we can understand how the systems work, we've actually created interfaces that say, "This is how we express what a person is, what a car is, what the various features of the world that we care about are." And then we basically decompose that model and separate it from the model that reacts to the world. The downside of this is that you're potentially throwing away information, right? If I were to take the bull argument for end-to-end, it's that the model may find a more expressive way to describe a person,
which would provide more subtle cues, which then you could react to. Whereas if I say, "This is how you describe a person," the model is only that expressive.
