The Exchanges

Q Yeah. I mean, GM is, is depending on how you measure the fifth or the sixth biggest car company in the world. Now with cruise as part of the GM family, what did that mean? I mean, did it sort of supercharge your ability to You know, to, to develop this technology at a faster pace and, and at a bigger scale.

A Well, we got access to, first of all, the, um, Chevrolet Bolt platform, which at the time was one of the really only the, one of the few viable EV platforms, um, that you could use for something like this. And, and that was helpful because EVs obviously have, you know, the, the battery capacity to run all of the computer and sensing systems we wanted. We also had access to, you know, essentially the funding we needed to build a larger team, uh, to do a lot of the hardware development and other things that were very capital intensive. So it did kind of take a lot of the restrictions off and let us accelerate, um, the development. And we did a pretty good job, all things considered, making sure we could actually unlock that speed and not become kind of swallowed by the parent company.

AI assessment note: “take a lot of the restrictions off and let us accelerate, um, the development.”

Q And why did you guys decide to start in San Francisco? Because you're based there, or because there, there are other reasons that you want the cars to, to be in that city?

A Well, both. Uh, the first thing to realize is that, you know, we're operating a fleet of vehicles, and to actually turn this into a business, which is our goal here, ultimately, I mean, we can't provide this, this benefit, um, to communities for free. We have to, have to be able to earn some revenue from it to pay for operations. And, uh, you want to fish where the fish are. And so San Francisco is among You know, a handful of major markets in the US where there's existing demand and willingness to pay for ride hail services. So it's one of the obvious first places to start. It's also a really difficult driving environment. It is a major city. It has high, uh, pedestrian and cyclist density. It also has hills and fog and many other things that make it, uh, challenging for a self-driving system to do really well. And so our attitude was, you know, if we want this system to improve at the fastest rate and be ready to go in other cities as quickly as possible, we should subject it to the most challenging environment so that, uh, you know, we know if it works here, the path to unlock other cities and deployed in other cities is going to be pretty straightforward.

AI assessment note: “Well, both. Uh, the first thing to realize is that, you know, we're operating”

Q During your junior year, um, you heard about these guys at Yale, Emmett Shear, Justin Kahn, and Michael Seibel. Um, they had sold something like a calendar software through eBay. You hear about these guys and Do you think that sounds interesting and what you just like email them and say, Hey, can I meet you? What's the story?

A Well, kind of. There was a, there's an infamous, um, uh, mailing list, email mailing list at MIT called the jobs list that was run by a faculty member in the computer science department. And it ended up being a place where Harvard and Yale business school students would basically send out a request for some engineer at MIT to kind of build their app or product for them. Um, and so most of those were sort of, you know, uh, half-baked attempt to, to, you know, frankly, to get some, get at MIT to do the work for them. But when I saw the, the note come across from Emmett and Justin and then Google them, I, I was impressed because they seem to have, you know, some actual experience building and selling companies. Uh, and so I was intrigued enough, especially by what they were wanting to do, which was to build this reality TV show that would stream 24 seven to the internet, uh, that I reached out just to figure out, you know, what they were up to.

AI assessment note: “when I saw the, the note come across from Emmett and Justin”

Q 2014, and that year, you start taking pre-orders for this, this, um, RP-ONE kit, this highway autopilot system. It turns out that you never ended up making these kits at all, because I guess you soon after, soon after you announced it, you abandoned it, and you kind of decided to focus on fully autonomous vehicles for urban environments, not, not these sort of highway autopilot cars. Why is that?

A Well, we, two things happened. One is, you know, our eyes got a little wider as we started looking at the complexities of From a legal and engineering standpoint to build these kits that would work on lots and lots of different cars. Um, and so he thought that was doable, but it was going to be a long slog and we'd spend a lot more time on adding support for new vehicles than we would actually making the self-driving technology better. Uh, so that was one thing. It was going to be a tough business to, to make it work. On the other hand, um, what had been happening in parallel is these ride hailing companies like Uber and Lyft had exploded in popularity. And, uh, if you look at the economics of those businesses, the vast majority of the revenue that comes in goes right back out to the driver. And so it became obvious to us that there's a massive and growing market for fleets of vehicles that, uh, You know, have a lower cost than, than what it would cost to employ a human driver. And, uh, we thought that was enough evidence of market demand that we could do a pivot, a big, bold move to basically abandon the business we had started building and recruited engineers for and raise money for and do something completely different, which was to build robo taxis.

AI assessment note: “two things happened. One is, you know, our eyes got a little wider”

Q and it will come and get you and you unlock the car and there's no one in there and it will take you to where you're going. Before we get into the details of how this works, how did you land on this concept of, of a taxi service? Why does that make most sense rather than, you know, trying to basically turn this into a consumer product right away?

A Well, it's a good question, and there's, there's several reasons for that. Um, the first, as I mentioned, is, is cost. You know, if a vehicle is operating 20 hours a day earning revenue, you know, we can afford to pay a little more upfront than you would for a car that, You go out and buy, but it probably sits in your driveway, 90% of the time. Um, it's not providing value 24 seven like these vehicles can. The other is that we think there's a benefit to getting this technology out there soon, as soon as it can provide a, you know, a safety benefit to the community, but before it could necessarily work everywhere all the time. So for example, we can deploy in cities and have these drive in San Francisco at say 20 to 30 miles an hour max speed. Even though that technology is not ready to go 80 miles an hour on the freeway or drive through a blizzard. And we'll get there eventually, but we can start, you know, getting some of those benefits right now, given the state of the technology. And then lastly, I guess the other piece is, is keeping these vehicles in really good shape. The vehicle we're building called the Origin is designed to last one million miles, which is four or five times as long As the average car will. And so, you know, when this, when this, uh, vehicle drives you and it's responsible for your safety, you want to know that every sensor has been inspected, that the…

AI assessment note: “The first, as I mentioned, is, is cost.”

Q One of the things that Tesla says about its vehicles is that it has so much data already because there's so many Teslas on the roads all over the United States that, um, they have this, this neural network that enables their vehicles to be, to have better information. Is that, in your view, a fair point given how many of those vehicles are deployed on the roads?

A If you're using machine learning in these systems, which most of us are, There's definitely a benefit to having diverse and large data sets. But, you know, the data itself isn't enough to make these systems better. It's actually the ability to analyze that data, extract insights from it, and then use it to improve the product. And so you've got Tesla that maybe has a large amount of, of low quality data. It's just camera data from a bunch of vehicles. Um, and then you have crews and others that have a lot of Very high resolution, high quality data from these lidars and radars and other sensors. And we found that we're not really limited by the data we have, even though we've driven maybe a tiny fraction of what Tesla has. It's really the ability to turn that data into insights and improvements in the product.

AI assessment note: “There's definitely a benefit... But, you know, the data itself isn't enough”

Q how did you raise the money to get started? I mean, just the capital costs of this thing from the beginning were going to be huge. You wouldn't be able to do this with like two or five million dollars. You had to raise tens of millions. Was that challenging, or were you able to do it? On the strength of what you'd already done at Twitch and Justin TV?

A Well, a little bit of both. So any big, bold idea, um, that carries a lot of risk, especially technical risk, you usually have to fundraise in stages. And so you take the, the, the biggest, scariest, you know, most formidable technical challenges and prove that you can solve those on a small scale. And that enables, you know, that builds some confidence with the Potential investors that, hey, if a small team can solve, you know, a big part of this problem in a short period of time, that gives me a lot more confidence that with, you know, a larger amount of funding and potentially a larger team, they could actually go after the full problem. And that was our philosophy. If we could build this proof of concept that showed highway driving, you know, really quickly with a small team, then that would enable us to raise the funding to go for something more ambitious, like, You know, what we ultimately ended up doing, which is full driverless robotaxis.

AI assessment note: “Well, a little bit of both. So any big, bold idea... fundraise in stages.”

Q Grand Challenge. And, and that, so you're thinking about a bunch of different possible things to do. You land on self-driving cars. This is a, an enormously challenging problem. But this is something that you thought, okay, it's 2013. I'm going to put at least 10 years of my life into building this thing. And what was the sort of the product or the technology that you wanted to commercialize?

A Well, I mean, I guess as I started thinking about self-driving cars, the first question that popped into my head was why, um, or, or, you know, is this worth doing? And it doesn't take long if you poke around at statistics, you know, even today there's 40,000 people that die each year in car accidents, you know, a million people injured in the U S and, um, almost all of those accidents, there's some form of human error involved. And so going back to this notion of automating the sort of Repetitive or mundane tasks. Like, you know, if we can chip away at that, that's a really big deal and a problem worth working on. And just thinking about the time we all spend driving, it is a catastrophic tax on society and productivity. The amount of time we sit and spend, you know, sitting behind the wheel. Um, so anyways, I thought that was worth, that was definitely worth doing. There's clear benefit to society. I also knew though, that Google had been working on this for a while. You know, after the DARPA grand challenge and, and rumor had it, they had spent something like a hundred million dollars, you know, to date and they didn't have a product yet. And so I thought if I'm going to go after this, I can't do exactly what they've done. I don't have that kind of resources. I don't have, you know, an army of brilliant engineers from Google. So I thought, you know, maybe that what's the sim…

AI assessment note: “designing a system that could Drive for you when you're on the highway.”

Q It did pull over. Okay. A group of cars that blocked an intersection for some time. There's a lot of attention that gets paid to those mishaps. First of all, what's your reaction to that? I mean, I mean, do you get frustrated? Are you like, God, no one's seeing the good side of this? Or do you understand why people sometimes focus on those things that cause disruption?

A Well, look, I think these vehicles, um, are very clearly, you know, autonomous vehicles. They have sensors on the roof, they say crews on the side, and there's no one in them. So they draw a lot of attention. And anytime they do anything that perhaps, you know, a human, if it were a human driver, you would not even really notice, uh, it makes the news. And that's, that's to be expected. But what I think is really important is that we not lose sight of the fact that, you know, we have a real problem today. With car accidents. You know, I have a four year old son that I, that's on my mind all the time. I have a grandfather I never met from a car accident. And I think as a society, our goal should be progress and we've fallen behind. And I think that, uh, self-driving cars are the only thing that I've really seen as a, as a potential solution to car accidents, because humans are always going to be human. They're always going to make mistakes. That's, that's part of being human. These AVs, autonomous vehicles, you know, they're safe. Our track record for safety is excellent. But then, you know, they won't make zero mistakes. And I think they'll make far fewer than we as humans do collectively. And so I think there's a distinction between things that look kind of odd that AVs might do versus the real safety impact that they can have, you know, currently and over the long run. And so…

AI assessment note: “it makes the news. And that's, that's to be expected.”

Q the full self-driving car, which is in, in beta, and many Tesla users have access to it. They only use cameras. They're, and, and their argument is that LiDAR is unnecessary, that it's expensive, and that actually, It can't tell the difference between a floating plastic bag and a speed bump or, you know, or a dog running across the road. I mean, what do you make of that argument?

A Well, it's a good question, and I think a lot of attention gets put on this debate around LIDAR, but if you don't mind, I want to take a step back. I think it's kind of the wrong way to look at it. You know, so our, our philosophy is, you know, in our business is we operate robo-taxis, and so it only works if they don't have a driver. And so we have optimized for vehicles that are well-equipped. They have all the sensors they need to do that job, like literally no driver, which is a very hard task. Uh, and that's where we start for, for our business model. Our goal, which I believe is also, you know, the goal of Tesla and other companies is to eventually get this to be low cost, you know, works everywhere and doesn't need a driver. And if you look at the business model of Tesla, they're just starting from the other side, which is these vehicles aren't well-equipped. And they don't yet work everywhere and they still need a driver, but they're, but they're starting on a lower cost point and they're hoping that, you know, maybe they'll build software in the future that will actually let you take the driver out of the car. Similar for us, um, we already have the drivers out of the car, but we're hoping in the future we can make it cost much, much less. Um, and so it could eventually be on a car that you could go out and buy. And so for LIDAR and cameras, it really just comes down t…

AI assessment note: “And so for LIDAR and cameras, it really just comes down to cost.”

Q regulatory rules. You needed the, presumably, the state of California to give you the green light to let you do that. I'm sure at a certain point in your, in the development of the technology, you felt, or your team felt like we're ready to go, but it was probably, that was probably a year, maybe two years before you were given the permission to do it. Is that right?

A Well, in this case, you know, there are regulators, uh, especially the California DMV who want to make sure that we followed best practices and taken obvious steps to make sure that, you know, we can be on communication with the vehicle and we put it through the necessary or appropriate tests, given how we intend to operate it. But that's not sufficient by itself. We, we went a lot further and, you know, safety has to be our top priority and it is. And so at that point, you know, with our Testing that we had done with this system running on the road, but someone sitting behind the wheel just in case we had collected millions of miles of data. And from that data, we have extracted hundreds of thousands of situations like close calls or human drivers doing weird things or all these strange, you know, they call them corner cases, things that don't happen very often, but the, this autonomous vehicle still has to handle them properly. And, uh, turned those into simulations where we could, each time we make a change to the software that drives these cars, we could run it through hundreds of thousands of scenarios, almost like, you know, going, taking a driving test, but, you know, times a thousand to make sure that, you know, no matter what kind of situation the AV is in, we understand how it's going to behave and what risk it may be encountering. Uh, and then we can make a much more…

AI assessment note: “But that's not sufficient by itself. We, we went a lot further”

Q So tell me where, where this is headed now. Now you're in San Francisco, and you're about to expand to Austin and Phoenix, and anyone can download the app, right? And then you can get on the waiting list to get a chance to use it. How, on average, how long does it take to, to get off the waiting list?

A Um, it's, we have a lot of demand right now. Um, demand far outstrips supply, so it can be a while. But, you know, for, for the longest time, for seven or eight years, our focus was just getting the first vehicles to operate as robotaxis without drivers, and doing all the work to make these products safe, and, you know, handle any kind of thing that could go wrong, and, and literally anything. There's, there's a lot of work that goes into that, whether it's a sensor malfunctioning, or computer crashing, or Even someone damaging the car from the outside, no matter what it is, the vehicle has to behave responsibly in those situations. Now we're at the point where we have the robotaxis out there in a city like San Francisco. And so the technical problems, the software and AI problems are no longer the biggest bottleneck. Now we're turning our attention back to manufacturing and scaling and building out charging and cleaning and maintenance infrastructure in cities. And so you're going to see us pop up in multiple cities and with more cars and more service area pretty rapidly over the next year or two.

AI assessment note: “demand far outstrips supply, so it can be a while. But, you know,”