The Exchanges

Q Um, I mean, at the time, did you feel like it's just a matter of time, just a matter of, you know, maybe a few years before this is normal, this is deployed everywhere, we're gonna see this all over the place? Or did you still think it was a long road ahead?

A No, I, I think, you know, it was, it was clear that both were true, that this was going to happen, and that there was a lot of work to be done. Uh, and I think in 2012, I said something about working to make sure that we had self-driving vehicles before my, my son had to get his driver's license. Uh, and so, we missed that by a little bit, uh, and, and I didn't miss it by much, because my younger, my older son still hasn't got his license, and my younger son just got his license, you know, two weeks ago. So, you know, I feel like in the grand scheme of things, that doesn't feel like a terrible estimate in retrospect. Yeah. But, you know, what's exciting is that today, you know, in, in Phoenix, in San Francisco, in, uh, L.A., You're actually able to get a ride in a self-driving vehicle.

AI assessment note: “it was clear that both were true, that this was going to happen”

Q zero, one, two. Um, let's, let's first talk about, like, level two, which I think is where, you know, cars that have some kind of automation, like, um, like Teslas, right? Um, which are called, like, the Tesla calls itself a full self-driving, you know, system or whatever it, it, it, it calls itself, but, um, it's actually not autonomous, right? It's not designed to drive Without a human intervention.

A That's right. And, and, and there's kind of a really simple way to think about it. So level zero is your, your classic Corvette. Then it's basically, it's the steering wheel and brake and gas pedal and a clutch and, you know, very manual. And then level one and two are systems that are designed to help a person. So there's a person who's responsible for driving and whether it's the Ford collision warning system or Forward collision assist, so if you're about to hit something, it hits the brakes a little bit before you do, or lane keep assist, you know, a lot of modern cars have these kind of features, and so if you have one of them at a time, then it's level one. If you have kind of in lane and along lane control, that's level two, and then we'll not talk about level three, because that's kind of a confusing place, but when you switch to level four and five, instead of the system assisting a human driver, The system is actually driving the vehicle, and the human is really along as a passenger, and so for us, the systems we've been building from day one are, uh, aspire to be, are intended to be those level four systems, so it does all of the driving, so our trucks will leave our terminal, drive the frontage road, get onto the freeway, drive down the freeway, get onto the frontage road, drive through the commercial neighborhood, and pull to their terminal, and they do that end to…

AI assessment note: “That's right. And, and, and there's kind of a really simple way to think about it.”

Q as somebody who's wearing Birkenstocks right now, I can say that was probably a good decision, because they're great shoes. But, but seriously, so you get to, to Google, and, um, and the challenge is what? I mean, you were given a challenge to, to, to create a fully autonomous vehicle, like, that can, can go on the roads. Like, what, what did they tell you they wanted to accomplish?

A Yeah, so, so one of the brilliant things they did was, Take the DNA of the DARPA challenges and kind of bring it internally, and so we were given two goals, really, to drive a 100,000 miles on public roads, and then to drive a thousand miles of very specific roads that, that Larry Page and Sergey Brin had kind of picked out as ways to kind of make sure that we had both the breadth and depth of performance, and from that we'd be able to figure out, like, is this 50 years away, or is it Five to 10 years away, uh, and that's whether it's worth kind of a company spending resources on it. And so, you know, the 100,000 miles was pretty straightforward. We, we mapped a part of the freeway in the Bay Area here. We had a bunch of cars that just kind of went up and down it, and we learned from that and refined it, and, and then the other was driving a thousand miles of really interesting roads where we'd have to kind of push the capabilities so we could really understand, like, is this viable? And so it was things like, Driving around Tiburon.

AI assessment note: “we were given two goals, really, to drive a 100,000 miles on public roads”

Q There was a huge leap between what happened in 2004 and 2005. 2005, a bunch of cars finished. You guys came in second to the Stanford team. Um, what happened in that one year to make that leap in technology possible?

A Yeah. Well, I think that, um, what was really exciting about the challenges was that it was an opportunity to focus a bunch of research that was happening broadly and kind of bring it together on one mission, if you will. Uh, and, and unlike most academic work where, you know, the idea is to kind of have the concept and you get the robot to do the thing once and you shoot the video and then it doesn't matter if it ever does it again or if it ever did it before, you've kind of made the point. Here, it was truly a race. And so the thing had to be reliable enough that when you kicked it out the, the shoot first thing in the morning, that it worked the first time and was off on its way. Uh, and so there was, you know, there's a lot of neat ideas we put into it. I think this was the first time where you saw the, the whole conglomeration of what is now a self-driving vehicle. We had LIDAR radar camera, we had machine learning, we had high definition maps, um, all this stuff that is now making vehicles kind of come together in a moment. To solve this problem. And, uh, it was still a bit surprising that we, we made as much progress as we did in those years, but there's a lot of, a lot of people kind of, uh, worked very hard, uh, you know, and enjoyed the, the thrill of making the thing happen.

AI assessment note: “We had LIDAR radar camera, we had machine learning, we had high definition maps”

Q Going back to that time, was it clear to you that this was, this was the future of, of travel, of driving?

A I would love to say yes, but, but the, the honest answer is not in 2004, 2005, right? It was, for me at the time, it started out as this is really cool. Yeah. Uh, right. This is an amazing technology. I get to go drive Humvees in the desert and lasers and cameras and computing, and it was just exciting and interesting and, and very cool. And then as I started to learn About the, the challenges our troops were facing out in the world and how complicated supply lines are, you know, and how vulnerable our troops are while they're operating those vehicles and how much, you know, we lost more people on the supply line than we did on the front line in, in Iraq.

AI assessment note: “the honest answer is not in 2004, 2005”

Q for Google by Sebastian Thrun, who was kind of like one of your, one of your rivals during these competitions. He was the head of the Stanford team in these DARPA challenges, um, and I guess he recruited you to come work at Google on this secret project at the time, which was An autonomous vehicle project. Tell me about, how did they describe it to you at the time?

A Yeah, so this was 2009 was when I started Google, and so Sebastian and I had been, uh, you know, competing in these challenges, and, but, you know, it was a friendly rivalry. We all kind of got that this, there was a race, but it was about moving this technology forward, uh, and we had talked about doing something together and thought about a couple of different ideas, and eventually Larry Page had said to Sebastian, you know, why would you do this outside? We, you know, we want to Do this at Google. Let's put a team together, and so I was there to help, help found the team at Google, um, and it was, it was a very secret project. Nobody knew that we were doing it, you know, because, you know, they wanted to be able to take a risk and try something out there and see how it went, and it kind of had this, um, Dread Pirate Roberts feeling to it. When I signed up, they said, you know, you, we'll give you, Two years to go do this, and then we'll probably fire you, but, you know, it should be fun in the interim, and so, uh, my wife and I and our kids moved out from Pittsburgh to, to the Bay Area, and at the time, it was a very difficult decision, because I'd just joined the faculty at Carnegie Mellon, and that had kind of been my aspiration up to that point, and like, well, you know, when are we gonna have a chance to go visit with the crazy people in California, and we're Birkenstock…

AI assessment note: “When I signed up, they said, you know, you, we'll give you, Two years”

Q So, um, I've seen that, that the expectation is that the commercial service will launch by the end of 2024, which is, you know, seems doable, but also ambitious, right? I mean, just because of the, there's still probably some regulatory hurdles and also just safety. I mean, what, what you're working on, like there's just zero room for error.

A Yeah. And, and this is where we're really proud of the work we're doing. So we've shared publicly how we think about safety, and we've shared this thing we call a safety case, which is how we're going to convince ourselves that this vehicle is safe to operate on the roads. And it, it spans everything from when it's working properly, does it drive in a way that's safe? If something breaks, does it know about it and kind of handle that in a way that's, that's appropriate? Have we thought about how it protects itself if somebody tries to attack it? Do we learn from things that have happened so that we're constantly improving? And do we have a company where we have the right procedures and the right culture and we're trustworthy to kind of deliver something like this? And one of the things that's really exciting is that we can use simulation tools to test the driver against more events than a, than a person will see in a lifetime. So today, one of the things we do is we, we look at all the ways vehicles can get into crashes, and we build simulations of those, and we actually expose the Aurora driver to them, and it's tens of thousands of these things, and we see how it responds, and that level of evaluation is something we, we just wouldn't ever be able to put a human driver through.

AI assessment note: “we've shared this thing we call a safety case, which is how we're going”