Rethinking Transportation Safety with Reid Hoffman and Nauto CEO Stefan Heck | Greymatter

Hi, everyone, I'm Reid Hoffman, and welcome to Gray Matter. At Greylock, we're trying to help build massive companies. So this means practitioners who work from the earliest stages of building these super interesting companies that kind of be a couple of people in a garage to scaling to global relevance. And so we try to be practical, we try to be entrepreneurial, we try to be detailed in the things we're doing. Some of it will be kind of company building, and some of it will be interesting trends in technology. And actually, I'm delighted to be here with my friend Stefan Heck. We've known each other for nearly three decades, which both of us are kind of hanging our heads. Boy, we're part of the old crowd now. We have talked in our three decades about everything that ranges from the nature of friendship and going through life to how organizations have great cultures, to technologies that change the world, to philosophy, neuroscience, and everything else. Here we will try to be a little bit more contained to the entrepreneurial community, and we will essentially be what is going on with this new world of autonomous vehicles with artificial intelligence and how Natto. Stefan's company is one of the really interesting kind of unique plays within the space. Stephan, it's great that you're here.

It's great to be here with you. Look forward to our conversation.

Let's start just a little personally, because one of the funny things about Silicon Valley biases, it tends to be the we're doers, we're not consultants. And part of the thing that I think that you've demonstrated so well has been the shift from actually McKinsey into entrepreneurship. But I think we should probably do very quickly the whole arc. You and I were undergraduate together. I stopped you to ask you why you were wearing an Austrian a raincoat that looks somewhat like an elf cloak to anyone who isn't from Austria. And you're like, how do you know what a wetter flick is? And I was like, well, I spent some time in Vienna, and then from there you would come to Stanford to do the same major. We were both symbolic systems majors, both artificial intelligence focused from back in the days when it was much less Hot Topic than it is today. We both decided to go to graduate school, but you actually did a PhD. So, first, what was the path to the PhD?

Well, I came to Stanford because of a book, actually, Douglas Hofstadter's Godolescherbach. That's what got me interested in AI and my algorithm for finding a college in the US Applying from Austria was, do you have financial aid for foreign Students and do you have a program in AI? That's how I wound up at Stanford. Of course, at that time was just past the 60s and 70s hype cycle of traditional AI. And so I arrived just in time for the disillusionment of this is a nice academic toy, but it's not really a practical application. So I got more interested actually in human beings. Did a lot of work on what was called connectionism at the time, which was neural networks, well before the term deep learning had been coined. Working with David Rommelhard and others in that era, and did a bunch of neuroscience and so went to graduate school because I found human beings were more interesting than the AI algorithms and approaches of the time, and started looking at philosophy of mind and how the brain really worked. Worked for my PhD with Antonio Damasio and Paul and Pat Churchland on how the brain works. And actually, interestingly, in a way that sort of life circles back. Worked on the origin of visual perception in human beings and how concepts are formed, and built a big simulator on what at the time was state of the art hardware, the next machine to simulate visual systems and see how concepts arise. All of that is toy problems compared to today's deep learning neural networks, but is at the center of how do we recognize objects, what are objects and what do they mean, and how do we relate to them? And the thesis I pursued was essentially that through interacting repeatedly, you could bootstrap a practice and you could get good at something and develop expertise. Which of course, 25 years later, almost 30 years later, is exactly what we're doing with autonomous cars. Having them drive around, having them recognize people, object trees, traffic lights, and hopefully avoid running into them. And from that, learn how to drive. Well, at Nauta, we have a little bit of a unique approach because we're actually watching the humans inside the car as well, to keep them safe, to realize when they're distracted. And one of the big insights is the entire thing, the driver, the external environment, the other traffic participants, the vehicle itself, they're all participants in one system. We add this element of cloud AI and onboard AI to that, and the whole thing becomes a learning system. So we're watching how if you encounter a dangerous situation as a driver, you can respond to that well as a human, or you can respond to that poorly, or if you're distracted by your phone, you may not respond to it at all, which is definitely a bad outcome. And what's interesting for us is by studying the human drivers, we can see who over time turns out to be a good driver because they're safe. And we have their accident history, we have insurance data. So there's about 15% of human drivers that will never cause an accident. They may still get hit, they may still wind up in an accident, but it won't be there triggering it. And we want to learn how those people drive because they've mastered a skill over 20, 30, sometimes 50 years. Particularly in our world. It's professional drivers driving commercial vehicles are really expert drivers. And we learn from them how to avoid all kinds of close calls. The other end of the spectrum is the not so good drivers, which is 15% of humans cause 86% of all accidents on the road. And we collect all those screw ups on the road because if you're trying to build a robust autonomous car, you want to make sure your autonomous car can handle getting involved in traffic with those people. And a lot of that is learning how to drive. Teenagers are notoriously bad, about four times as risky, but they're just individuals who are bad drivers. And all of us turn into bad drivers when we're distracted, which is the number one cause for accidents. So it's funny how that old history of doing a PhD in Philosophy and cognitive science and studying neuroscience actually comes full circle now when it comes to cars. I was not a car buff as a kid, but I'm really passionate about transportation now because it's such a big part of our life.

Well, we're going to come back to Naoto and the target of building AI that's kind of symbiotic with the human being. And part of what that is not surprising for your background. That's the target of the company and the product and the mission. But let's first actually dispel one kind of classic Silicon Valley myth, because one of the things I think is useful for people to hear is everyone goes, oh, there's consultants and there's entrepreneurs. And part of the thing that I've actually learned from working with you and a couple other folks is, yeah, there's a swath. And generally people who go into consulting because they'd like to talk more than they'd like to do, but there's also people who are also reflexively doers who then bring the conceptual frameworks from McKinsey. And so you went and built a practice, it was the green practice, how business can be better by being green. You wrote a book on it, Resource Revolution. And then you kind of said, okay, this is interesting, but I want to actually build something new. I see something here. So what was that Transition like from consulting to entrepreneurship. And then what were the things that you quickly learned is the predictors for how your experience in consulting actually helped you with entrepreneurship. And then what were the additional things you needed to learn from the kind of consulting background?

I've heard a lot of this. There are different worlds, they don't mix. And certainly having been in consulting, I met my share of consultants who only talk but can't do. But from my life, my own experience, it's much more a blend of building and reflecting and thinking throughout. So certainly doing a philosophy PhD is more reflection than doing. But the hidden secret is even in graduate school because at the time the UC system cut budgets, I had to pay my own way through and I built my own web design company in the early 90s to pay my way through graduate school. So buried inside that philosophy and cognitive science PhD is actually a little e commerce programming shop, certainly in McKinsey. A big part of what I did was very analytical, helping advise other companies, solving tough intellectual problems, but not having to be the one to implement and then to execute. But a part of that always was building a practice. As you alluded to. I grew the semiconductor practice globally. It's running a mini business inside a larger company, right? You have a P and L. You have a recruiting challenge of drawing people in. You have operational issues, deploying them. Now it's all inside a big mothership. And there are a lot of things that you have to do as a startup entrepreneur that I didn't have to do because you've got huge support infrastructure of it and people you can call legal experts. So it's much less bootstrapped. But you're still building. You're responsible for growing a mini business inside a larger company. And mini is not that small. They're a couple hundred million dollars in revenue for each practice and global in scale 80 countries. For me, that mix has just shifted again. Now I'm doing even more hands on building. But even at Natto, I do a lot of reflecting about where's the space going in these very leading edge nascent industries like autonomous driving. A lot of it's around looking ahead two or three moves, right? What is going to happen when we have the first serious innocent bystander killed by an autonomous vehicle and the backlash from that? What questions are people going to ask? Similarly, we have focused ahead of where the industry is really focused. Most people today are trying to get highway autonomous driving going. We're deliberately looking at urban autonomous driving because there's a different set of problems there that everybody will ultimately have to get to. And that has served us very well. And we're approaching that by looking at human driving because you can do that safely today, not have to worry about what is the vehicle going to do if it's a prototype autonomous car. For me personally also, the blend is really satisfying. I'm a very hands on builder and I mean literally physical. Right. I love to build tree houses, furniture, woodworking and so I've always had that side. It was more a hobby during the McKinsey years. Now it's more primary, we're actually building product. But similarly, I get dissatisfied if I have to just mechanically do stuff over and over again and I can't innovate and transform and reflect. So I love that duality. I've met a lot of people who say, well, consultants can't be entrepreneurs. And while that might be true for some, I don't think that's a good rule to follow.

I completely agree, which of course is one of the reasons why I'm an investor and on your board.

Yes.

What would be the one piece of advice you would give to people who have a bunch of experience with consulting with a kind of normal background to say, make sure you're paying attention to X or Y when you choose, change into entrepreneurship and startups.

I think there's an obvious piece of that and there's a not so obvious piece. The obvious piece, and a lot of people leave consulting because they want to be more of an operator, is to live and die by the recommendations you make and actually implement them and have to carry them out. And that's in a very operational way. What product decision are you going to make, what operational implementation you're going to make? Less well known about McKinsey is 40% of the work is actually operations work. I did a lot of work in semiconductor fabs, actually tuning machines, recalibrating, doing equipment maintenance. So it's very operational even though it's consulting. But you're not writing PowerPoint documents, you're actually doing stuff on equipment. So that part is more obvious. And I think a lot of people actually have that yearning. After a couple of years at McKinsey, okay, I've seen a lot, but now I want to do it. The hidden part that really surprised me more is how much you're dealing with people issues in a startup environment versus in a consulting environment. Consulting has a kind of rhythm. Right. People get reviewed every six months. McKinsey is explicitly up or out. So you either advance to engage your manager partner or you're expected to leave. And if you don't leave on your own, then you're asked to leave. Startup environments aren't quite that way, although we've implemented some of those practices of giving regular feedback and thinking very dynamically about the roles. But there's a lot more just every ordinary issues of who gets along with whom, what are their goals, even though that may not be the thing you need to do in the startup right now. So making an engineer happy, working on something they think is cool, but also something that just needs doing, that may not be sexy at all. I spend a lot more time on that kind of people management than I used to in a firm like McKinsey, where you've got a particular, relatively narrow profile of the kind of background, people that are hired and how people view that. Tour of duty, to use your term. Consulting for almost everybody is inherently a tour of duty for many startups. People don't necessarily think of it that way yet. And so there are challenges about how you get people deployed in a way that is both fast enough for the company, but does allow them some amount of growth for their personal aspirations and their professional desires of what they want to learn. One other thing I would add is, you know, in the McKinsey world, every once in a while some team would not function. There's personality conflicts, the problem's too hard or you don't have the right people background for that project. So it happens typically in a McKinsey environment, you could reset, you could even redo the project, offer to pull the team out, stop the work, all those things are perfectly fine. In a startup you don't have that luxury. And so mistakes in hiring, mistakes in how you configure a team are much more costly. And we've had our share of near death experiences and setbacks where a single person behaving in a way that's not constructive in the team can make the whole thing slow down dramatically and you don't get a do over.

In most startups, do over is the next startup.

The next startup. That's right.

Which is painful. Let's shift to Naoto. Let's start with two of the angles in which I thought Naatto was particularly clever. One of them is that you're a systems thinker, right? So it's kind of a how does this whole system come together? What is the system of the road, the transport, the people? And then the second was it's not AI or human, it's AI and human together. So why don't we start with the second, the AI and human, because this will build to the systems. One of the illusions that people have is they kind of think there's going to be a switch flipped and then everything's going to be AV on the road. There's certain places that would be a much easier environment, but it's never going to actually play out that way. That's just not going to happen. How did you start thinking about, okay, this combination of AI and human, how AI can help human, how AI or autonomous vehicles on the road together with humans on the road, what is that blended environment going to be? What are the necessary components of that? And how is that central to Naodo's vision?

I mean, I've come very naturally to looking at the human as part of the system. And so for me, when I approach the problem of what is safe driving, I look at, well, how do people drive safely, right? And most of the time, most of us do it pretty well. It's about 100 million miles between serious accidents where people get hurt or die. That's actually an amazing performance when you really think about it. That's a lot of driving between serious events. And so I also, just broadly, going back to our undergrad days, I think of AI as a tool, not necessarily as sort of general AI that will replace people and turn us all into cyborgs. I don't come to AI with the immortality dream of it's going to replace my body and make me immortal. That's not my aspiration. I do think it's like any other phone or hammer in the old days. It's a tool that helps us be better, helps us ultimately, in the spirit of enlightenment, helps us improve ourselves. AI, of course, as you and I have talked over the years, is unique in that way because it's also a reflection on US consciousness. What does thinking mean? So it's an especially unique kind of tool. Not like a hammer that way, but it makes us reflect. So part of driving for Nauto, for me, is you have to study humans to learn what they do. A lot of driving is interaction between humans. And to come back to your question, we could deploy autonomy today if there were nothing else on those roads. So there are great examples of autonomous systems. You go to Heathrow Airport or a lot of other systems. Enclosed tunnels with no pedestrians, no animals, highly constrained linear tracks. We have autonomous vehicles today. Now, they're not doing very much. They're not navigating intersections. They do stop at stations and onboard and off board people, but that's about it. As you open that up to more real world where you have kids, you have dogs, you have bicyclists, you have trams or buses. It gets a lot more complicated. And that's really the hard part about autonomy. And I think when people worry about, oh, we'll all be replaced and all drivers will be replaced, that certainly is a concern for the longer term, because eventually the interesting question is, and it's an ethical question, we'll get to the point where autonomous vehicles are better than most humans. We do kill, you know, 43,000 people every year in the U.S. through driving. Most of those don't need to die. If you look at airplane travel, it had the same fatality rate in 1956. And with the last plane crash that killed anybody was more than 10 years ago now. So we've gone for a whole decade without any. And I think the same thing is largely possible for ground vehicles. There'll still be corner cases of freak accidents. So that ethical question of should we no longer allow humans to drive? Should it be all vehicles, or should we allow humans to drive? I call it, you'll be driving in safari parks, you'll go someplace for human driving that'll be fun and entertaining, but it won't be an ordinary commute to work experience anymore. That question is still way down the road for us. We're really looking for 25, 30 years at a very mixed environment. The autonomous vehicles will take a while to be developed at all. They'll take a while to be accepted from a policy, regulatory environment point of view. They won't come in all vehicles and types immediately. People are much less comfortable about a huge truck driving autonomously than about a little shuttle bus. The problem on highways that are already limited access is much easier to solve. Low speed pedestrian zone driving is easier to solve than kind of vehicular traffic that sits in the middle. Cities are harder than rural roads. So we'll see it kind of come in piecemeal. There's a long transition period. And so for anybody to bet holistically on, oh, I'm going to spend a couple billion dollars and build a fully autonomous level 5 car at the top level of autonomy that really drives all on its own all the time, that's a very ambitious project with a very long, uncertain return. And part of the beauty of NATTO is we provide safety systems for humans today where the humans are the ones driving and the AI is really just observing, watching, warning the human when they're getting into a dangerous circumstance, watching and warning the human when they're repeatedly engaging in risky behavior. So we warn commercial drivers if they're persistently tailgating. The number one cause of accidents is distraction in the government statistics is about quarter of all accidents. In reality, it's more like two thirds, three quarters of all accidents. And we can warn you if you take your eyes off the road for two and a half seconds or more, and it matters, you're moving, there's something coming up. People find that helpful. So that's purely using AI to augment what you're doing. That's like AI on your phone, sorting your photos for you. It's not replacing you at all, it's helping you. There's a middle world where the car may start to drive on its own some of the time in certain circumstances, you know, boring stretches of highways or mobility at low speeds for people who either are too young to drive or don't have the sensory perception to drive. Driving at very old age. My grandfather, when he was 95, was a very risky driver. We got him a little electric scooter which was fine, low speed, but if there were autonomous vehicles, that would have been the right answer for him. And so that's another stage till we get in your car and tell her where to go. That's actually still a ways off. So for me, the humans are an essential part. Both because mobility is really a service, it's really meant to enable. It's almost to the level of a basic. Right. I mean, it's interesting when you look at productivity data in the us Transportation is embedded everywhere. Every time you go buy something, every time you get something shipped, grocery, wholesale trade is all driven by transport. And yet transport as a system is horribly inefficient. So we haven't actually optimized that, which means we're underserving that human need. And just one very poignant example, if you look at the unemployment statistics, almost half of the unemployment is structural because people can't afford to get to where the jobs are. They live in a location that doesn't have enough jobs, and they can't afford a car because they don't have a job. And then in the absence of public transport, they can't afford the trip to get to a job. So classic catch 22. If you had shared autonomous mobility as a service, which is where I think we want to head, they can just buy that per trip and actually begin to solve some of those problems. So that's a good example just of how rigid our current model is, but very interesting opportunities for AI to help rather than to, you know, it's not Skynet within the cloud as a threat at all.

Well, one of the particular things that I thought was exciting and unique about Natto and literally everything I've seen in the world, because I've seen a lot of autonomous vehicle pitches and everything else.

And ridden in some, I'm sure, and.

Ridden in a few, and kind of gotten a sense of what the design strategy and the kind of what the target, the product, the service, the technology, the design goals of these things are. You actually said, look, this is going to be a blended environment. It's going to be humans and machines. There'll be cases where there'll just be autonomous vehicles, but there's going to be a very strong, almost considered, maybe even a human lifetime or a human driving experience, length of time where it's blended. All of the autonomous vehicle approaches are taking a physics oriented approach, just like I just purely predict it like it's a rock hurtling through space with some interesting characteristics and nothing about the fact that it's a human environment with human strengths, human weaknesses. Humans are very adaptive, as you've mentioned. They drive actually surprisingly well across the average driver circumstance. And yet that blended circumstance is what's going to make successful autonomous vehicles. That's what's going to make a vehicle pattern work. And so you said, this is how I'm going to attack this. So unpack a little bit of the theory of kind of what this future blended environment's going to be like and why Nauto has some of the interesting early steps towards it.

The motivation is really that humans are an essential part. We're not going to ban humans from our roads. That's not going to happen. And so we need to plan for that environment. It's interesting if you look at the history of where Autonomy came from. It came from robot races in the desert, funded by darpa. And there was literally no obstacles. I mean, you had to deal with terrain, but there were no cities, there were no people. We moved to the DARPA Urban Challenge, which had buildings and mock inhabitants. So it got more difficult. But a lot of the early efforts on Autonomy were really centered on the highway use case. And in the highway case, by law, you've eliminated the pedestrians, you've mostly limited the animals. Lots of deer strikes in the Midwest, but generally you've gotten rid of any other hazards. We've made it a nice smooth road with nice even gradients and you've got clear lane markings. And as long as you stay in your lane and don't hit the guy in front of you, which is a classic physics problem. Definition, right Object hurling through space. What's the velocity and how do I guide it? It works. And that's why we see autopilot and autocruise and some of these modes having solved that particular constrained problem actually quite well. That's not how cities work though. Cities are messy. There's people going every which way. And so the interesting thing for us was to say, can we attack that problem? And what you very quickly realize is a couple different things. One is it's no longer just a physics problem, It's a social coordination dance. When two people pull up to an intersection, their timing might be governed by a stoplight in some cases. And what do you do when it's yellow? Do you quickly finish going across or do you stop and yield to the other person? More complicated four way stop signs. We have this rule about if the person is there first, they get to go first. But if you arrive there at the same time, then you yield to the right, which a lot of people don't even know. Well, what counts as the same time, what counts as stop? We all know the famous California stop. Most people don't actually really arrive. They kind of roll through at low speed. And so you have this social turn taking behavior. And there's really interesting sociological studies about who really gets to go first. And it turns out in reality, among humans, it has to do less with who arrives first and more who looks more hurried and who drives the fancier car. And then you get into uglier racial elements of it as well. Right? Men get the right of way earlier and so on and so forth. So all of society intrudes on what should be a relatively simple rule based system. So the DMV says yield to the person to your right. Human reality says we negotiate. And all of the expectations play a role. Even more interesting are the fact that some parts of our transport system recognize reality and some parts deal with abstractions. So a good example is speeding. All of our roads have some kind of speed limit. The traffic engineers building the road are the realists. They say, okay, the post speed limit is the 85th percentile of the actual speed on that road. They know everybody speeds a little. You know, you and I do, right?

I don't admit it.

And so we edge to the limit and maybe a little bit over it. Nothing really dangerous. Obviously if you speed a lot, you'd be much more risky and much more dangerous. But actually, literally going the speed limit, as many current prototype autonomous cars do, turns out to be Less safe because I've had this experience on i5 and my Tesla, you know, it's sitting there going the speed limit in autopilot, and it leaves a long gap for safety to the car ahead of you. What's the result as everybody tries to get around you, to pass you and cuts in and your car keeps backing off to create more space and then the next guy cuts in. That's less safe, both for you, because now you've got all these people cutting in in front of you. It's also less safe for everybody around you because they're now essentially trying to avoid a slower moving bottleneck. So we look at that and say, how interesting here, the autonomous car should be thinking about the human dynamic, which is we actually want to coordinate. We want to belong into one queue that moves at a harmonious speed. In fact, the great promise of autonomous cars is, of course, we can reduce the distance between the cars and we can move more coherently. And a lot of these traffic jams we suffer from today are because you have built up shock waves, right? You're in front of me, you hit the brakes, I hit the brakes a little bit harder because I don't know how much you're going to hit the brakes and whether you're going to come to a stop or just slow down. If you and my car were talking, they would say, hey, we're going to reset from 65 to 55 miles per hour and do that together. And there'd be no shock wave passing back. In reality, it keeps amplifying. So there's a huge opportunity, if we think of this as a social activity, to make congestion much better. The other example, of course, is sharing rides. If our phones know where we're going and can coordinate that through data in the cloud, and the cars can ride share. Autonomous makes that even easier because they can deviate slightly to pick us up. We can get into the same vehicle and reduce traffic even more. But then you have social and safety issues around that. How do I know it's safe? Whose car is it? By having all of this data in the cloud and having the systems in the cars to ensure that the two humans inside the vehicle are getting along, the vehicles with each other are getting along, we can dramatically improve the efficiency of the system. Coming back to, why do we see this as a human and machine? It's not just that we have to coexist for 25 years, but actually, intrinsically, cars should fit into the human environment. I've talked about how what we really want to build is not an autonomous car because auto means self in Greek. So it's the self mobile device, automobile. But now autonomous means it's following its own law, nomos being law. So the original vision is, hey, it's isolated. It does all its own computing, all its own intelligence. It can run it all by itself without talking to anybody. That's the wrong concept, actually. We want a harmonious car that fits into the rest of traffic, fits into the rest of society, that takes human needs and human response times into account and knows if that's a kid, it's going to behave a certain way. If you're in front of me and you're coming up on a traffic light before you even hit the brakes, if the traffic light turned red, I know what you're going to do. As humans, we take all that for granted. We know so much about our environment, we know what we would do in that situation. Empathy allows me to predict what you're going to do that is way beyond autonomous cars today. And so we have to actually program that social world and coordination into the vehicle as well. I'm not claiming we've solved that problem, but we're at least studying and gathering data on that problem to begin to build some of these in. So a very simple one is, when do you break the DMV rules? When do you break the law for the benefit of everybody? That sounds bizarre. At first pass, right? And I'm not talking about driving drunk or driving at 100 miles an hour. I'm talking about very ordinary things where there's a double yellow line that technically the DMV says you're not supposed to cross. But if there's a child playing at the side of the road or a bicyclist that just stopped and has to dismount their bike, or even a person in a wheelchair or somebody who needs more time to cross, you might actually deviate across the yellow line. Again, assuming there's no oncoming truck. That's a very complex judgment. We see that human drivers do that all the time and do it very successfully. They factor in all this context, and they will routinely ignore a literal law. And of course, they don't get ticketed because the police officer makes the same judgment. Right. If they're standing by the side of the road, they would say yes, in that case it was okay. And a lot of the carmakers talk about cones. Why? Because cones are basically indicators that you should deviate from the normal rules. And cones are very hard for autonomous cars to process appropriately because it suspends the normal traffic rules. I drove on the Autobahn recently, and at the construction sites, they paint a set of yellow lines over the normal white lines. And you as a human have to figure out which set of lane markers do I know, follow. Incredibly confusing for any kind of autonomous system to figure out which one is now the relevant one. So lots of work to do.

And the complete arc describes a bunch of the systems thinking that actually really requires. It's not a individual autonomous vehicle. It's not the Western liberal. We're all a bunch of individuals. It's actually, in fact, we're a bunch of individuals in a group, in a society. And a networked harmony is actually, in fact, the target together with judgment around, questions around. All right, well, we've all crossed that double yellow line going, oh, I'm going to give this unknown thing a little bit more room because I can see behind me, I can see there's no traffic coming. I can go around it. And if you had an AV that just stopped and then the car behind it gets frustrated and irritated and it pulls around it. I mean, one of the things I think that people don't realize about a blended autonomous vehicle environment is humans are likely to behave much worse around autonomous vehicles. Like, oh, I can cut you off, I can do it. Because it's like. Whereas they realize, oh, no, in the human circumstance, we don't do that because we understand what the parameters of retaliation, being able to call the police. So I think people are planning, the pure AV environment is planning very badly for that future. And part of the thing that NATTO is doing is not just that blended future and the harmonious future, but even an empathetic one, which is, I think, key.

It goes back to our philosophical discussions, right? We want a social system that has rules, that has guidelines, but we want it to reflect the human needs and we want it to reflect human judgment. You don't want blind application of a set of rules to situations where they actually don't apply. As humans, we automatically calibrate. We know it's more important to protect a life or to help a child out. It's very interesting. A lot of the work that Wendy Ju and others are doing on robotics. The robot that asks for help winds up having a different level of capabilities than the robot that pretends to know all. And it's because of the way the other people respond to the robot that pretends to know all versus the robot that is asking for help. We're empathetically tuned to help out the robot to learn, acquire skills and to do things it can't do on its own yet. So in some ways, we need to make these autonomous cars also ask us for help and be helpful to each other, which is a difficult thing for machines that have for 50 years now served to isolate us from all of our surroundings. Right. And I talk about one of the things we need to invent is the like button for driving. Because right now we have the horn, which is generally used in its negative connotation, and we have the middle finger. We don't have a like button for, yes, you yielded to me or you let me in. Thank you so much. Like, sometimes we have the wave, but even that's not universal.

And you can't always tell, is that a wave or a finger, not quite sure which.

How many fingers were you waving exactly?

So let's step down from the system level, which I think is critical in these kinds of things. Transport as a service. What is the future going to look like? How are products going to fit into it, into some of the more tactical details? Because part of the entrepreneurial journey is, say, I envision this future. I see what the actual navigation path, the road to it looks like in this particular circumstance and what it needs to have as part of it. So you get all the way to building hardware. So frequently in Silicon Valley, we tend to go, well, bits versus atoms, because bits are easy and atoms are hard. And so the advice that entrepreneurs get is, no, no, stay as close to bits as possible and stay away from atoms. And you kind of looked at this and said, no, no, no, actually, in fact, we need to do atoms. Now. Part of it is to build that social environment. But what were some of the early experiences and learnings from building Adams products?

Well, I have a lot of battle scars from Adams Products, including your fellow VCs that don't love Adams, who all said, ooh, hardware, automotive, yuck. This is before, of course, autonomy became a hot thing to invest in. And we said, no, we actually need to retrofit additional sensors, which the cars don't have. We decided early on to focus not on luxury vehicles, but on ordinary commercial vehicles, because, as I already mentioned, that's where the best drivers are today. But that's also the vehicles that get the most mileage, they have the most driving experience. And so we're using commercial vehicles as a way to capture lots of learnings quickly. But generally, when you're a business owner, you don't buy a fancy BMW or Tesla for your commercial delivery. You're buying a Prius or a Ford van or a basic pickup truck, depending on what your business is. And so we're working with those kinds of vehicles, any kind of bus, any kind of truck, sedan, pickup truck. The hard part is getting something to actually survive in the real world of an automobile. Not only is it hot and cold, you're baking in the sun, people beat it up, it has to survive lots of miles. So we've spent three generations of hardware, beginning with what was essentially the guts of a mobile phone put into an automotive cooled fan, cooled box and so forth, evolving that to really understanding what does it take and little things that we take for granted. Right. Our mobile phones basically are designed to live in our pocket and to have these very brief peak usage while we play a game, but then rest for a while. Most of the time you're browsing or doing email, you're not doing something that intensive. If I want to run real AI on a device like that, I'm talking about fully loading the gpu, all the CPU cores, all the time, constantly for a 12 hour drive. Very different kind of problem. So if you look out there, there's a lot of people making dash cameras today. Most of them claim they do some kind of intelligence feature or adas, but we've tested everyone that anyone's ever recommended to us and nearly all of them really don't do that. Or they do that for five minutes but then overheat. And so that's what working in the world of atoms really means. You've got thermal constraints, you've got mounting issues, you've got tampering issues, you've got varieties of cars. Cars are actually not very standardized and so just fitting a device into a vehicle is a big challenge. How do you do installation? The bit side is also hard. So we see great example is, you know, we're using face detection to understand whether there's a driver in the vehicle and if it's the same driver for that trip and to associate that driving record with a particular driver and to associate the feedback also. So you can see your record and what you do well and what you, you do poorly. And almost all of us have some good aspects of driving and some bad behaviors we should change. We used existing data sets of faces. Well, it turns out a lot of them are built out of TV content and TV content turns out to overrepresent certain white male kind of standard faces. So as we got into taxi drivers, they tend to be a very different ethnographic makeup and all of the existing face algorithms are deficient. So we're now training with a lot of real world data. So that's more a bytes example. Data example. But it's actually the same real world lesson of you've got to get into the real environment. So we ran for an entire year with our beta program before we really ever sold any devices. Just learning from real driving in the real world. And it ties back to the philosophical approach of learning from real human driving. But it's also an engineering reality of until you've seen a bunch of real experience, your stuff isn't going to be good. Right. And you've talked about if your MVP is perfect first time you launch it, you've launched it way too late. So we went through all that.

Yep. Learning curve and then accelerating through it. Exactly. Another thing that Natto has done interesting. Well, which is challenging, is building up good relations with the OEMs. So classic entrepreneurial advice tends to be. And kind of we're going through the contrarian parts of where it's contrarian. And right. As part of what Naoda is doing, his classic advice is, oh my God, the big companies, you can't work with them. They slow you down, they kill you. A metaphor that's sometimes used is in a breakfast, the chicken is interested, but the pig is committed. And it's kind of like, well, in the breakfast, that's the partnership between the startup and the large company. The startup is the pig and the large company is the chicken. And so it tends to end badly for the pig when the breakfast actually doesn't really deliver. And yet actually, in fact, part of the system's thinking, part of being pragmatic, Natto has built a set of thriving and productive OEM relationships. So talk a little bit about that.

There's a couple things coming together. One is just to be humble and to realize that cars are an incredibly complex machine that have been optimized since the 1880s and with billions of dollars and hundreds of thousands of engineers globally thinking about how to improve them. So it's naive at some level to assume that a dozen men and women in Silicon Valley can actually make any difference. Radically. Now, I think that's still the case because the system optimization opportunities, the opportunity to improve how these vehicles flow, how they coordinate, coming at it through a lens of collaboration and data. And optimization is a very different lens. But we knew from the beginning we didn't want to build cars. And so we're working with people whose entire expertise and profession is to build cars. And they're great at it. We work with BMW, Toyota, gm, they're all fantastic car builders. They have different positions in the market. Some are creating a unique luxury driving experience, some are creating a very comfortable family experience. Some want the thrill of the road. And so they have different backgrounds even within the same overall engineering discipline. We wanted to make sure we got access to the best of that expertise, but also brought in the software AI, deep learning thinking. And that's hard for a lot of automakers. They have 100 years of history of fantastic engine building, mechanical engineering, suspension design. How do I make this a really great ride? But they're not used to hiring fantastic software teams or algorithm engineers at the same level. So there's actually a match made in heaven. There's the other part goes back to having been a consultant. I worked with a lot of very large companies and you realize very quickly that there's a huge spectrum in every industry. There's some very high performing large companies and there's some laggard large companies. Silicon Valley mythology is all large companies are laggards. And I don't think that's true. I worked with some extremely amazing, very rapidly moving, turn on a dime large companies. Now they're rare. So if you picked an average large company, yes, you're talking about them moving slower than a startup. And certainly some elements of the organizations will always take more time, just like large startups move slower than small startups. But we went in through previous experience and through people that I knew at the board and senior executive level. It's interesting at the board level, a lot of these companies are thinking about the future of mobility, not just thinking about the cars they ship today. So whether that's the founder of Toyota, whether that's the Quant family at BMW or Bill Ford at Ford, they're all out there advocating for actually we have a social responsibility to make mobility great. We're not just a car company shipping cars. We're providing a mobility service. That mindset shift is fundamental. And I give the auto industry huge credit for beginning to think that way in a way that utilities, for example, haven't, or many other industries. But in the automobile industry, I see a huge change in the last few years. We deliberately picked partners that we thought were at the forefront of this, both technologically but also in terms of the mindset, in terms of the willingness to experiment and pioneer. And we've structured collaboration programs that are very open that way. We share the data with each of our customers. We don't try to say everything has to come to us and we have to control it. It's a Shared platform. And we got all five of our OEM partners excited about the idea that yes, each of you will see certain parts of America, certain parts of Europe, certain parts of Japan. But wouldn't it be better for everybody if the insights about traffic risk, dangerous spots, or even just risky maneuvers and what's good and what's bad behavior in different situations were shared? Because you don't want every carmaker to have to relearn all the same mistakes that every other carmaker has already learned to solve. That's kind of what we do today with teenagers, right? Everybody starts over, but you don't want to do that in the large company world. So they're all contributing data, of course, their algorithms, their patented technology stays private, the personal information of the driver stays private. But the abstract anonymized data patterns of how people behave, we pool and benefits everybody. I've had a great experience. It's not easy working with large companies, but I think it's a huge advantage if you can pull it off. It took a full year to get it to work, which is a long time in a startup. But now that we've got it up and running, it's a huge advantage. And I think for the long haul that autonomy really is. We will be much better off getting our technology into vehicles through this shared collaborative approach. At the same time, building the atoms and building our retrofit device allows us to innovate extremely rapidly. We do an over the air upgrade every couple weeks. We can rev hardware on our own schedule. So the duality of there's a built in car version and there's an aftermarket version that is very Silicon Valley consumer electronics, like we live in the best of both worlds in that case.

Awesome. When people tend to encounter new technologies like autonomous vehicles and AI, they tend to worry about risks first. So you get a whole bunch of stories about oh my God, is this going to take away jobs from drivers and robots will come for our cars, then they'll come for our jobs, and then maybe they'll come for our lives. Tends to be the Hollywood version of these stories. And they tend to miss the actual real human transformation that's possible here in the evolution of our societies. So why don't we first start with what do you think how cities change? Like what the mission in Naoto to say, if I can help this path, this is the way we lead more human lives.

I think it's really important to realize how deeply intertwined mobility in cities are. The cities that we know were actually created very recently all in the last 150 years. There were a few big cities before then in ancient Babylonia, Chang' an in China. But by and large they're all products of the post industrial revolution. It's fun to reflect. The cities used to be unpaved roads until bicycles came along. That paved the way for cars ultimately, of course, which take advantage of paved roads. But the initial transition was because of pollution. All the horses for horse drawn delivery of goods and horse drawn carriages created a huge mess. And there's a famous conference in New York in 1893 about how we will die in many feet of horse dung because there's a geometric expansion of ever more horses to remove the dung from the other horses. The automobile was the great environmental savior, it turns out, of that era. Now of course, we see the downside of that. We have internal combustion engines that cause a lot of pollution. We have the traffic fatalities in the U.S. it's the leading cause of death right up there with guns and well above infectious diseases until you get to be very old. Right now there are real pain points, congestion, finding, parking, not being able to get around fast enough, conveniently enough, and then equal access, good portion of the population can't afford the right kind of mobility. And in the us by and large, New York and a few cities accepted we haven't invested in the shared public infrastructure to provide great transportation. So I see autonomy as one element of that transformation. I've been very clear it's not the only element. And the data that we are able to extract about how transportation works is not just helping autonomy, but also helping sharing, also helping electrification. We're doing work with one of our OEM partners about how electric vehicles are used in their profiles and when they charge in studying that as part of our using our device to understand those patterns because that will help them design an electric vehicle experience that actually is better. Electric is important because you don't want the pollution right in the center of the city. And it also reduces noise and makes the system much more efficient. In terms of energy sharing is really important because just more autonomy can actually mean more vehicles, miles traveled, more sprawl. And I describe the scenario for people here in Silicon Valley. Imagine you're on 101 and it's total gridlock, not because there was an accident up ahead, but because the car next to you is delivering somebody's tie. From the dry cleaner to the right is an autonomous Amazon truck. Making package deliveries. In front of you is a vehicle that's coming back from dropping junior off at private school. We could live in a world where instead of 1.2 people per car, we're down to 0.4 people per car and you're stuck in congestion hell. So autonomy by itself deployed without thinking about how it helps the city, can actually make things worse. Another example is we think autonomy will eliminate all parking. And maybe that's true for parking structures. But if parking is expensive and driving is free, you could actually have people park their car by just having it circle around the block, which would be a total nightmare for city environments. So I think it's important for us to be thoughtful about how these technologies deploy and to understand the patterns of needs to engineer vehicles and these technologies in a way that they don't just come in as a luxury feature for the super rich. I like having a Tesla and I like the features on it, but I actually want those features to be available for somebody buying a VW Beetle or as a shared service for anybody, even a kid who can't drive yet. And I think that nudge to get it to being a clean, shared, autonomous system is really important. And we want our cities to embrace that innovation and to do it in a way where we're actually beginning to change our planning and our land use to reflect that as well. A lot of US cities have built airports that have poor connectivity to downtown because they're actually incented to have traditional car traffic. If you go to a lot of European cities, the train pulls right into the airport and you can get on. And people talk about autonomy taking over all mobility. I don't think that's true at all. And that's really important when we talk about cities because if you have a one or two mile trip, walking is still the best way to cover that. If you're doing a regular commute to school, to a job, and you're within the 3, 4, 5, even 9 mile range, actually an electric two wheeler scooter bicycle is by far the healthiest, by far the most fun and most efficient way to do that trip. And if you have a really busy corridor, even only autonomous vehicles, you're going to get to roughly eight times the current carrying capacity per lane. Today it's about 2,000 vehicles per lane per hour maximum. Before, before you get a traffic jam. With autonomy, you might be able to take that to 14, 15, 16,000 per lane per hour, which is huge. That's a four lane highway turning into a virtual 32 lane highway. But a great train system can actually be 40, 50, 60,000 trips people per lane per hour. So between 15 and 45. That's another step change, increase. Those are all technologies we have today. Trains, by the way, in many places are autonomous because they it's a more restricted environment. So I think the potential for transforming our cities, making them healthier, making them more livable, making them safer is huge. But it takes thoughtful deployment about how we use that. And that's why we see our data from Nauto actually also providing service to municipalities in terms of understanding what people are doing in terms of being able to improve the road infrastructure. Right now a lot of cities are on a rolling upgrade of every road in the state city. Instead of prioritizing the most dangerous spots, the most needed spots, the biggest repair locations. So exciting times.

The city obviously needs to be reconceptualized and needs to be conceptualized in a human centric way. The networks is a platform for that. Let's go to another concern that people have in autonomous vehicles, which is job creation and loss. And the obvious thing is people go, okay, the robot's coming from my car and therefore I'm going to lose my job because I'm a driver. Part of Naoto is to say, actually, in fact, we're trying to make humans much better drivers and trying to create a blended environment of autonomous vehicles and human drivers. So that's actually, in fact we're helping humans and making them safer and more productive and all the rest. What do you think of the general nature of the autonomous vehicle revolution and job creation?

It's a mixed story right now. On the one hand, you can see in the Data we have 4 million people whose job is called driver in the US today. It's the single largest job category, followed closely by retail and other service jobs. And so clearly autonomy is going to affect that population over time. At the same time, if you talk to anybody running commercial fleets today, there's an acute driver shortage. Delivery services, Uber and Lyft taxi companies are all battling for drivers. Because reality is it's not a wonderful job that's drawing, you know, the best and the brightest of the next generation to be a driver. It's a job that people use to bootstrap. It's to supplement your income while, you know, you're studying or going to night school or. I talked to many Uber drivers who, you know, they're opening a restaurant, but they're driving for Uber in the daytime because that's the only way they can make ends meet. And so it's a fill in job for many. It is a professional job at the high end. If you're really a commercial driver of trucks or delivery trucks, then it's a full career. And I think in that case, if you look at those roles, they will evolve more gradually. First of all, our approach to autonomous technology is really to use the sensing perception and the intelligence to augment the human driver. So in a way that's actually extending the life and capabilities of the human driver, not only for safety, but also into this sort of blended autonomous era. I think on a lot of vehicles there's still a role for humans for the other functions they do. A lot of trucks will eventually have an autonomous mode for the boring long haul interstates, but have a human driver for receiving and loading the contents, for maybe driving the last bit in the city where we won't feel as comfortable with large autonomous trucks early on. Airplanes are a good example of this. Right? There's actually no technical reason anymore for us to have pilots on airplanes. They can in fact fly the entire route autonomously on autopilot. But we all feel safer by having somebody there just in case. And I think there'll be a lot of that. Another great historical example is there were elevator people for about 20 years after automated elevators were invented. They greeted you, they asked you what floor you wanted to go to, they pushed the button for you, even though the elevator actually went there on its own. So I see the change actually happening much more that way. We might see autonomous taxis or autonomous lift vehicles that still have somebody in them because they will take on other roles. They will provide information, be a tour guide, you know, provide other entertainment or convenience functions. And it might even be specialized. You might get your hair cut while driving in the car. So when I wrote the Resource Revolution book, I looked back at the second phase of the Industrial Revolution, 1870-1920. And for a period of time there, we had both the Victorian era of lots of labor and increasing automation of factory and machines. And I think we're about to face another era like that. You'll have robot assistants, but you'll have high touch human service as well. Now, long term, people are right to be worried about driving as a category will become a smaller role. All the examples I've described, it's no longer technically a driving job for much of that trip. It's doing something else while being in a vehicle. And I think that's the most likely evolution. We'll see an actual increase in transportation demand, an increase in the number of services that are being provided while mobile. Right. And the fact that we all order from Amazon is an example of that. Far more things are coming to us. Won't necessarily be more jobs that are called driver. And so there's an adjustment period. But if you look at history, as long as that happens over 20, 30 years, we're actually used to those generational adjustments. Very few of us are still farmers, even though three generations ago half of us were farmers. I'm not going to advise my daughter to go into driving as her skill set, but she will definitely do something in her future career that involve a lot of mobility and a lot of being in vehicles, even if she may no longer be at the wheel. So I'm quite optimistic about us being able to make the transition, as long as we're thoughtful and we help people find new roles and we continue to create jobs. There's a broader version of this, though, that I want to acknowledge, because it's not just driving. If you look at AI and robots broadly, if we start to say, hey, I can replace reading CAT scans and I can replace legal document editing with AI, and I can replace the last leg of delivery for the robot from the truck to my house factory automation, over time, we will find that more and more job categories are not only augmented by AI, but some of them will be replaced. And so I don't know what that net rate's going to be. I certainly think the educational bar for the skills you need to have for the jobs will rise. And so that's a huge challenge for our educational system. And I'm a huge proponent of redirecting a lot of our investment there as a society to enable that. Driving is one piece of that story, but it's a much bigger story we urgently need to address. And so I don't want us to think about driving as a unique element. I want us to think about the broader reinvention we've made, the shift from agricultural to industrial society. We talk about being in the information society, but a lot of our school systems, college systems, haven't caught up, and a lot of our government systems haven't caught up with needing to sustain a different mix of industries of service roles.

Well, I think one of the things you're gesturing at there is today when people talk about driving school, they talk about a school by which you learn to drive. I think at some point in this future, what we will have is we'll have school that you're having. You're getting school while driving. While driving, while being driven. It's while you're being driven, you're actually learning new and other interesting things. And I actually think there Might be this driving university, which is the. Oh, as I get here and I move there, I'm actually picking up other skills and experiences and perhaps even creating productive work and all the rest.

In fact, if I had one piece of advice for the next generation of entrepreneurs, you know, we've all focused on how do I make driving safer, how do I make it better, how do I improve the vehicle? Or if you're really thinking ahead, how do I improve the system as we've talked about, right. Nauta is very much focused on that system improvement. But if you take driving changing to riding for granted, thinking much more creatively about what are you going to do while riding, the standard answers I find very banal. I want to watch a movie, you know, I'm going to do telephone calls, I'm going to. Okay, that's sort of what I'm doing in a car today. Let's get really creative. Can I start to have health care services? Can I start to have social interactions in these vehicles, particularly if we want them to be autonomous and shared? I could have really interesting coffees, lunches, social encounters, get my hair cut while moving somewhere. And we can think about that in both ways because now I go to someplace to get service in. The world of Amazon has showed us actually the stuff can come to me. And so what shows up at my house in a vehicle to do something temporarily for 10, 15, 20 minutes that I get into have that service done? You know, the dentist might be road mobile coming to your home, you get on board for your 15 minute dental checkup and then that moves on. From a societal point of view, massively efficient. Because instead of 20 people driving to one dentist office, you can actually now do an optimized route every day for that one mobile dental office. So I think we're only at the beginning of creativity. I talk about the next big Uber is actually the person who invents the way to move stuff around, matching it with people. Because the number one barrier for people giving up car ownership today, even though they'd like to, because it's expensive, inconvenient, hard to find parking, all that stuff is actually they use their car as storage. So some way to dynamically match rides, stuff and people, it's clearly a winner in the future.

Well, I clearly see that I will also be getting my manicure and my haircut en route to dinner as part of this future. It's been awesome talking with you. Naoto has this unique approach of human and vehicle, of making humans more powerful and capable, but also designing for a blended environment. You know, you could call you a philosopher builder, you could call you a thoughtful pragmatist. There's a set of these dualities, which the blended middle is actually the right entrepreneurial path to are the great leader of. And so with this, I will bring an end to this Gray Matter podcast, and I look forward to future conversations.

And I look forward to the rest of the journey, however it may be.

Thank you.