Skip to main content

Meet the self-driving brains working with Volkswagen and Ford

Argo AI CEO Bryan Salesky on a new partnership with Lyft and setting realistic expectations for self-driving cars

Share this story

Photo Illustration by Grayson Blackmon / The Verge

Today I’m talking to Bryan Salesky, the co-founder and CEO of Argo AI, a startup that’s trying to build the tech stack for self-driving cars. Argo was founded in 2016, but there’s some solid foundations there: Bryan and his team were veterans of Google’s and Uber’s self-driving groups, and they have major investments from Ford and Volkswagen.

Bryan is actually one of my favorite people to talk to in the car industry — he’s a podcaster himself with a show called the No Parking Podcast, and he’s unusually direct for a tech CEO about what technology can and can’t do, and on what timeframe. He thinks self-driving cars will be “ready when [they’re] ready.” You might compare that to Elon Musk, who in 2019 promised a fleet of 1 million Tesla robotaxis on the streets by the end of 2020. 

It’s August 2021, and while that hasn’t happened, Argo has launched a small fleet of robotaxis in Miami and Austin in partnership with Lyft. There are still safety drivers in the cars, but it’s another small step toward self driving in the US — which, as Bryan points out, is actually trailing behind other countries.

I wanted to talk to Bryan about his partnership with Lyft, but I also wanted to know if the pandemic accelerated any of his investment or development. And, of course, I had to ask about 5G. Is 5G enabling any of Argo’s current self-driving technology? His answer might surprise you… unless you’re a regular listener of this show — then it won’t surprise you one bit.

The following transcript has been lightly edited for clarity.

Bryan Salesky, you’re the CEO of Argo AI, welcome to Decoder.

Thanks, good to be here.

I spoke to you in February of 2020 on The Vergecast, and then the whole world flipped over, but I’m glad to have you back on Decoder. I feel like maybe the frame of our conversation is a little different — Argo makes self-driving hardware and software for cars.

And I’m curious, this past year there’s been just a lot of interest in having computers do more for us. You see all the big tech companies are doing better than ever. Did the pandemic change the velocity or direction of self-driving?

For us it didn’t really change the direction. You’re right though, when I went to New York, I didn’t realize that would be like my last normal — pre-COVID — business trip. Actually, I was back to New York last week and it was neat to see the streets busy again. In fact, it seems like people don’t want to use the subway right now, so it seems like the surface streets are actually even busier than they had been.

Yeah.

But [the pandemic] did not really change the direction. I think there is more urgency to get this technology out there. The convenience and the ability to have a private ride — to have the car take you somewhere and not be with anybody else in the car — seems like that would be a pretty good value proposition right now as people are getting more comfortable with being around people again. 

I think on the goods delivery front we’ve seen a huge uptick in change in customer habits that we knew was happening, but it seems like COVID just really accelerated it, right? It’s kind of now an expectation among customers that there’s a delivery option for most retail businesses. We knew that was always the case. But it just seems like all of this has been accelerated.

Yeah, that’s why I wanted to start there. I have a lot of questions about Argo, and you have some news that we should discuss, but we’ve just seen so many technology products get uptake in a way that people predicted would take five years.

Zoom is the canonical example of the pandemic. Self-driving is one of those things where it’s connected to everything from the future of work — you’re gonna live farther away, and you’ll have a self-driving car, and it will tootle you into the city whenever you need to go to a city, but mostly you’re gonna telecommute and work, live wherever you wanna live — [to] package delivery, e-commerce.

The cost of goods delivery is a big deal. We’ve heard about this in the self-driving context for years and years and years. But it seems like you can’t make the products innovate any faster. Even though there’s all this pressure to get there, you’re still kind of on the same cadence as you were before?

Yeah, I think that’s true. We’re obviously working as quickly as we can because we know that this has real convenience, affordability, and safety benefits. But at the end of the day, it’s ready when it’s ready and there’s a process that has to be followed. It’s a safety-critical product as much as it is a fun and exciting product that folks will really want to use. I think it’s good to know that when we are ready and when we do start scaling out that there’s a real market for this.

Right now there are a lot of small businesses wondering, “Wait a second, how do I kinda get my piece of the e-commerce thing that’s happening?” And part of what they struggle with is, “How do I get my products from A to B, and to the customer’s doorstep as efficiently as some of the giant retail brands?” What’s cool about self-driving tech is that it has the promise to provide that A to B delivery efficiently, and give them a chance to compete.

And, nothing I say on this podcast will ever happen overnight, but I think this is where it’s going, and the fact that we can give small businesses more of an equal footing at some point, is interesting.

So let’s talk about Argo; how many people work there?

I think we’re approaching 1,400 now.

And how is that structured?

We have engineering offices here in the US, as well as in Munich, Germany. We have a fairly large workforce across our test cities that help us test and operate the vehicles, as well as a sizable product development engineering workforce that builds the technology.

You make hardware and software that allows cars to drive themselves. But I was just looking at your website, and your website’s really interesting. It has your podcast on it, your company values, the description of how Argo works, but it doesn’t name your technology. Does the Argo tech stack have a name that you use to describe itself?

We don’t have a name for it yet, and I think that’s some more of what you’ll see in the coming year as we start to get closer to launching. We do a lot of explaining around self-driving, and that’s what the podcast is for, that’s why we have groundtruthautonomy.com. These are properties to help folks that see our cars in cities today and they wanna know more: “Why are you here?” “What are you doing?” “What’s the benefit?”

That’s a lot of what we focus on at the moment. At the end of the day, what we are is a platform that enables lower-cost and safer mobility, and it has the potential to move not just people, but also goods. And, really, we have an extensive footprint, probably larger than any other AV developer at this point. We’re in seven cities and counting, and we’re really testing in some of the hardest areas in the country. Whether it be in the core of Miami, Austin, DC.

It’s been a great learning experience and it’s also given us an advantage. We have a lot of really important and valuable data in those cities now that helps us build the self-driving tech stack. Where it’s all going, though, is it’s a platform. And we’re working on a number of vehicles, with both Ford and Volkswagen, and [our platform] will enable those vehicles to do real work in these major cities across the world.

Ford and Volkswagen are obviously gigantic car companies. They’re good at things like alloy wheels and leather seats and all the other things that go into making cars. They make cars. You don’t make cars. What is the Argo product? Where does it begin and end?

The product is really, at its core, a whole lot of software that runs on some pretty specialized hardware, that connects to a car in a safe way. And I would say those car companies do a lot more than just make leather seats and alloy wheels — I don’t know if you’re setting me up here, but the car companies are increasingly becoming software companies in their own right. If you look at the car as a digital device, there’s actually an API, and a really important one that we interface with to be able to control basic things like steering and braking, and being able to do that in a safe and secure way is actually not trivial. 

So, heavy respect for what they do, and working in concert with the automaker makes sure that those interfaces are done right and in a secure and safe way.

So they put a specialized computer in their car, it runs your software. Do you have any hardware demands? Or is there a set of sensors that you require? Is there stuff that you make, or is it off-the-shelf? How does that part work?

It’s sort of an amalgamation of things. So, they certainly have computing that their control software operates on. We have something that almost looks like a mini data center in the car that’s able to process data from sensors that are positioned all around it. So the car is able to see through sensors that we make as well as buy — it’s able to see 360 degrees around it, 400 meters away, day, night, and is able to pick up on things that, I would venture to say, most human drivers don’t even necessarily see or notice.

“The advantage with self-driving tech is that our software stack can reason about literally thousands of objects at the same time.”

So, many times a second, the car is reading that information and making decisions about how to navigate through the street. People ask me all the time, “Well, how is it any different than how a human thinks about things?” Well, the difference is a human’s sort of picking the top two or three things that are relevant at the time. And if they make a mistake in that judgment, and they pick the wrong thing to focus on, or if they’re distracted, typically that’s when collisions happen, right?

The advantage with self-driving tech is that our software stack can reason about literally thousands of objects at the same time, and be tracking each individual bike, pedestrian, and car that’s in a busy surface street, and be able to extrapolate not just what are they doing now, but what are they going to be doing several seconds in the future. It doesn’t get tired, it doesn’t get distracted, it’s always learning and improving. And this is where the safety proposition comes from.

It’s funny that you use that set of phrases to describe a safety proposition. Because that is also the description of the Terminator. It doesn’t get tired, it won’t ever give up. Like, it’s the same.

I think that that movie is more about self-awareness, like being completely freely autonomous in the world, and that’s not what we’re talking about with Argo. These vehicles follow a set of very strict rules and are highly supervised, this is not like generalized, artificial intelligence. This is a highly skilled device for doing sort of one thing and one thing really, really well.

The last time we talked, you had this great line about how a self-driving car would perceive and react to the world around it. And you were like, “It turns out the rules are encoded all around us. It’s just traffic regulations. And there’s signs and roads and we just have to understand it, and the car will know how to drive in any city that it’s in.”

 I feel like it’s been a winding journey. Has that proven out? Is that still how you see it? Or do you need more and better structured map data about the world around you, to make self-driving happen?

Yeah, I don’t immediately remember what I said in that regard. You certainly need to know those rules, but our maps really encode a lot more than just what is encoded in a street sign. We make notes around what the behaviors you would expect at certain intersections. For example, at certain peak times it is socially acceptable, in the heart of a city, to do what’s called blocking the box, or basically pulling into the intersection in order to get all the way through. And if you don’t do that, you’re never going to get a turn.

“The customer is not just the rider in the car, the customer is actually the bike that’s riding next to us, the pedestrian that’s crossing in front of the vehicle.”

There are other places where, if you do that, you’re going to get honked at and run off the road. So we do encode certain contextual information so that we can act as naturally as possible with what the norms are in that local area. That is a painstaking process, but certainly something we’ve learned is important for people to accept it.

It’s one of these unique products in that the customer is not just the rider in the car, the customer is actually the bike that’s riding next to us, the pedestrian that’s crossing in front of the vehicle. The customer is really the environment around us in addition to whatever work the vehicle is doing at the time. It all matters, for a community to be okay with having self-driving cars.

It feels like we started 10 years ago with the idea that self-driving cars are going to change the world. And every company, and every executive I talk to working on the problem, somewhat naturally, and I think extremely understandably, developed a much more nuanced appreciation of the problem.

And ideas like, “All I need to do is OCR the local traffic manual — that a 15-year-old kid has to learn to get their driver’s license — and tell a computer to do it, and we’re off and running” turned into “Oh, there’s a bunch of norms in various cities that we need to encode in a semantic way, into the actual maps of the car.”

That’s a new kind of understanding, I think, for the industry. At the same time, it feels like the dream of self-driving cars has gotten much farther away.

I don’t know if that concept is new. I think something that we’ve understood over the last decade, is that it isn’t enough to just see and understand the world, you also have to predict what it’s going to do. If things are in motion, you have to operate in and among other human drivers. And you don’t want to be moving in an overly robotic way where the car is gonna be constantly starting and stopping, and moving left and right in ways that an external observer would say, “What the heck is this thing doing?”

I started doing robotics in like 2004 or 2005, in that neighborhood. And I remember one of the first assignments I had, building a simulator. Ok, so we were at the whiteboard and very little prior art existed, in the application we were in. And it was like, “All right, let’s start making a list of all the things that we’re gonna encounter in the world.” And I’m new to this industry. I’m staring at the whiteboard with a marker, and I’m looking at a guy that’s been doing this for 15 years, and I thought to myself, “Is this really the starting point?”

That’s hilarious.

So, yes. We have come a long way since those times. I think that the dream was always suspect if the dream was, “We’re gonna deploy millions of cars and go from literally zero self-driving cars on the street, to millions in a couple of years.” I mean, change doesn’t happen that quickly.

On the other hand, if you look at it from a basic compute and storage and silicon standpoint, we’ve made substantial progress in the 17 years I’ve been doing this. And I’ve now seen probably three major generations or inflection points of how we approach the autonomy problem in a fairly short period of time. So this all depends on your frame of reference, Nilay, and I guess, which headlines you’re reading and believing.

Yeah. I would say even in the past year we’ve seen a lot more solutions get deployed to consumers. They’re maybe level two autonomy, but they’re much more widespread. 

Sidewalk robots. My brother was on campus at University of Pittsburgh, he sent me a picture of a sidewalk robot that was delivering books. 

Yeah.

This was a bit earlier, right? And I mean, I think it’s cool that we see this now coming into everyday life. 

Yeah. I think DoorDash has robots in San Francisco, delivering food. So, we’re seeing it, but given your sweep and your experience, has this past year been as much of a jump as it felt like from the outside?

For me it’s not so much a jump, I think it’s just the natural evolution of things. But maybe I have a different perspective. I still stand by what I wrote a few years ago. Everybody said that I was “not bullish enough.” And I was just being a realist! I wrote that fully autonomous cars — cars that don’t require any operator — are going to start with modest deployments in cities and then scale up from there. And I think we’re seeing that across the industry. And, at some point we probably will look back and there will be an inflection point where we can say, “Well, that was the point where [autonomous vehicles] hockeysticked in terms of growth and scale.” We’re not there yet, and I still predict we’re a few years away from that.

Okay. One more sweeping question and then I want to talk about Argo itself. You are a unique CEO in that, obviously, you’re not deploying at scale tomorrow. We’re talking years, maybe decades. How do you make decisions operating on that timeline?

Well, for me, I focus on the customer. When you’re building something this complex and you’re sort of immersed in the tech day in and day out, I think it’s easy to deliver a solution that doesn’t quite meet what people need, right? So you have to really be customer-obsessed to steer and direct the company in the development.

So I look at it through a customer’s eyes. Is this the right thing to do? Is this meeting the promise that we made and the promises embedded in our core values? That’s why they’re on our website. Safety is number one.

So I use those values, and I try to put myself in the customer’s shoes in all the decisions we’re making to guide what we do and how we do it. I think we’ll build a stronger company and a stronger culture that way, and hopefully be around for a long time if we continue to follow that principle.

What’s the customer feedback loop here? Is it the Ford program manager for a car? You mentioned even the other people on the street. Who is the person saying, “This product isn’t good enough. Make it better.”?

Yeah, I mean, it’s the whole company. We all try it out. But it’s also partners, it’s investors, we have [given away] rides, even though they are not at scale yet. There’s a lot of feedback that comes in every day. We even get people that write to our company email addresses, “Hey, I saw your car. Does it do this? Does it do that?” There’s a lot of interest.

But, one of the things that we did, we announced recently, is a partnership with Lyft, where later this year, we’ll be offering rides on their network in Austin and Miami. I’m really looking forward to that, because that is going to be an opportunity to get true customer feedback that will be sort of real-time, if you will, as we’re delivering the rides. It’s gonna be interesting and something that we’re all looking forward to.

You’ve been testing in Miami for a few years. Was the Lyft partnership just a natural extension to actually go to consumers in Miami?

I think so. We looked at Miami several years ago as a really welcoming market to autonomy. The folks there are interested in technology, are interested in trying new things. It has good weather, which is a good place to start.

Well, it has some pretty bad weather too, right?

It has some bad weather. I will say, the bad weather we can handle. We can handle pretty heavy rain at this point. Our system was actually just operating in a really heavy rainstorm in Austin yesterday, and it was remarkable how well it worked. So, we’re cracking that, and eventually we’ll get to snow and other things.

But, yes, Miami has just been a fantastic city for us.

How many cars are you deploying in Miami with Lyft?

The number varies. Right now in any given month, we have anywhere from 30 to 40 vehicles.

If you’re a Lyft customer in Miami, and you want to try out an Argo car, can you request it specifically or is it luck of the draw?

So, not yet. So we just announced the partnership. We’ve got some integration to do to hook up our systems, but later this year, yeah, in some areas, you’ll have an option to pick an Argo AV.

And do those vehicles have safety drivers in them?

They will have safety drivers in them. You have an opportunity to ask questions about the tech and learn more about it.

Is it multiple people?

So we have two people in the car at all times today, yeah.

And are they Argo employees or Lyft employees? How does that relationship work?

They are Argo employees. And then, when they’re driving folks around on the Lyft network, they’re, I guess wearing two hats, if you will. They’re also Lyft drivers at that instant in time as well.

Can you describe the actual product? Are these cars that you have built, are they Lyft’s cars that you’ve retrofitted?

So these will be Ford Escape hybrid vehicles, which is a third or fourth generation of vehicle that we’re on at this point from Ford. It’s sort of an evolution of the vehicle. It has all the sensors and computing hardware that you need in order to drive fully self-driving on its own. We just are not taking the drivers out yet, because the product is still, frankly, being tested. We’re still adding new features and as the vehicle gets more and more capable, we work our way toward a state where we can start to remove the drivers.

What does “more capable” mean?

Just being able to handle various corner cases, whether it be emergency vehicles doing different maneuvers, or we recently added in fairly substantial updates that allow us to operate really well in areas where there’s heavy construction or unmapped signage. All of those things are things that need to be handled that go beyond basic driving functionality, and they are the realities of the real world. So we’re very much in the tail of that development.

You described cars as sophisticated digital devices with an API. Do these Escapes just come off the line? Do you figure it out on the dealer side? Do you have to haggle with the dealer? 

Yeah, Ford’s investment in us prevents me from having to haggle with a dealer, thankfully. No, these are, these are prototype vehicles that are—

So they’re special. 

They’re special, but they’re off the production line. They have modifications made to them; those modifications are heavily tested before the public would ever be offered a ride in them. So the public is seeing a very well-tested or robust system when they ultimately would get a ride in our vehicles.

I’m just curious where the line is between Ford and Argo. So Ford pulls them off and modifies them with a bunch of Ford stuff, and then you add the Argo computing and software?

So we co-design with Ford the modifications to the vehicle to accept the self-driving hardware. So we work together with them on the locations of sensors and what the best way is to power them, and how to take care of all the different potential failure modes. What if power goes out? What if brakes fail, what if— lots of what ifs. And we work with them to sort through those things, and make sure we have a safe product.

And then ultimately, that design gets vaulted and sent ultimately to production. And that’s where a factory will produce a base vehicle that looks very much similar to what would go to a dealer. Before that step, all the self-driving tech is applied sort of as an upfit, or an option on the vehicle. And then, ultimately they come to us for deployment.

So there’s a lot of brands here, right? There’s Ford, there’s Argo, there’s Lyft. Do you want the Lyft riders to know they’re in an Argo vehicle? Does that matter to you?

I think they want to know. They want to know “Who’s driving me?” Right? “How is that built?” “Tell me more about that company.” Those are natural questions. And we’ll be working to answer those as part of the partnership.

This initial phase is a relatively small number of cars. It’s less than a hundred across a few cities, and it’s an opportunity for, frankly, Lyft, Argo, and Ford to collectively learn about the experience we’re offering and to find out what customers want to see in future cycles of the product. And I think all three brands are going to learn a lot.

The reason I ask is, I think about the other pieces of a car. And some of them, car companies cannot help but label, right? The stereo in a car always has some brand that has been licensed, right? But it’s just a label.

Yeah.

And the label can actually change. And I always wonder if those stereo companies have anything to do with the stereo in a car beyond selling a label.

But then there’s stuff like standard cruise control. I have no idea who the supplier for the cruise control is in my current Ford. It’s just part of Ford’s system. Is self-driving going to be the stereo with a prominent brand? Or is it going to be an intermittent wiper control?

What you might be asking is, is it going to become commoditized? Right? Because you would say that wipers and basic cruise control and so on are commodities at this point. You work with a tier-one [manufacturer], you get the component, put it in the car, and it’s done.

Maybe sometime in the distant future, this technology is a commodity. I think there’s going to be very few companies that ultimately unlock the promise of full autonomy: A to B, no fixed routes, can literally take you anywhere within the confines of a city.

The self-driving software could be a brand in its own right

I think there’s only so many companies that will come up with that. I think there’ll be a lot of interest in how the tech is built. Like, what’s the team that was behind this. And I think the brand, the self-driving brand, if you will, or the platform is going to be something people will want to know. And, who knows, maybe in the future, there will be a certain driving style or DNA that people associate with. They could say, “Look, I like the Argo driver, I feel comfortable, I feel safe. I like the way it handled X, Y, and Z,” and they commit to it.

I don’t know, we’ll see where this goes. But the driver, whether it’s a computer or not, has a huge impact on the consumer’s experience. And so I think they’re gonna care.

That’s all in tension with what we talked about earlier. There are lots more level two, level two-plus driving systems on the market now. You said Ford has a big investment in Argo. Ford also has BlueCruise, which they’re going to roll out, and that can drive a Mustang Mach-E or an F-150 down a bunch of highways all by itself. Do you have any involvement in BlueCruise? Is that a different product? How does that interact with what you’re doing?

So, we don’t have involvement in BlueCruise or any of the driver assistance technology that you would maybe call level two or level two-plus, some people use that phrase now. We don’t have involvement in those products today.

On the commoditization point: do you think as brands like BlueCruise get more prevalent, and companies like Ford and Mercedes and VW and whoever else starts to market those features as brands, they will ultimately subsume full self-driving like Argo is doing?

I don’t know. It’s a good question. I get asked this all the time. And, the answer I’ve had for years now is that there really is no clear stepping stone from basic driver assistance to full-scale autonomy. You just need way more sensors, way more computing capability, and so on and so forth to achieve full autonomy where no human needs to be in the loop.

Over time, I think there will be a point where there’s a crossover, right? But it isn’t going to look like a stepping stone, it isn’t going to look like, a year or two from now we’re going to get there. To me there’s still huge advancements that would need to be made, for that to be practical. I mean, look, just even the cost of the technology today, regardless of your technical approach, is going to be pretty high for fully autonomous vehicles. To the point where it makes most sense to deploy those in captive fleets that are maintained, curated, cleaned, serviced, etc., and deployed on a sort of per mile, per trip basis.

At some point though, the volumes will go up in these fleets. Component costs will come down over time. I can’t predict when this will happen, but it will happen. And at some point, you’ll be able to offer this on personally owned vehicles. And at that point, have there been major technology changes where the sensor footprint of the car has decreased, or where it’s starting to blend more with what level two systems may have become at that point? I don’t know. I’m so far in the future though, Nilay, that I just put a bunch of dots on the screen for you. I have no idea how to connect them yet, time-wise.

Let me push on those a little bit. So you talked about fleets, that’s obviously taxi fleets, commercial fleets, that makes sense. You have a high-cost vehicle, you want to run them all the time to get your money back out of them. No average person needs to run their car that much, or uses their car for commercial reasons that much.

Let’s face it. Most people in cities today have no desire to drive or even own a car, because it’s so punitive cost-wise.

So yeah, I totally buy that. At the same time, Lyft is pursuing its own autonomy research. Uber raided your alma mater, Carnegie Mellon, for all of their autonomy researchers.

This was their promise, right? The ride-share companies were sort of explicitly founded on, “We’re going to build the distribution network, we’re going to get people using the system, we’re going to invest heavily in autonomy, we’re going to put the robot cars on the road, our costs are going to drop. And that’s how we’re going to become giant profitable companies.”

That just didn’t happen for them. They couldn’t get there. Uber has walked away from it. Lyft is obviously partnered with you now. Is that a data point for or against your position that it’s going to happen in fleets first?

I think it’s a data point that says that most companies that got into the self-driving game in circa 2015 or whatever it was, just had no idea how hard the problem was. And I think as talent moves across the industry, and folks get educated, they come around to what people like me who have been doing this for a while have always known, which is, this isn’t a light switch. It’s not going to happen overnight, and it’s a time- and capital-intensive thing. 

When I started Argo, in 2016, the fundamental sensors we needed didn’t really exist. At least not in any production capacity. We’re now finally getting there, because there has been an investment across the industry to get what we need. But, I never believed back then the companies that were saying, “Yeah, we’re going to deploy in 12 months or 18 months” — it was just nuts to think that.

But they were heavily incentivized, and they were spending lots of money. Why do you think they failed?

I don’t know. You have to ask them, right? And by the way, I don’t know that any one thing failed. I mean if you look at the partnering that’s happened, by and large, each of these teams has invested, built teams, and there’s been M & A — I mean, value was created in all of these deals. So I always hesitate to call things a failure. It may not have taken the turns that everyone thought it would, but, hey, it was new.

I spend a lot of time thinking about the 5G hype cycle, and how it didn’t pan out, and I was made to believe that there was some sort of race to 5G that was important. And now it’s here, and I have middling AT&T mid-band 5G on my phone. It’s fine.

One of the things we were absolutely promised about 5G was that it would be a key enabling technology for autonomous vehicles. And I have probably watched more cell-carrier marketing reels, where, like, the Wi-Fi beams were hitting silver cars in animations. It was just the promise. It was there. 5G is here now. It’s widely deployed. Has it made a meaningful difference to your roadmaps, to your capabilities?

No. So I think the value prop was, “Okay. All this expensive compute you have in the car, move it to the cloud, 5G is gonna be this super-resilient robust network that never fails, offers a huge amount of bandwidth and extremely low latency.” That was the promise. In reality, the latency is better. The bandwidth is improved.

“But, [5G] still is not an infrastructure you can count on for doing the basic decision-making.”

But, it still is not an infrastructure you can count on for doing the basic decision-making — like, where is the pedestrian in front of my car this instant? —  so I can steer and brake appropriately. I just think it was uninformed. It has its place as a value-added part of the ecosystem. It is not overnight and on its own going to fundamentally change the architecture and how we approach building a self-driving car for those reasons.

Do you think we need to build custom networks — vehicle-to-vehicle, vehicle-to-infrastructure — to enable self-driving?

I’m gonna put self-driving aside for a second. I don’t understand why we don’t have a set of standards that everybody has embraced, where a vehicle is prevented from running a red light. That’s like table stakes. Like, why can’t we get that done?

Okay, so my view [on custom networks] is yes. If you want to talk about some of the simplest applications, let’s start there, and get some traction on a standard, and get funding to deploy this as part of our infrastructure. Let’s start with things that everybody can understand and get in, and get around.

Now, long-term, for self-driving, that does have a benefit. … The way I look at it, it’s like we’re adding an additional sensor that provides additional information about what’s in and around the vehicle, that is always helpful.

Did you see the crazy test in Las Vegas, I think it was, where some company said they were gonna run self-driving cars on T-Mobile’s 5G network?

I don’t know that I saw that one.

Well, it just, at the end of the day, it turned out that they were drones. Like they just had a bunch of people like, in a data center, with joysticks driving the cars.

Okay.

Is that a viable path forward?

No, I don’t think that helps with the economics of the equation. Also, driving with any amount of latency remotely is really difficult. So there’s lots of ways you can try to mitigate that, but that’s not going to have the promise that we all want to see in terms of safety and efficiency.

Well, what was particularly funny about that is they announced, very proudly, this was T-Mobile 5G-powered cars. And T-Mobile was like, “We’re not really a part of this.” It was just one of those odd, strange, tech-demo announcements.

I missed this one. Okay, this is good. I got to look this up.

I’ll send it to you afterwards.

All right.

But what struck me about it was, we did hear all of this hype about 5G, specifically as it related to self-driving cars. And that was the only pure application of the network technology, to enable something that even remotely looked like 5G, and it doesn’t seem like it was a great idea.

It doesn’t sound like it, no. I’m not sure what the point of that was.

We’ve talked a lot about Ford. We’ve talked about Lyft. Your other big partner is Volkswagen. Describe that partnership and what kind of products you’re making with them. Are you building taxis with them in Europe?

The plan is to deploy in Europe. They’ve picked an initial city in Hamburg. We also have an office in Munich, so we’ve started some small testing in Munich. We’re launching on the ID.Buzz platform. They announced that but they’ve not shown pictures of it yet, so that will be forthcoming. I think it’s going to be an outstanding vehicle. Really consistent with their DNA, what they know how to do best, but also very unique, very flexible. I think we’ll have a platform that’s going to be fantastic for both moving people and goods. It’s an all-electric vehicle, which we’re really excited about. I think it’s going to be great.

We work with them very, very similar to how we work with Ford and what we discussed earlier: co-create, co-develop, really work very closely with them on the modification of that platform for autonomy. We’re definitely excited to get their cars on the road later this year.

It’s interesting you call that the ID.Buzz. Great name, by the way — the ID.Buzz is all electric. The last time we spoke, you actually mentioned that you didn’t think the propulsion system of the car was important. And the Ford Escape is a hybrid. It’s got a gas engine in it.

I’ve talked to other folks in the self-driving and related industries and they’re pretty adamant that it has to be an electrified car; you don’t want the complexity of the ICE (internal combustion engine) engines to do all this other stuff you’re doing.

Yeah. What we talked about before was a very specific issue around utilization and how much time the car sits at a charger, right? And the business dynamics around that. The car needs to be electrified, without question. Whether it’s hybrid or electric, you need that power for the computing and sensors that are on board the vehicle. That was never up for debate. Electric was always the pathway. but you’re right. I was definitely saying in the short term, if you were to build a business and deploy it in the short term. This is now a year and a half earlier. What I was looking at was that hybrids still had a lot of advantages because you can keep the car on the road for extended periods of time earning revenue.

I mean, at the end of the day, everybody has to make that calculation of, when do you think you’re ready to launch? And, what is the state of range, charging time, can you fast charge it without destroying and replacing batteries all the time? As battery technology improves at a rapid rate, I think all of that is certainly changing.

Do you think the individual cars you’re going to operate in Miami, or Austin, or Hamburg, are they going to be profitable individual cars? Or is this a data collection subsidy?

Boy, profitable. I mean, in the context of all the investment that’s been made to date, no, I mean, we’re looking at this certainly from a revenue perspective. Revenue is king. That’s important to get when you’re first starting out. “Profitable,” and “break even,” those are things that we’ll talk about a few years from now.

I understand the first car is the most expensive one. That’s got all your R & D costs in it. But just from the really dumb Excel spreadsheet math of, “This is how many dollars it costs to make this car and buy all the parts, and we’re going to run it for however long we’re going to run it in Miami, and pick up customers with Lyft.” Will that car pay for itself?

If that car lasts on the road long enough, yes. What exactly is that horizon? I’m not in a position to give those numbers, but it’s conceivable. I don’t know that it’s realistic in the very first generation.

Let me push on one of the other dots you threw up for us to connect. One of the things you said was it’s going to be fleets first because you can get the economics right. I think the pushback there is that the big tech companies that operate fleets are a little shakier [at rolling them out]. And I take your point that they did generate some value. The other thing you said was, eventually the cost will come down and it will hit consumer vehicles, and it will be small enough and cheap enough that you can just put it in a car that you can buy.

A few months ago, I interviewed Austin Russell, who is the CEO of Luminar. Smart guy. His point was, “We’re not doing robo-taxis. My whole goal is to drive that cost curve down, and make small sensors, and make a package, and the place you want this is in a consumer car.” And he was pretty adamant about it. What is the back-and-forth there? Is that sort of an active “industry is going to sort itself out” conversation or is that two companies that just aren’t in the same lane?

Yeah. I didn’t listen to Austin’s episode, so I don’t know all the context around it.

Decoder is the hottest podcast in town, man.

I have listened to many of your others.

What do you do in the back of these self-driving cars?

I didn’t happen to listen to Austin’s — you’re doing a great job, by the way. Your interviews are fantastic.

I think Austin has a product that he’s saying is going to be ideal for driver-assist applications, and more power to him. Sell into that. But the bottom line is that to do full autonomy, a car needs to see in all directions out to a really long range, especially to operate at all speeds. And so, I don’t know that the small sensor he was telling you was going to go into a driver-assistance application is really going to scale to what’s necessary for full self-driving. I just think they’re two completely different applications.

And, again, this gets to everybody trying to make guesses as to, “Okay, is there a crossover point where those tech stacks eventually converge?” I don’t see that happening for many years from now. And I guess history will see if I’m right or not.

Let’s unpack that a little bit for people who are listening, because I feel like I have an intuitive understanding of the difference between driver assistance and full self-driving. I’m not sure it’s rigorous, and I feel like the industry is doing its best to muddy the difference there.

Did you see Alex Roy’s article on this? The Roy’s Razor. He’s calling it Roy’s Razor. Did you see this?

No. What’s Roy’s Razor?

Roy’s Razor. He has a very simple test for “Is it full self-driving or not?” which is, “Can you sleep in it?” 

That’s a great test. But, I’ll just say it. People do dumb things in their Teslas, which are advertised to have a thing called Full Self-Driving or Autopilot. There’s a lot of confusion there, so what, in your mind, beyond just, “Can you sleep in it?” What is the difference? What is the line that you can evaluate?

Do you have to pay attention and keep your hands on the wheel or not? Is it hands off, mind off, or not? To me, that’s the fundamental question. And regardless of what marketing terms you use, at the end of the day, Tesla’s very clear that you need to be paying attention, have your hands on the wheel and be prepared to take over if the technology’s not going to do the right thing. It would appear people are free to use that product however they want, at their own risk. My view is that, regardless of what marketing terms we use, we’re never going to speak to the everyday consumer when we use terminology like L2, L2-plus. I mean, I listen to myself on this show this last hour: “L4,” “Well, Nilay, I’m not sure that L3…”

“Do you have to pay attention and keep your hands on the wheel or not? Is it hands off, mind off, or not? To me, that’s the fundamental question.”

I can talk all day long and people’s eyes are going to glaze over. My mother was asking me this question the other day, and her eyes glazed over. So, we need to come up with a framework to communicate “What is it, and what can it do and not do?” I remember hearing a story a while back about Volvo and some pedestrian-detection technology that they had. They were running through a test for a customer and the salesperson leapt out in front of the car and got run over.

Well, you know what? The car wasn’t equipped with it. It turned out it was an option and that car did not have that option. I don’t know if that story is accurate or not, but we clearly are not putting this information out there in an easy-to-use, easy-to-understand way. To me that’s a thing the whole industry needs to solve. And not just autonomy. I’m saying this is an automotive industry issue that we need to push on.

There is a big trust component to this, right? Take your hands off the steering wheel. Put your mind away. A lot of people are not comfortable with that. Speaking of mothers — I tried to get my mom to use the adaptive cruise control in her Mercedes and she just didn’t want to. She was like, “No. I don’t trust this car to stay away from the car in front of me.”

My mother doesn’t like to use the backup camera, Nilay. And I say, “But it’s so simple. It’s right there.” No, she’s using her mirrors. She’s just like, “Eh, it confuses me.” I’m like, “How is it confusing?” She says, “There’s lines on it.” I said, “ The lines help you understand where the steering wheel’s pointed.” She’s like, “No, no, no. That’s confusing.” And that tells me there’s a “feel’’ thing that I don’t really understand. But, yeah, this is education. It shows you that change is actually not that easy for a consumer that isn’t immersed in the tech day in and day out.

So this is the big question. Do you think Tesla rolling out Autopilot in beta, letting people do dumb things with it, constantly generating news cycles about it going sideways — there was a video the other day where it was getting confused by the moon in the sky. I don’t know if you saw it.

I saw it, yeah. It saw it as a yellow light...

Yeah, do you think that stuff is just poisoning the well? Is it making change harder?

Maybe. Here’s what I wonder. Out of the Autopilot accidents, the collisions that have occurred, I’d really be interested in those who can be interviewed to just say, “Did you know how the product was intended to be used and did you knowingly use it differently? Or did you knowingly ignore the instructions?” I wonder.

I’m going to make the counterpoint to what I was just saying. Maybe it actually isn’t so much an education thing and It goes back to the old adage with consumer products, which is: If there’s a way to abuse it, it will be abused. I don’t know. Maybe that’s what’s at play here. I think it’s worthy of discussion, though, because it’s clear that there is a problem and we need to get to the bottom of it. 

Maybe it’s all of the above. Maybe it’s education and it’s also, “Hey, you know what? Maybe a driver monitoring system’s not a bad thing.” Like if your car really does require a person to stay focused on the road, maybe you actually need to monitor and make sure that they are, for it to be engaged or stay engaged. There are conversations here that I think are really important to be having right now because at the end of the day, the core foundational case for all of this in my view is safety. There is a convenience element to it as well, but these are safety products at the end of the day.

The cars you’re deploying with Lyft, do they have driver monitoring for the safety drivers?

Yes.

Are those safety drivers eyes front, hands on the wheel?

Absolutely.

So they have to have their hands on the wheel?

They need to be hovering their hands over the wheel. They need to have their eyes focused and trained on the road or in their mirrors. They need to be scanning.

So those are rules that I’m assuming you and Lyft have written together about how safe you want the cars to be. Another way of writing those rules is for governments, and cities, and states, and countries to write a bunch of regulations about self-driving cars or what they want safety drivers to be doing. That’s also connected to public perception and fear of change.

The last time we spoke, I think you said, “We need a unified way of thinking about this so this actually becomes a market.” I feel like, again, a lot has changed in the past year and a half. Have the regulators gotten antsier in your view? Have they gotten more open to it? What does that look like?

I think they’re asking a lot of questions, and they’ve been asking a lot of questions over the years. I think at some point we need to convene because it isn’t trackable, nor is it a good thing for the consumer or for the companies building [self-driving cars] to have 50 different rules across 50 different states. Plus, then city-specific guidelines. It just doesn’t work. It makes a really hard, complex problem even harder, so I hope that we’re able to come together. I’m optimistic. I think the administration has said positive things about self-driving and acknowledging the benefits that it has in terms of safer, cleaner transportation. But now we have to take that and use it as inspiration for getting a framework that works.

What do you think that framework looks like?

Well, I think we all need to convene and have a good discussion around it. There was a bill that was, I forget the exact stage in the process that it was in, but a bill that was close. I think it was part of the Endless Frontier Act. It at least provides for a deployment pathway and increases the cap on the number of vehicles that any one company can deploy. It also clarified a number of things. That’s a start. I mean, it at least acknowledges, at the federal level, that self-driving is something that we support and we want to move forward with.

Right now, I think we’re just stalled, and I don’t think it’s good for our country, especially when you look at Europe, you look at Germany that’s now passing regulation not just for the testing of cars but also the commercial deployment. You see what China is doing. We are falling behind, and it’s not good.

You obviously work with two big partners in Ford and Volkswagen. We see a bunch of change in the car industry. There’s new car companies spinning up that are very exciting. There’s all kinds of investment activity. There’s also a lot of consolidation. As you look across that landscape, who do you think your truest competitors are?

Well, I always leave those things to fellows like you to call out for me.

But I have to ask you.

But you follow this really closely. You have an answer. I think we’re in such an exciting time where a lot of good ideas are getting funded. I saw a chart in the Wall Street Journal the other day that showed just enormous growth in startup capital. I think it’s a really exciting time. And the way I look at it is, this technology, whoever gets it fielded, and deployed, whoever makes it at the end of the day, as long as they’re following the right practices and processes, and view it not just as a convenience but also a safety function, I think is going to have a huge transformative impact on the world. And it’s going to mean safer, cleaner, easier transportation worldwide.

That’s what we’re after. And regardless of who my competitors are or who’s going to end up deploying it, I think those that stay true to that vision and deliver on it, I think are going to be incredible, long-lasting companies.

Well, one of the reasons I ask you comes back to that commodification question. Is there just a first-mover advantage — where, if you happen to get to the finish line first, everyone else is going to just start buying Argo’s stack, and deploying it? Many markets are winner-take-all. Do you think self driving is winner-take-all?

No. I don’t think it’s winner-take-all. There are 3 trillion miles driven in the US alone. A huge percentage of those are actually passenger miles. So I think that the opportunity is enormous. I think you’re gonna see different companies specialize in automating different fractions and slices of those miles. And I think there’s incredible businesses to be built, even just within those slices.

One way you can measure this competition — I don’t know if I love it, but you’re just going to tell me it’s my job to figure it out, but we’re gonna try anyway — one way to measure this competition is miles driven. How many miles did our robots drive this week?

And the reason I bring it up is because Waymo, which I think is a competitor of yours, loves to tout how many miles it’s driven. It’s billions of more miles than anyone can even comprehend, Waymo says. Is that the right measure? Is that a measure you use?

No, it isn’t. Depending on where you operate, you’re driving at different speeds. And so the unit of work for any given vehicle on any given day is very different, right? I could put up a huge mileage number, and have it all be on the highway, and that’s very different from driving Ocean Drive in Miami. So not all miles are the same. It’s really an apples to oranges measurement.

I do think that being clear about where you’re driving and what the complexity is and what the industry terms the operational design domain — which is just describing the types of streets and roads you’re driving on — is really important to disclose just as much as mileage, if you’re gonna use mileage as your benchmark.

Mileage is obviously a proxy for data collected. The sensors are reading a bunch of data, they’re running machine-learning algorithms. They need a vast amount of data to make those algorithms smarter. Is data collected the right measure?

I mean, data’s useful, but again, data on I-75 is gonna be very different than data on Broadway. So, I just think that, again, it’s all about the environment, and if we want to start looking at the diversity of data, maybe that’s a way to look at it. What is the diversity in the different environments and cases that you’ve seen and recorded, and are testing against? And how well is the system performing against those test cases? That’s ultimately what autonomous driving companies are looking at internally. It may not get published, but that’s what we’re looking at.

Whenever we talk about data collection, we have to talk about privacy. We’ll use the Lyft car as a tiny example. Lyft knows who you are. You sign into Lyft. You call the car. It maps a route. Lyft can kind of throw that away after a while. They’re running a logistics business in some way, and they want to know where their cars need to be. But for a giant machine-learning dataset, you need to hold on to that data.

Are you anonymizing? Is that something you’ve had to think about?

It’s totally anonymized for us. We aren’t trying to identify anybody. We’re purely using the data to understand what we’re seeing: is it a vehicle? Is it a person? Is it a bike, a scooter, is the pedestrian motorized, as we’ve called it.

We’re trying to understand that so that we can predict and understand, what are the movements you’re gonna make to ensure safety? We are not handling the PII (personal identifiable information) in that engagement.

You don’t make the sort of user-facing interface of the car, right? You’re very focused on autonomy, tying into the car’s existing systems. I’m sure with Lyft there’s gonna be some consumer software interface element. But one of the things I’ve noticed, as I’ve talked to car company CEOs — the CEO of Ford, the CEO of Polestar — they’re aware that they can’t just give the center stack of the car away to CarPlay and Android Auto anymore. It’s got to be good, and they’ve got to win that back, because as cars start to drive themselves, they can’t just be like, it’s CarPlay. They need people to use their maps, their interface, to understand how the cars operate.

Where in that dynamic do you see Argo playing? Is it all in the background? Do you want an Argo interface? How does that work out?

Yeah, it’s a little bit in the background. The part that we concentrate on is, we want to provide displays that assure the riders that the car is operating safely and correctly. So what we do have is a display that we build at Argo that shows the scene in and around the vehicle, and the location of the different objects around the vehicle. We’ll show the upcoming trip and routes, and the next turn it’s going to make, that kind of thing.

In terms of the dashboard for a consumer vehicle, and getting into sort of the map wars and the different companies screaming for your attention and screen time, that’s not a battle that we’re in the middle of right now.

But you need to have maps to run your system. Who’s your map provider?

We do need maps, but they aren’t the types of maps that you would give to a consumer. They’re maps that help the car drive. They’re kind of high-resolution maps that help the vehicle figure out what location it’s in, like exactly where it is within a lane and so on. It isn’t navigational. I mean, that’s an element to it, right. But that’s a little bit of a different use case.

I don’t think I understand. So if I get into one of these Lyfts in Austin or Miami, or one of these ID.Buzzes — it’s still a great name — in Hamburg, and I want to go to a bar. There has to be some sort of consumer-facing map that’s showing me things like, here’s how long it will take, here are the roads we’re going to take. That has to get translated into instructions for the car to operate.

Yes.

Does that happen based on the map that I’m seeing, or a different map?

You picking where you want to go, and seeing your current location, and trying to find points of interest and all of that; at least with the Lyft engagement, that will happen through the Lyft app with the Lyft map. That will be the Lyft system. Ultimately trip data gets sent to us. And yes, we do have an understanding of addresses and locations, and how to navigate the vehicle.

My point is that those systems, though, are kind of optimized for the driving function. The actual UI presented to the user is more on the Lyft end of things. We don’t have involvement in that part of it.

As you think about making more consumer vehicles or think farther into the future, do you think about partnerships? Ford has a big partnership with Google now. Volvo uses Android Auto. All these companies are kind of headed in that direction. Is that a partnership that you foresee needing to navigate? Is that just something happening over there, and you’re the provider of a capability?

I think we’ll see how all this plays out. I mean, you’re pushing on an issue that is sort of big-picture. The interior real estate of the vehicle is a huge opportunity. And we’ll see where we want to play in that space. But your point is totally valid, right. It is extremely valuable real estate. Especially when you’re no longer driving, and can be staring at a screen.

Do you think we are all going to be facing ads in the interior of our car while we’re not driving?

I certainly hope that however ads enter, that it’s done tastefully.

Every year at CES I get pitched on “Come see the future of ads for self-driving cars.” I don’t go to those meetings, but I know they’re happening, and they terrify me.

Yeah. I think Minority Report did a great job of showing what the future could be, and I’m not sure that’s what we want.

All right, last question. It’s a total toss-up, because I don’t think anybody really knows. What do you expect, if and when Apple enters this game?

The iCar?

Yeah.

I don’t know. We’ll see what they come up with and what they’re working on. It seems like it’s pivoted a lot, just from what I see in the press, I don’t have any privileged information. It just seems like they’ve pivoted a few times and I’m not sure what they’re doing.

I am an Apple user. I revere their design philosophy and I think it’s really incredible, the products they put together. If they made a custom Apple car I would want to use it and try it. I’m sure it would be nothing short of amazing from an industrial design perspective. So that alone, I would nerd out on it for sure.

I’m not sure that I would trust a car driven by Siri.

I’m not sure Siri’s driving.

Fair. Fair enough.

I think Siri is there to be your personal assistant when you’re in the car. I don’t know that Siri’s doing the driving.

I certainly hope not.

I hope not.

All right, what’s next for Argo, and what should people expect?

Well, what’s next are rides in Austin and Miami with Lyft, and I’m excited about that, and looking forward to deploying here in the next couple of years. It’s gonna be exciting.

Awesome, man. Well, Bryan, thank you so much for being on Decoder. As always, it’s just a great conversation.

Awesome, thank you.

Decoder with Nilay Patel /

A podcast from The Verge about big ideas and other problems,

Subscribe now!