The Stack Overflow Podcast

Can AI solve car accidents and find you a parking space?

Episode Summary

In this episode, we chat with Eran Shir, CEO of Nexar, a company that provides smart dashcams. Using machine learning and the power of crowdsourcing, Nexar aims to help everyone drive smarter without sacrificing privacy.

Episode Notes

Graybeard conference alert! Eran and Ryan both started their technology journeys on the venerable Commodore 64

During his academic days, Eran helped to map all the BGP (background gateway protocol) gateways in the world. This got a fair bit of press recently during the six hour Facebook outage.

Nexar provides smart dashcams and an app that help cars understand the roads around them. 

While networked cameras on every car could be a privacy nightmare, Nexar says that they have privacy as a foundational part of the SDLC.

Episode Transcription

Eran Shir The edge is interesting for use cases that have to be solved at the edge, have to be solved in real time have to be solved by localized data that is not accessible in the cloud. In general, it's great for physical problems. And in particular, what we cared about is actually solving collisions, solving car accidents.

[intro music]

Ben Popper What’s costing your business time and money? Inaccurate address data. SmartyStreets provides blazing-fast address validation, autocompletion and rooftop geocoding APIs. StackOverflow listeners get a kick-awesome t-shirt in one of three designs when you start a free trial. Get yours at

BP Hello everybody. Welcome back to the Stack Overflow Podcast. I am Ben Popper, the Director of Content here at Stack Overflow. I am joined as I often am by my colleague and collaborator, Ryan Donovan. Hey, Ryan. 

Ryan Donovan Hey Ben, how you doing today?

BP I'm good. Thanks. So I have two young boys, they're seven and eight. And I've told them many times, like, I really hope when you're 18, you don't have to drive a car, like I hope that the cars drive you. It'll make it safer, I'll worry less. And they have like some vague knowledge of this and are very excited about it. And then they also know, we just like just least a relatively new car. And so it has the lane assist. So sometimes, you know, like, if I'm on a country road, it's easy, I'll take my hands off and like, show them that the car like you know, can stay in the lane on its own things like that. You know, where this comes from is just an increasingly large number of sensors, specifically cameras, but some cars also have radar. My car has radar and LIDAR all kinds of things in a car that can sense you know, map the world around them. And then onboard smarts, you know, that can decide what to do with that data. So today, we have a great guest coming on Eran Shir who is the cofounder and CEO of Nexar. Eran, welcome to the show.

ES Thanks for having me.

BP So this is the Stack Overflow Podcast. So usually we like to start out by just asking people, how did you get involved in the world of technology, computer science, programming, any of that stuff? Like take us back as far as you want to go? And let us know sort of what your journey has been in that world to hear.

ES Wow, now I need to really think how—

RD Search the memory banks, right? 

ES Yes. Well, my first computer was a Commodore 64. 

RD Me too.

ES Oh, cool! So that kind of dates us pretty well.

RD That's right.

ES So my first experience with software was to build games for Commodore 64 because I was like a elementary school and then middle school, and what you care about is games. So I remember distinctly, building a ninja game for the Commodore 64 and Commodore 128 with kind of sprites and reading and writing directly to memory and learning assembly and peek and poke and all kinds of things like that. And then, when when I turned 13, the Amiga came out. Now the Amiga was like this massive—it was like a science fiction computer compared to the 80s. It has a dedicated graphics chip. It has a dedicated audio chip. It was crazy. Unfortunately, it was too expensive. So my parents said, well, you can get a PC if you want. [Eran laughs]

BP Did they have that stuff because like they were looking at the gaming market? Like that's what people would buy. People wanting to play sweet Amiga games?

ES Yeah, in 3d. They had 3d games. And we're talking about like the 80s. It was crazy. But my real passion at the time was actually physics. So gaming was kind of a hobby, computers was kind of a hobby. But my real passion was physics. I studied physics all through kind of, basically since I was like in fourth, fifth grade, and then went out to do a first degree in physics. I'm Israeli, so I did my army service. And then I went back did my Master's in physics, and over that time, computers was sort of a something I enjoyed. I used as a non programmer programming for different research needs. So I was kind of doing Perl and all kinds of stuff like that. No one really knows how to read. And then in the early 2000s, I started doing startups. My first startup that was a company that tried to do a AI for search, like Google like search. I was sort of enamored by Google and searching in general, back in 98, 99, when they just started and so try to build a personalization engine for search. So you will get different results than I will get based on our different kind—

BP The filter bubble before the filter bubble was cool. 

ES Exactly. Obviously that wouldn't fly in, like 2000, 2001 didn't really work. But I learned a lot, went back to academia, doing a PhD in complex systems and complex networks. And there, I really dove into hardcore stuff because we were building the large internet observatory. So basically, I had 1000s and 10s of 1000s of people all over the world and install my software. And through that software, we could map the underlying network of the internet. All the routers and all the connections, what's called the autonomous systems, not the new autonomous, but the internet networks. If you ever heard of something called BGP.

RD It's very popular right now. 

ES It is. Yesterday, it's been all the rage, right? [Ryan laughs] We mapped my research project called Dines mapped all of the BGP in the world through crowdsourcing. And so I fell in love with crowdsourcing. At the time, it was like the superpower. All of a sudden, you know, a young guy in Tel Aviv can get 10s of 1000s of people all over the world, to install the software and to basically say things that the biggest companies in the world couldn't see.

BP Did you have like a fully global view? Like it would, you know, if some networks cordoned off, like a, you know, North Korea or where like, you know, some places wouldn't let you do it? Or like you really were able through user participation to get like a global view of sort of that internet backbone?

ES Yeah, I'm sure there were holes like North Korea, I can't say that we had a lot of coverage there. I'm not sure that there was a lot of internet in North Korea back in the early 2000s. But we could see pretty much everything because it was like a virus, right? Like we were going, whatever defences ISP is put, they put them looking outside, right? If you're a customer of some ISP, and you install my software, then I circumvented all of the defenses. I get all that information.

BP So the ISP is, did they hate you? Do they sue you? Yeah, this doesn't sound like something they'd be into.

ES No, they didn't. They we, you know, it was all for research, academia, writing papers, instead, etc. And it was all open, like the data was open, we gave it to the research community. And even after I left, the academia, it ran for like 10 years. We invented some really cool fun things. So we invented something I met at something called Tree Route, which basically use the fact that I have all of those agents to map unlawful routes, right? So basically trace wide world through what is approved through BGP to be the right route. But if I can coordinate, you two can coordinate so that when a packet comes to Ben, and he says, I'm actually Ryan, and I'm not Ben, you can actually trick the network to think something else. And then everything is available to you. Anyway, long story short, it was fun. Eventually, I realized that I enjoy startups more and have more impact. So I left academia and started a company called Dapper that was sort of trying to build a semantic web. This is like the web tool days, we allow you to build an API from any website in like five minutes. Doesn't matter if they have an API, or don't. You come in some machine learning some some clicks, and you got yourself a full fledged REST API that you can query from any any website in the world. And that turned out to be quite popular, turned out to be a foundation of some cool stuff, including dynamic advertising and widgets, and all kinds of stuff like that for a few years. And then about 10 years ago, I sold it to Yahoo!, spent a few years at Yahoo! building their next gen advertising platform, did my stand, spent a year with a VC and in 2015, I launched Nexar, together with my cofounder, Bruno who was actually my boss at Yahoo!

BP So yeah, let's poke around a little bit about why you chose Nexar, you know, what interests you in this space and how you and your cofounder, what sort of expertise you brought. I know a little bit about this world for my time at The Verge I used to cover some autonomous driving stuff, I did a piece on the history of LIDAR and I briefly worked at DJI which has lots of pretty advanced technology in the field. So you know, took, yeah, from a world of API's and dynamic advertising to a world of physical objects moving at high speed, trying to understand the world around them and perform safely. 

ES So we looked at the world in 2014 and early 2015, so about seven years ago. Looking at the world then we had one key insight. Intelligence is coming to the edge. So we knew How to do machine learning and deep learning at scale, at Yahoo! scale, which was pretty significant, and how to leverage it for all kinds of problems. It was the beginning of the deep learning revolution those years. And we were very excited about it. But the key insight that we had was that intelligence, deep learning is all kind of going to get to the edge eventually. And now, if you get to the edge, you ask yourself, why? Why it's important to be on the edge and not in the cloud in the cloud, I have infinite amount of compute, I have infinite amount of data, supposedly, why is the edge interesting? The edge is interesting for use cases that have to be solved at the edge, have to be solved in real time, have to be solved by localized data that is not accessible in the cloud. In general, it's great for physical problems. And in particular, what we cared about is actually solving collisions, solving car accidents. Back in 2014, and 2015. And still to this day, there was a conundrum, right? You mentioned your car has lane assist, and had all these kind of fancy features, right. And still, if you look at the statistics, from 2010 onwards, the number of fatalities keeps going. It went down for years and years until around 2010. And then it started going up, we're told that the cars are getting smarter and safer, etc. The number of collisions, the number of fatalities going up in the US, how come? We thought that the main reason why that's happening, aside from the obsession with SUVs and trucks, which obviously contributed significantly to it, was that we need to extend the time budget cars have and people have to react to the crazy, chaotic world that is out there. Right? The fundamental problem that we have in making cars safer, is that the budget time that you have as a driver is around a second or two, if you're lucky, because that's basically the distance that you maintained from the next car. And that is just not enough. It's not enough for you as a human, it's also not enough, and especially for computers. And so what can you do this is the physical that the truck is there, like two seconds ahead of you, there's nothing you can do, supposedly. Well, what we thought you can do is actually network all the cars together, right? So if we network all of the vehicles, with real time networking, with eyes that understand what they see, and can communicate that in real time across an area, across a region, all of a sudden, you are aware of something that's happening 100 yards away, you can react to it in beforehand. You know, it's like that sentence that a wise man doesn't fall into a ball that the smart men can get out of. That's basically what we envision. We've envisioned a network of vehicles with eyes open constantly, that calls the public space, calls the world like Google calls the web. And that's what we set out to do for Nexar. Build that network to unlock the value of the data for smart city, for mapping, for real time Driver Assist, for all these kinds of use cases. That's our trajectory. That was the vision then, it's still the vision very much today, only we managed to do a few things in them in between.

RD You know, you're talking about having cameras all over the cars and networking, all the cars. This is a big data machine you're constructing. How do you prevent this from being abused? I have heard of repo men going around scanning license plates to find repossessed cars. And then that data being used to track people by their cars and participate in things like union busting.

ES So the way that you could prevent your system for being abused is that you put privacy as a foundational part of the system. Right? So for example, we blur all license plates and blur all faces. We have a sort of a wall between your personal account where the raw data is and any what we call shared memory and insights that are aggregated and anomaly anonymized by the swarm of vehicles. We have a privacy pledge, we're doing a lot in that realm. And the reason why fundamentally is because, you know, we believe that there is a third path, third business model. So you know, typically we will told and we told others that if you don't pay for the product, you are the product right? That's what we learned in the last 20 years. And that was because it was the—I only had the widget selling business model, right, like Apple. Selling your product, you pay me good money. And then I don't need to resell your data. Or Google, were like I'm giving you services for free. And I'm gonna sell all your data, or the way and Facebook optimize that even further. But in the physical space, in the physical world, there is a third way, because you're seeing things that are super interesting, but are not private. Right? The fact that a traffic light stopped working is very important. But you wouldn't consider that as the private information. Right? You will say, oh my god, my camera updated someone that the traffic light on for thinking is stopped, stopped working? No, you think, oh, I contributed to the world, actually, this is good. And so there are lots of services, whether it's around Smart City, and mapping updates, and construction zones, and the potholes and I can go on and on, parking, etc, that you can build a great business on them without sacrificing anyone's privacy, anyone's private data. And that's what we set out to do. Believe me, I got lots of offers over the years to do just that, those applications have been repurposed for or even even more legitimate application, like giving people that double park fines. Eventually, you don't want people to double back they blocked the traffic, it actually makes sense, right? But I say I'm not playing that game. Lots of companies that can build that that's not my, what I'm interested in.

RD In order to blur faces and license plates doesn't the vision system have to understand what those are?

ES Yeah, we actually developed and trained, very efficient model, very efficient networks, to detect faces and detect license plates and actually do more. Like for example, we, we detect that on your windshield, there's a reflection, of a document that you put, your left on the dashboard, and at night, it reflects and gets into the view, we detect these kind of things, and we'll remove them, but we do all of that at the edge. So data goes up to any centralized cloud data set that you can attack.

BP Right, so the cameras doing its work, then there's some processing time on the edge work and make those decisions, and then it's going to share it with you. 

ES Yes. Now separately, the raw data is stored for you in your person in your Dropbox. Basically, you have a Nexar account, that is for you, it's like Nest or Wing, or anyone else that stores the videos that are of interest to you. And there they are. We don't own them. And we don't touch it, we don't even have access to it, other than in support situations. And we don't use it for any data application.

BP Is there going to be a tipping point in the near future where some of sort of our more ambitious goals for self driving cars or smart cities are going to come to fruition? Or, you know, does that stuff have just too much friction with local politics, federal regulation, etc, etc, to be achievable in the near term?

ES I will say, I don't think it's local politics. I think it's the technology, it's really hard. You know, I think that we are still very far away from solving the generic level five, kind of, I can drive wherever use case. The world is a super complex, look, we have collected over 25 million house events over the last four or five years. All of them are different. It's a super long tale that we have to solve. As an industry, it's non trivial. So you will see things popping up in limited regions in limited scenarios. And they will slowly very slowly expand highways, then maybe highways and a bit of the primary ways to do that. It's not going to be something that is going to happen overnight. Sadly, I think your kids in like eight years, also will still be needing a chauffeur or drive themselves to many, many, many destinations.

BP Maybe a different question to ask, which I'm equally interested in is, if the data were to already show and maybe it does, I haven't looked too closely, that being in a car that's fully autonomous at level three or four is as safe as being one that's driven by a person. Do you think we could start to change people's minds and say, look, it's the system is never going to be perfect, and it's not as near perfect as we want it. But it's already safer than many people driving around making, as you said many mistakes because their reaction time is impaired or just slower than what a top line machine can do today.

ES I think that the two issues here one is psychological, right, like you are willing to be killed by a Ryan, you're not willing to be killed by a robot.

BP I tell him that all the time. Yeah. 

ES So that is something that is just, you need to be aware of it, like people are not willing to be killed by a robot. So that's what, take it as you will. But that's the that's the situation. The second point is I think the issue is not really safety. And the reason why we don't have cruise cars everywhere, is not because of safety necessarily, is because of them getting stuck and don't know what to do. If imagine downtown San Francisco and all of a sudden someone puts a an orange cone on some lane. And that causes a chain reaction that gets a cruise car to be stuck. And then under it, there's a line of drivers kind of shouting and congestion. And all these are the hard problems that the companies haven't solved yet. And it will take time because the problem is truly complex. A lot of things can happen in different parts of the world. That's why it's not about convincing the people, because regulators can adjusting, I think yesterday, the California DMV approved the cause and way more to provide autonomous rides. It's because the people in the companies truly understand the complexity of the problem, right, and see how many times they need to engage. And they know that if they just launch it, it will create still lots of havoc. And that's why we need to continue kind of progressing gradually and not kind of do a one time kind of launch.

RD Yeah. And I think there's also a liability question, too, that for companies to launch this, they have to be very sure if their algorithms because if there's an accident, who does the insurance company go looking for money from? Do they go to the car manufacturer.

ES But I think that what happened in the last year that is very interesting, and very exciting for us is that the industry is starting to realize the power of crowdsourcing. The typical way in which AV companies went is we're gonna take them the meanest, baddest cow that we can engineer with lighters and cameras and everything, and we can put doesn't matter the price, we'll figure out the problem. And then we'll scale. We actually went a totally different way of saying we're gonna take the cheapest sensor we can get, and we'll put it everywhere, like hundreds of 1000s of people and crowd sourced all this data. And what's happening these days, is that the companies understand that they need orders of magnitude more data in order to finish the other 20%. You know, there is the 80% initially, and then there's 80%. Afterwards, for the other 80%, you need a lot more data. And that will come from cartels and things like that. And I think that's the thing. I'm hopeful for the industry because the penny has dropped and you see all of these initiatives and see also the coverage of these things. And OEMs getting into the picture.


BP Alright, everybody, it is that time of the show, I am going to shout out the winner of a lifeboat badge. Someone who came on Stack Overflow, found a question with a score of three or less gave an answer and got up to a score of 20 or more. Awarded September 27 to user engineer, "how to break out of the if statement." It's probably a deep one. I'm gonna have to put it in the show notes. I am Ben Popper. I'm the Director of Content here at Stack Overflow. You can always find me on Twitter @BenPopper. You can always email us with questions and suggestions podcast@StackOverflow. If you liked the show, leave a rating and a review. It really helps.

RD I'm Ryan Donovan, I edit the blog here at Stack Overflow. You can find me on Twitter @RThorDonovan. And if you have a great blog post idea, you can email me at

ES And I'm Eran Shir, cofounder and CEO of Nexar. You can find me on Twitter @EranShir. And if you want to learn more about Nexar, is the way to go.

BP Alright everybody, thanks for coming on and everybody who's listening, I'll talk to you soon.

[outro music]