SPEAKER_00: Season 2 of The Next Unicorns is brought to you by Dell for Entrepreneurs. Save up to 43% at launch.co slash dell. And while you're there, register for a free IT consultation and be entered to win a $200 Amazon gift card. LinkedIn Jobs. A business is only as strong as its people. And every hire matters. Get $50 off your first job post at linkedin.com slash unicorn. And Embroker. The Embroker Startup Insurance Program helps startups secure the most important lines of insurance at a lower cost and with less hassle. Save up to 20% off of traditional insurance today at embroker.com slash twist. And while you're there, get an extra 10% off using offer code twist.

SPEAKER_02: Hey everybody, welcome back to The Next Unicorns series.

SPEAKER_04: This is our eighth in a series of 10 companies that we think have unicorn potential. What does unicorn potential mean? It's just an arbitrary thing here in Silicon Valley. When a company becomes worth a billion dollars, people tend to think, well, that's a milestone. And I guess it is. It's an arbitrary one, just like trying to pick the next unicorns is an arbitrary one. But sometimes forcing yourself to pick the next 10 requires a process. And our editorial team and researchers went to work looking at what are the signals for a company to be a large, sustainable, humanity-changing company. And we try to challenge ourselves every year to do this series. And we're eight episodes in. And today's going to be fantastic. We're going to talk a little bit about self-driving and autonomy, which we've been talking about for years on this podcast. And it seems like we're slowly getting there. If you want to look back at the other episodes, which started around episode 1089. Yes, we've done 1,000 episodes of this podcast. Daphne, the co-founder of Coursera, was on talking about InCitro, which does drug discovery by machine learning. That was really great. Nikki Peckett was on talking about homebound and building new homes from the bottom up. We had a bunch of other episodes. One that came up that people really seemed to like was Cody Friesen from Zero Mass Water. And they do those hydro panels you put on your roof and magically two cases of bottled water appear. Not literally, but that amount of water appears. And that was pretty inspiring for folks. And if you would like to have your CEO or founder on this podcast, you can have your PR person email. Don't email us at This Week in Startups. That's again, don't email us at This Week in Startups. And we'll consider it now. We don't consider any pitches from PR people or other firms. We do listen to the fans of the show, and they at mention us on Twitter and suggest and CC people to be on the podcast. But generally, we don't take pitches. Thank you, though. We do appreciate it. So today on the program is John Hayes from Ghost Locomotion, which they just go by Ghost. Welcome to the podcast, John. Hey, glad to be here. Thanks for doing this. I know we're in the middle of a pandemic. And here in California, we're having wildfires. May I ask, where are you and are you safe?

SPEAKER_13: Well, I'm at home right now, so I'm feeling pretty safe. Great. And so is most of our staff. They're at home too.

SPEAKER_15: Where's safe? Are you in Los Angeles, Arizona?

SPEAKER_13: Oh, no.

SPEAKER_16: We're in Mountain View, or I'm in Mountain View. Our company's Mountain View. Most everyone else is within a few miles of there.

SPEAKER_15: How's the smoke down there? I'm mid-peninsula. So how's the smoke down there?

SPEAKER_17: It's much better than two days ago. So now it's only slightly orange and a bit foggy, but it's quite breathable now.

SPEAKER_20: It is a crazy moment in time. How is your team handling this?

SPEAKER_04: What I would say is just an incredible run of, I think we're all having, humanity in 2020

SPEAKER_22: between bad luck, the pandemic. Obviously, we had some of the race protests and racism protests. I think it's an accurate description of them.

SPEAKER_04: The election is causing a lot of people anxiety, work from home, shelter in place, and the economy. I mean, these things are now these fires. People are, I don't know if you're having this experience in your circle or your company. Some people are cracking a little bit under this six months of just bad news. And let's face it, staying at home and being isolated and then having to stay indoors now with this, you know, 10 days of really unbelievably horrible air quality. It's breaking people. Are people holding up okay over there? And are you dealing with that as the chief psychologist officer at your startup by default?

SPEAKER_13: But I have help on the psychology front. People are handling remarkably well.

SPEAKER_17: And I think that because we have a huge software focus in our company, people have been mainly able to just keep doing their jobs. And, you know, it varies for a lot of people. It's like working on a great project like this is a huge positive focus for them. And the team is kind of split up. Like there's some people who have to go in because we do do some hardware development. We do some operations. But overall, it's like I went into this. I'm personally not a fan of remote work. And previously, we basically had a policy that everyone has to be in the office. And we actually have turned on people who wanted to be remote maybe a couple of days a week. And so the performance of the team has just been incredibly positive.

SPEAKER_29: So I mean, this kind of changed my mind.

SPEAKER_30: Yeah. So I think we're both old school and we like people to come to work if they have a job.

SPEAKER_32: Well, it's not just that. It's that this is an intensive engineering project where you're closely collaborating.

SPEAKER_17: There's a lot of like really hard problems and you need to talk to the people around you. It's not like we can just write out stories and send them off to anyone to develop.

SPEAKER_37: Right.

SPEAKER_20: So now some number of people, I guess the software team might be working from home and the hardware team needs to be in the office. You have to take certain precautions in the office. How is that all gone? I'm curious. And then we'll get into sort of the mission of Ghost.

SPEAKER_41: And what you're doing.

SPEAKER_13: That's kind of okay. I mean, part of it is that there's enough people out of the office that it's pretty sparse.

SPEAKER_17: But when you're working on hardware, you have a bunch of equipment and you have labs. And of course, we're building stuff into cars. So we have a car lab.

SPEAKER_13: And so, you know, you take precautions, you take breathing precautions, you check, but it hasn't really changed the actual work. And I think that, you know, almost everything has gone on.

SPEAKER_17: For some self-driving companies that do a lot of driving, they had to suspend a huge part of their operations. And that didn't hit us. And to some degree, I think that it hit us at a lucky time. If we had been at an earlier stage where we were, like, say, raising money, say, in March or April, we would have been in trouble.

SPEAKER_13: If we had been at a later stage where we are, like, just beginning scaling up consumer operations, that also would have been a huge problem. But we're just in this, like, really concentrated R&D stage where, like, most of our staff is engineers. And so engineers just keep engineering.

SPEAKER_37: So when you're in the laboratory phase, it's actually heads down and actually not a bad time to, yeah, have never a good time to have a pandemic.

SPEAKER_45: Yeah, it's not a bad time to, like, hunker down and just do some hard work and get it done.

SPEAKER_37: So the company, Ghost, is building add-on hardware that you can put onto any car to make it self-driving.

SPEAKER_22: That's correct?

SPEAKER_49: That's correct.

SPEAKER_22: Now, why take that approach as opposed to building from the bottom up as Tesla is doing as Waymail was adding equipment, but they are, I believe they were also considering, and they had built some test models of their own cars.

SPEAKER_04: How does one come to the, or how did you come to the conclusion that the best thing to do in 2017, I think when you started, was to add on to existing fleets?

SPEAKER_12: So, in 2017, we looked at the AV industry, and there was a bunch of companies, and one

SPEAKER_17: of the things we saw was a lot of them were doing the same strategy. This was a very robotics-oriented strategy where you put a quarter million to a half million dollars of equipment on a car, and then you have a fleet, and you drive them around to

SPEAKER_13: test, essentially test your system. And one of the insights we have, I don't have a background in robotics. My background is in software, both enterprise software and consumer software. My co-founder, his background is in verified systems and machine learning.

SPEAKER_17: But we looked at AV and said, what if it's not a robotics problem? And if you look at like California accident reports, like you can read them because every

SPEAKER_13: single collision is there. One of the things that you see is when you, when you take sort of an engineering eye, when you look at them, you realize that all the problems are caused by software. No one actually has any hardware problems in this space, and a lot of the equipment that they're using is 10 plus years old.

SPEAKER_53: Fascinating. So you looked at what I would call maybe the fourth or fifth inning of self-driving, you know,

SPEAKER_22: where we're like sort of, I would, I think we'd all agree we're more than halfway there. You looked at this halfway mark and said, wait a second. If I look at Waymo and Tesla's crash reports and Uber self-driving unit, every single time there's a problem, it has not been a failure of LiDAR or the camera or the other sensors. It has been a software problem. So we should relentlessly focus on the software part.

SPEAKER_20: So then what is the business you're going to be in eventually? Are you building this enabling technology to sell it to other self-driving companies and have components of the software solved? Or are you going to sell it to car companies? Or are you going to sell it to consumers and let them retrofit their cars? What is the business model here?

SPEAKER_13: So the first business model is to actually retrofit consumers' cars. Wow. So you come in with a car and it's, it's kind of like getting a stereo installed, like back in the day, where we put down wires.

SPEAKER_61: Benzibox, what do they call those? The ones that slide out? Benzibox?

SPEAKER_13: Yeah, we actually, we put a computer in the trunk where you used to put a CD changer. Perfect. We put cameras in the car, they're, they're glued onto the car onto the front on the sides. And it's all very small. It's connected to the controls of the car. And so that we can transform someone's car. Wow.

SPEAKER_54: With essentially consumer equipment.

SPEAKER_20: And is there a specific type of car like that you've white labeled and said, hey, we're going to do the Honda Accord, the Civic,

SPEAKER_22: the Prius because those are so plentiful out there and that's our target market. Or are you looking at more high end cars and saying, hey,

SPEAKER_41: this is going to be a $10,000 or $5,000 option. We should put it on these other cars. Answer that question when we get back on This Week in Startups.

SPEAKER_65: Hey, everybody. You know, Dell has been sponsoring This Week in Startups and they've been a tremendous supporter of me for many years. And I have been a tremendous supporter of Dell's. Long story short, Dell for Entrepreneurs has really been trying to help every single one of our startups. And we're very lucky today to have Mobilaji Sokumbi on the program and he runs Dell for Entrepreneurs.

SPEAKER_68: Yeah, Jason. Yeah. Thanks. Thanks for having me. As you said, name is Mobilaji Sokumbi. I oversee strategic partnerships and the Center for Entrepreneurship for Dell Small Business in the US.

SPEAKER_65: What do entrepreneurs tell you they need and what do you provide for them?

SPEAKER_69: With the current partnership we have right now, Dell Small Business in the States, we have 500 IT advisors, startup IT advisors dedicated to our program. Half of those are based in Nashville, Tennessee, our second largest physical location in the US. And the other half is based here in Round Rock, Texas, where we have our global headquarters. And the conversation with entrepreneurs is more trying to get an understanding regardless of where they are on that entrepreneurial journey, right? Whether they're just starting off or they're about to scale, we simply want to have a conversation with them, you know, to figure out what their needs are. Okay. That team of 500 tech advisors has conversations with our startups and entrepreneurs around IT consultation. We're having conversations with them about how they can accelerate their technology, how they can partner with Dell, you know, to make sure that they can save money in the short term and the long term. We're having a conversation with them about having, making sure like, you know, they can use the capital they have towards driving the business. And that way they don't have to worry about burning it, burning to their cash. We also want to make sure we provide rewards to them

SPEAKER_71: so that they can put something back in their pocket. Save up to 43% when you take an extra 5% off at launch.co.dell. And while you're there, you can also register for a free IT consultation and be entered in to win a $200 Amazon gift card. Hey, everybody. Welcome back to this week in startups. My guest

SPEAKER_20: today, John Hayes. He is Ghost Hayes on the Twitter, G-H-O-S-T-H-A-Y-E-S. And he is the CEO and co-founder of Ghost Locomotion or just known as Ghost. Their website is gh.st. And when we left our hero, that's you, John, we were talking about, um, maybe the cost of this to put on a car

SPEAKER_22: and then what cars this will be available for, because am I right that this is going to have to be adjusted on a per car basis or are cars similar enough that you could pretty much slap it on any

SPEAKER_17: car? Well, cars are pretty similar. We do have to do some per car engineering because you can think of cars as like there's an American platform or German platform or Japanese platform. And so that's sort of the basis of how we integrate into the cars. And then they have a pretty narrow range of things like, um, say front windshield angles, uh, you know, various angles around the car. It's all pretty narrow because cars are built to the same safety standards, same aerodynamic standards.

SPEAKER_13: Um, but to get to your question about the cost, like right now, our anchor is about $3,500.

SPEAKER_81: And so if you think about- $3,500, so cheap. Wow.

SPEAKER_13: Well, you know, when we looked at this, the comparison is what do people today pay for driver

SPEAKER_17: assist features? And what do people, and we know we did a lot of surveys around what is this sort of

SPEAKER_84: worth? But if you translate- And Tesla charges 8,000, I think? Is that the right number or 10,000?

SPEAKER_13: They change it every month. I think right now it's, I think right now you get the first one built in,

SPEAKER_17: like basic autopilot built in, and then there's an extension. Yeah. So, so if you think about that in terms of, you know, what model of car, obviously you're not going to put this on a 10 year old car.

SPEAKER_13: That's only that itself is only worth $5,000. We see that the buyers have extremely expensive cars. Like if you're buying a, you know, $60,000 plus car, they usually just don't want to do anything to them. And so you're kind of in that mid range and the average new car in the US is about $35,000.

SPEAKER_88: So it's like sort of centered around there. Got it. And so what is your test? What is your mule?

SPEAKER_89: What do you, what do you use as your muse right now in terms of in the lab, what are you chopping

SPEAKER_17: up in the lab right now and Frankensteining? Yeah. So right now we're using a Toyota Camrys because they're really common. No one else uses them because they have no features in them that enable any automation at all. Like, so it's a super basic car, very efficiently built and

SPEAKER_20: LiDAR or cameras. That was Elon's big bet. He said, I think this can all be done with cameras.

SPEAKER_22: LiDAR is too expensive. Waymo said, no way it has to be this giant 3D model. And we have to create a

SPEAKER_04: LiDAR 3D model. It seems like, am I correct that the world is proving Elon right on this crazy bet? And are you taking the same bet that this can be done with cameras and sensors, not LiDAR?

SPEAKER_40: Yeah, we're on team camera. A lot of that is we're on team any sensor that you can buy

SPEAKER_13: that has large scale. And that isn't LiDAR today. What does that mean?

SPEAKER_17: Well, so it's large scale. Yeah. Yeah. So the world market for LiDARs today is a couple hundred thousand units. So it is, you know, outside of self driving cars, it's primarily used for like architects and land surveyors. Like it's just, it's a very, very specialized piece of equipment. So the idea of high volume LiDAR doesn't exist. There's lots of companies working on it.

SPEAKER_13: So maybe it'll exist in the future, but even then it's never going to hit the scale of cameras. It's like cameras are sold on the scale of billions a year. Like every phone has five to

SPEAKER_98: seven cameras in it now. Right. And so they're also getting better all the time.

SPEAKER_44: Is LiDAR getting better all the time? Or is it just incrementally getting better?

SPEAKER_40: It's incrementally getting better. Like, and I think you're right about the cameras,

SPEAKER_13: like imagine the R&D pressure of like the billions of dollars that are spent on R&D to improve cameras, both the hardware and the software. But LiDAR, I mean, the first LiDAR came out in 2004. And that was a 64 laser LiDAR. Right now, the best ladder you can buy is 128 laser LiDAR.

SPEAKER_45: Like that's twice as good over 15 years. Yeah. So it's very incremental.

SPEAKER_101: Price has dropped from 50 to maybe 10,000 to put it on a car. Am I correct?

SPEAKER_95: Yeah. Well, that's for one. Most of these cars have about a dozen LiDARs on them.

SPEAKER_37: So they're still up in 100,000 range, which is ridiculous. Why are people obsessed,

SPEAKER_04: or certain folks obsessed with having LiDAR, when LiDAR hasn't proven itself, and is not growing at an exponential pace, and doesn't seem necessary? What would they argue is the reason that LiDAR is so essential, and they keep, Waymo keeps, you know, investing year after year in LiDAR?

SPEAKER_13: Well, there's definitely an evolutionary track, because LiDAR was put in the first AV,

SPEAKER_17: like back in the DARPA Urban Challenge days. But it also answers an obvious question, which is, it does do ranging. And it is very, very accurate to say,

SPEAKER_13: I can avoid colliding with something, and I have a sensor that directly detects whether there's an object. But that's existed, and we still have collisions, right?

SPEAKER_109: So LiDAR isn't perfect.

SPEAKER_13: Well, what I would say is that the LiDAR itself is very, very good, but that's not the problem. The problem is, what do you do once you get that signal? Because there's still a lot of prediction that you have to do, because it's not enough to say that you're going to avoid an obstacle. You have to predict the motion around it. And then there's like a lot of control choices that you have to make. So we're kind of right back to the software problem, which is what caused all the

SPEAKER_48: problems.

SPEAKER_37: What if we look at the problems of identifying objects on the road in the path of the vehicle,

SPEAKER_22: that seems to be the problem that people have not yet solved. So is this a plastic bag flying across the road, a garbage bag that just flew off the back of somebody's truck? Or is it a bicyclist?

SPEAKER_04: Like in the Uber tragedy that we saw in, I think, Arizona, where the safety driver was not actually looking at the road, they were playing Candy Crush or doing WhatsApp.

SPEAKER_22: When you look at that specific acute problem, how close are we to solving that one? And then what percentage of the autonomous vehicle accidents have been because of trying to identify and what is

SPEAKER_51: that problem called? Is it called the floating, you know, plastic bag problem? Or what is that

SPEAKER_84: problem called in your world? Is there a name for it? If not, we should name it now.

SPEAKER_40: There's not a name for it. But if you look at the Uber scenario, okay, you had a clear night,

SPEAKER_13: you had a perfectly straight road, you had exactly one obstacle. This was not a problem of combinatorics.

SPEAKER_93: Combinatorics means multiple problems at once?

SPEAKER_13: Right. This was not a complicated scene. This is one car and, you know, one person. And if that car had made any different decision, the collision would not have occurred. So this is, we're not in the world where, what you see from these platforms is trying to solve subtle problems like floating

SPEAKER_22: bags. What you need is- It's not a floating bag problem. This was like a major object going across the road problem. So did they ever discover what the problem was there?

SPEAKER_17: Yeah. There was a pretty thorough report on that. And a lot of it was actually around,

SPEAKER_13: what you discussed is like they had a pipeline which is fairly standard to say, let's identify an object and then we'll put it to another complex piece of software to say what to do about that. And then they had like other control software that tries to override it. And they had a problem where their camera and their lidar, even though it definitely appeared on the

SPEAKER_40: lidar, was kept reclassifying. It was like, is this a vehicle? Is this a bicycle? And it kept kind of

SPEAKER_109: forgetting its history. Wow. Whereas any normal person, a human being has an instinct. It has an intuition. There's something there we should slow down. So a human might be more precautious and say,

SPEAKER_51: you know what? I slammed on the brakes. I thought I saw something. It was actually a floating bag. Whereas a computer would say, I looked at it. Is it a bike? No, it's a floating bag. No, it's a cloud. No, it's fog. No. And then it got caught in that loop.

SPEAKER_125: Yeah. And it didn't make a decision. It was a, it was a non-decision problem.

SPEAKER_40: Well, and it was also a classification problem, which is fascinating. How about not colliding with

SPEAKER_13: anything? Like, instead of identifying an object, like, how about you don't collide with it? Or you have a much simpler system that, that has very simple rules about things you must not do. So when we think about the software problem, it's like, how do you create layers of software where you can have, you know, sometimes very sophisticated software at the top,

SPEAKER_54: which is dealing with complex scenarios, but then you have much simpler software that has very

SPEAKER_130: basic rules built into it. Like a prime directive. Like it, Yes. Don't kill your creator. It's like, don't kill your creator. No matter what they are.

SPEAKER_37: RoboCop. Don't kill the creator. Like, don't slam into something. But what I want to understand is

SPEAKER_22: how you deal, was we talked about the floating bag. I think it's a nice way to name it, the floating bag problem. When we get back to this quick break, I want to know about the slamming on your brakes and causing an accident behind your problem. When we get back on this week in startups.

SPEAKER_134: Let's get down to brass tacks here. I'm going to give you 50 bucks for your LinkedIn jobs. That's right. Go to linkedin.com slash twist and you're going to get

David Friedberg: $50 off your first job posting. If you don't know about LinkedIn jobs, where have you been? It is the best hiring platform in the world. And one of our, uh, twist listeners, Aaron Mason is the founder and CEO of Emma AI, a startup that uses AI to optimize travel time on your work schedule. And Aaron used LinkedIn jobs to hire a machine learning engineer who started back in July and he received over 110 relevant applications in only four days with a very small, modest budget. He got amazing value and from job posts to offer accepted in only a few days. You've been dreaming about a solution to your hiring problems. Well, here it is LinkedIn jobs. You know, LinkedIn has over 690 million members worldwide and they screen candidates for both hard skills and the soft skills that you're looking for. And they like to put those jobs in front of qualified members every day. Some of them are actively looking. Some of them might be passively looking. You never know. Maybe they're just opportunistic. LinkedIn jobs is the most effective hiring platform in the world. We all know that when you're ready to get that perfect tire to solve all of your problems, you want to get them in there and get them working next week. Well, here's what you do. You go to linkedin.com slash unicorn, you get the $50 off, you post the job, and then you can email me, Jason at Calacanis, and we'll put you in the ad for the next, uh, time we talk about LinkedIn jobs, uh, because you've had so much success. Go ahead. LinkedIn.com slash unicorn, linkedin.com slash unicorn to get 50, 50, $50 off your first job posting. What are you waiting for? Terms and conditions, of course, apply because they're giving you 50 bucks. Okay.

SPEAKER_20: Let's get back to this amazing episode. Hey, everybody. Welcome back to this week in stars. My guest is John Hayes. He is with ghost gh.st. Did I get it right? gh.st or drive ghost.com or

SPEAKER_22: drive ghost.com. And, uh, he's got 80 people. They've raised tens of millions of dollars from our friends over at Khosla, uh, a friend of the show, Keith Raboi, who's been on, I think four episodes now, great guest, uh, and our friend Vino Khosla. We had a great interview with him, uh, also on the podcast and they've got about 80 people here in mountain view and they're trying to solve software and build add on, uh, components. So you can drive your, you know, F-150 maybe even into, uh, the shop and then, uh, put self-driving on it. When do you think you're going to launch the self-driving when we, when will I be able to drive my, you know, uh, 1990s, uh, you know,

SPEAKER_04: Land Rover Defender into the shop and, and have you put stuff on it or, or whatever it is, Camry, as it is. When do you think you'd be ready for market to have trial customers on it?

SPEAKER_13: So I think that we're going to begin doing trial customers next year.

SPEAKER_32: Wow. And I don't, we probably won't be able to support a, a 1990s Defender. Uh, we need some, we need some electronics in the car. Right. Um, so there's a minimum stack,

SPEAKER_147: right? Like the OD, what is it? The ODB port or something? You need to have that?

SPEAKER_13: Yeah. It has to be on CAN bus. That's like 2008. The other thing we integrate with is electric steering as opposed to hydraulic steering. That's the breaking point there is about 2012. So it's

SPEAKER_148: most cars past 2012. Got it. All right. So we, we talked about the floating bag problem. Is there

SPEAKER_22: anything we didn't cover in the floating bag problem of why that's so hard? Because we do have fog and you do have, um, you know, air is got particles in it. And I understand air, it can be difficult for computers to understand what's happening in air. So have we solved air

SPEAKER_20: and does air and air density and fog and humidity, is that still an issue with self-driving and the floating bag problem? Or have we kind of solved that? I think that air is not that big a problem

SPEAKER_40: because the nice thing is it has a very linear relationship. And you can even see this in video

SPEAKER_17: game engines where there's fog in the video game engine. And what it does is says further away

SPEAKER_13: objects become less colorful. So it can actually be used as a positive cue to figure out how far

SPEAKER_151: away something is. Got it. I think further away objects are less colorful. Yes. To the camera and

SPEAKER_77: to the human. To your eyes as well. It's right. Just a physical effect. Yeah. Um, and so

SPEAKER_22: how close are we to the floating bag problem being solved? Because that's going to solve that bag problem is going to cause the slam on your brakes problem, which then you get rear-ended. So I guess we'll call that the rear-ended inadvertently slamming on the brakes problem. Is that inadvertently slamming on the brakes problem happened with Teslas and Waymos on the road and the Uber soft driving programs because you're studying those problems? Or is it generally clipping something

SPEAKER_13: that it should have stopped for? So the majority of collisions that occur with with Teslas, we don't

SPEAKER_17: have good data on Uber, uh, are usually we would say false negatives in that they hit things that they shouldn't hit. And there was a spate of them running into emergency vehicles. And some of that is a consequence of automotive radar. And so, you know, what you're raising is a real concern. So these vehicles that have these safety systems tend to be tuned to avoid slamming on their brakes even if it means

SPEAKER_13: that they're going to hit more things in front. And so, what they have is they're typically using automotive radar which has trouble distinguishing, uh, things that are stopped in your lane from other fixed objects in the scene like signs or overhead bridges or other things that also look stopped because its resolution is very low. And so that's why you tend to see that type of failure. Got it. The plastic bag thing, I don't believe that that's going to turn out to be a big problem.

SPEAKER_22: Got it. So the stationary objects, a police officer pulls off to the side of the road, but one third of their vehicle is still in the road or some painting truck is painting the left most lanes, left most stripe or putting down cones. That's what we see on some of the Tesla accidents is they just clip and it's the front side driver is going to get that impact, but not the middle of the car or the side of the car. And it, the reason you're telling me is the radar, not LIDAR, the radar doesn't know, hey, that's the overpass that you're coming up on versus a police cruiser in the side lane.

SPEAKER_40: That is the problem. Yeah, because they often filter out everything that's not moving because

SPEAKER_16: they're trying to filter out all the fixed objects in the scene. Got it. And then you can see that warning

SPEAKER_22: directly in the owner's manual. In the, in the, in the Tesla autopilot owner's manual is that, hey, you have to keep your eyes on the road because we haven't solved this problem yet.

SPEAKER_20: You're trying to solve that problem. What is the solution to that problem? Is it using LIDAR?

SPEAKER_58: Would LIDAR solve that? Or is it using more cameras at different angles or more radar? What

SPEAKER_12: solves that problem? So for us, we're solving it with cameras and particularly we're solving it with

SPEAKER_13: video. So most of the time, like camera systems, like go back to autopilot one or what's available in most cars, they go through two stages. They try and recognize what a vehicle looks like in a still frame and they put a box around that. And then they use that box size to estimate how far away the car is. And what we do instead is we use sequences of video and search for parts of the scene that are expanding. And this is a lot of, this is just like how your retina works, where you have specific sensors for looking for expanding or contracting objects. And so we build a model that's similar to that. And what's nice is that provides first close collision avoidance that isn't dependent on the texture of the car. So we don't have to learn every single type of car. That's really important when you want to make a universal collision avoider on a video. But the other thing is it's also perceptual. And so if you think about the rag quality, some of the times you get like autopilot or other company's products can be laggy in how they respond to vehicles looming

SPEAKER_17: because they don't use the same perceptual model that people have for deciding what's a comfortable distance or what's a comfortable rate. What does looming mean in this context when

SPEAKER_13: you say looming? Looming just means, you know, you don't see the world in 3D, you see it in 2D. Right. And so it's just how quickly is that expanding over your retina?

SPEAKER_163: Got it. So like Jurassic Park objects in the review mirror may appear larger than they actually

SPEAKER_22: are smaller than they are, whatever. When something is charging at you like a T-Rex, it's getting bigger. Yeah. So this is why we have this in our system specifically is because we fear predators, predators come charging at us and it does something to our somatic system when something gets big real

SPEAKER_125: quick. It makes us alert.

SPEAKER_45: Yeah. You have a direct sensor that connects from your retina directly to your motor cortex, just for looming objects.

SPEAKER_33: Oh, wow. I didn't know that. What a great feature by Darwin. When did Darwin add that one? It's like,

SPEAKER_30: by the way, there's going to be a lion that's going to charge at you. We don't have the time for you to process it in your brain. We're just going to go right to your spinal cord and make you run.

SPEAKER_22: It is true. You see, and that's why you get a jump scare in a movie because in a movie, they'll just have some object come charging at you from the closet of a ghost or something, or a serial killer or something, and it makes you jump. It's called the jump scare because you don't even have a chance to process it. It's just the size of it getting bigger that makes

SPEAKER_20: you jump. That's fascinating. So I guess the big question on everyone's mind when they're listening to this is, when do we think that full self-driving point to point, which I think most people would refer to

SPEAKER_22: as level four, this is when you don't have to have your eyes on the road. Level one is you give over one function. Level two is you give over two functions, stay in the lane, adapt to cruise control. Level three is you can be somewhat distracted and level four is you can basically sleep, right?

SPEAKER_77: So the, yeah, the way it works is level three and level four are hard to distinguish. We would

SPEAKER_13: classify ourselves as level three, meaning that there's a protocol for when the system is engaged,

SPEAKER_17: you can be disengaged. And if the system decides that it can't continue to the destination,

SPEAKER_13: it has to pull over. It's like it's still, it can't demand that you wake up and pay attention. I don't think that's a very good product idea. So it has to come to a safe state. And so the, the main difference between level three and level four is how much redundancy there

SPEAKER_17: is in the system with the main issue being powertrain redundancy. So no one has really

SPEAKER_20: made a level four product today. Ah, interesting. So level four means if it has a problem, it can take over itself and try to resolve the problem.

SPEAKER_54: Yeah. So if there's certain types of failures, it has a limp home mode. It can be more artistic

SPEAKER_40: around getting to a safe condition. Um, but it's, it's a pretty fuzzy distinction. And I look at it from a product point of view. It's like, what do people want? Well, they want

SPEAKER_13: to make sure that when they start the system, that it's not going to get them into a state where they

SPEAKER_22: suddenly have to take over. Okay. So this is what's on everybody's mind. We all like driving to a certain extent. What we don't like is having to exclusively drive because we all got ADHD and we'll have smartphones on our dashboards. So what everybody wants to know is not when I'm going to remove the steering wheel because we kind of have an affinity and we don't mind driving. I think by and large, but we would like, we would like to be able to pick our podcast or maybe reply to a text message or perhaps scroll through YouTube and pick a video to watch. We want to be distracted for a minute to two minutes at a time. I want to know in your best estimate, what year we will be able to

SPEAKER_51: be distracted for a minute or two while we're driving when we get back on the swing of service.

David Friedberg: You need business insurance for your startup without insurance. You failed one of the earliest tests of properly running your company. You know, some of the examples, I'm going to walk you through them right now because I do this when I'm a board member of every single company. And I see people make every mistake you can imagine. And one of the biggest mistakes is not having insurance for when you make mistakes like cyber insurance, in case you get hacked. How many companies do you know? How many of your peers who are founders have been hacked? What about directors and officers insurance in case somebody does something stupid and you get sued in your company? You want to make sure your board of directors and the officers, i.e. the top executives of the company, that they have insurance to pay for lawyers, to make sure you're defended properly. And there's E&O insurance, which covers errors and omissions. When you're scaling, you need to have E&O. And a lot

SPEAKER_51: of big customers are going to ask you, do you have cyber? Do you have D&O? Do you have E&O? And of course, there's employment practices liability. That's EPL. And that covers bad stuff like harassment and

David Friedberg: wrongful termination. And a broker's technology saves you time and money. The prices are up to 20% lower and they give you better coverage. And you go from sign up to quote and a purchase in just 10 minutes, as opposed to dealing with large, slow incumbents in the insurance space who take weeks to get you quotes. And you don't even know if you're getting a good deal. The process at Embroker is transparent and there's none of that opaque pricing. Here's your call to action to instantly buy custom built insurance for startups. And that's the important part here. This is for startups. Go to Embroker.com slash twist. E-M-B-R-O-K-E-R.com slash twist. What a great name. Embroker.com slash twist. And while you're there, you're going to get an extra 10% off, which is very generous of them. Thank you for that. By using the offer code twist. That's right. You go to Embroker.com slash twist and you get 10% off. If you've been ignoring this for two or three years, it's time to stop what you're doing. Put it at the top of your punch list and go to Embroker.com

SPEAKER_51: slash twist. Use that code twist to get another 10% off. Thanks to Embroker for supporting independent media like this week in startups. Okay, let's get back to this amazing episode.

SPEAKER_20: Hey, everybody. Welcome back. It's our next unicorn series, number eight of 10. We've got two more to go. Today on the program, John Hayes from Ghost. And you can go check out the website, gh.st.

SPEAKER_22: You're working on adding these units to existing cars. You're going to be able to do that starting next year. So if somebody has a Camry, they're going to be able to get it. Great. And you're working on the software using cameras, LIDARs. You're not opposed to it. You might add it later, but you think you can get it done without LIDAR, correct? That's correct. Okay, so Elon made the right bet there. Waymo was wrong. Tesla was right.

SPEAKER_176: I think at least the same bet. So we'll take that.

SPEAKER_22: Yeah, you're taking the same bet as him. I don't think people would doubt that's the wrong bet. It sounds like it's the right bet. So in your estimation, this like level three, where I can change the podcast, and I can watch a video or I can do an SMS, which candidly, let's be honest, we know that people are in all likelihood doing this already with their Hondas and their Audis and their Teslas, which have autopilot like features, people are already starting to maybe take their eyes off the road for five or 10 seconds at a time. And they shouldn't, but they do. Is that correct?

SPEAKER_13: Just like the Uber driver did? Yeah, that's obviously true. And everyone knows that it's dangerous, but they do it anyways. People are extremely compelled to do things other than driving.

SPEAKER_179: Right. I mean, and this is tragically the Apple engineer who was using his Model X and who crashed

SPEAKER_22: into a divider. He specifically was not paying attention on an exit ramp of all places to really

SPEAKER_04: push the envelope on a Tesla's. One thing in any of these cars, because my Honda Odyssey also has, you can stay in the lane and adaptive cruise control. So it's very much autopilot-ish. It's just not as good, not even close to as good, I would say, as the Tesla's. I don't think they've

SPEAKER_22: really worked on it that much. But if you're on the 280, it's one thing. If you're on a straight shot, these things are dialed in. I think they're close to perfect, right, on that kind of scenario?

SPEAKER_32: You would agree with that mission? Yeah. From the Rhodes point of view, yeah.

SPEAKER_22: Yeah. So it's only if there's a wild card that happens. But to my question before the cliff hangar in the break, thanks to our sponsors, when will we be able to legally and ethically, morally, and safety-wise, do what pilots of airplanes do? When the pilots put the autopilot on, they'll look down and have a bite of their sandwich, have some coffee, they'll look at their maps, and then they'll keep looking out the window. They're not looking out the window constantly to make

SPEAKER_04: sure they don't collide with another plane. They have autopilot on, and they have sensors, and they got a lot more space, obviously, with their x-axis, or their z-axis, I guess. When are we going to be able to do that? When will we be able to take a minute or two off,

SPEAKER_22: do you think? You had to put a year on it in the United States, in California.

SPEAKER_41: When are we going to be able to, for God's sake, flip through our SMS messages without risking our lives?

SPEAKER_13: I think that'll be 2022. Part of the reason why that date is much earlier, because the question is when you can do it, where? Because what you have on the roads is you have a huge variability in the complexity. When we started, people thought that the highway was so simple in 2017 that the auto companies were going to be able to incrementally innovate their way there. We've seen that that's not straightforward. From our perspective, the highway is plenty complex, but you go from there from a freeway to a highway, from a highway to a regional road, from a regional road to a structured sort of suburban road. Each time you take one of those steps, your decision complexity goes up. When I think about this from a product point of view, it's like, well, the simplest product is to go where the decision complexity is the lowest. And that is where the speed limit is the highest. Because we're humans that have constant rate brains. And so, the reason that we can put the speed limit up is because the environment has been so simplified that we're not going to have to make

SPEAKER_22: a lot of complex decisions. We've reduced the number of variables. On a highway, you know you're not stopping. You know there's not going to be cars or people on the highway. And if there are, there's something tremendously wrong. They've made a bad decision.

SPEAKER_40: Yeah, and you can deal with them as if they're obstacles, in a sense. Right.

SPEAKER_13: Like, because your delta speed between someone walking around on highways is so high that they're effectively not moving. Right.

SPEAKER_45: Compared to if you're in Mountain View and you're driving on Castor Street at 15 miles an hour, well, someone on a bicycle is kind of going the same speed as you. Or faster.

SPEAKER_118: So, your peers in that case. Yeah. In Silicon Valley, they're definitely going faster. These like crazy venture capitalists on their

SPEAKER_04: like road bikes are like going 40 miles an hour. So, you're really focusing on providing this

SPEAKER_22: opportunity on the highway. And are regulators also kind of behind this concept of like, hey, let's master the highways. Let's give people that freedom and time back on the highways. That's a great milestone for everybody to work towards. Or have they not, or they're just leaving it up to the companies working in the space to sort of submit their application that you want to do this.

SPEAKER_04: Like who's driving this concept of, let's solve highways first.

SPEAKER_13: So, I think regulators, you know, are interesting. Because what they tend to do at the federal level is they tend to document existing practice. And so, for example, lane keeping features that you have today do not have controlling regulations. And that's fairly normal that you introduce a new feature, you let it get engineered, you learn the best practices, and then you codify those best practices. And that's what constitutes the regulation. Got it. And so, a lot of what you'd be evaluated on is what is your engineering process? Have you done the reasoning that says you have a positive case for whether this is going to work? And that's kind of the opposite of a lot of other technology companies where you kind of test something to prove that it sort of doesn't work, right? You're making tests for the reason it does work. Here, you have to make a positive case. And that's what you're evaluated on.

SPEAKER_22: Interesting. Now, do you have to apply for that? Like, are you going to say to them, hey, we want permission to go level three on these specific highways? What is that process going to be like as we navigate through this? Because I know there are some states, it seems, or some regions, I don't know if it's city or state level, but there seem to be people who are aggressively courting your companies to make their cities more attractive for that business, both in the short term, I think, attracting companies like yours, and then in the midterm, attracting, you know, individuals to have a better quality of life because these features are allowed and be progressive. So who's the most progressive? And then how does it work that you actually would, by 2022, be able to say to California, hey, can we, are you going to say these specific highways, we think we've got mastered? Can we get

SPEAKER_12: permission for level three here on 101? So the way it works in California is you apply for various levels,

SPEAKER_13: degrees of testing permit. Now, we're not operating a taxi company, nor is anyone else. And there's an additional level of permit where you could say I could sell this. Now, people are fundamentally allowed to just modify their cars. And so you get permission from California to have your product deployed. Now, when you talk about progressive, I would say California is the most regulated of any state in that there are controlling laws for autonomous vehicles. In almost every other state, there are no laws. And so I think that it comes down to a direct negotiation with the executive of the state and

SPEAKER_17: probably the state police to get sort of permission or at least acceptance.

SPEAKER_22: And how are you going about that? Who do you think, you know, what is your process? Do you think it's, it's California is going to be the way to go? And they've been reasonable and you coming

SPEAKER_04: up with examples of what you want to do and saying, hey, here's our plan. And, or do you think you need to go to Texas or someplace that's like a little more, you know, wild west, so to speak?

SPEAKER_40: I think that right now we're on the well-worn path of getting testing permits,

SPEAKER_17: that you have to make a lot of certifications around your process. How much did you test off the road before you test on the road? No one has proceeded much beyond that point. Got it. And so I think that we're going to discover that probably in 2022. Hopefully we discover

SPEAKER_13: it before then 2021. But right now, no one knows what the acceptable standard is. Okay. And I think that this is the case where AV is such a frontier that no one can point to a formula and say, this is exactly how you engineer it. So we need an existence proof. And then you can, you know, take that apart and figure out how they did it. And then that's how you build a standard.

SPEAKER_20: Michigan has approved an AV only highway. What is the nature of that corridor? And is that material

SPEAKER_41: for what the AV industry is doing, the autonomous vehicle industry is doing? And do you see that as

SPEAKER_40: the future of this? I'm not familiar with what Michigan is doing specifically. I think that

SPEAKER_17: that for at least the medium term, at least the next 10 or 15 years, you generally shouldn't make

SPEAKER_13: a startup that depends on government building expensive infrastructure. And so I believe it's like you, we're going to be in the world immersed in human driven cars for a very long time.

SPEAKER_20: Is AI the way to go about building this rule set in software and just plugging in all of the inputs,

SPEAKER_22: all of historical inputs. So you have all the times it's been disengaged, right? All the times autonomy has been disengaged, all the times there has been accidents documented, all the GPS data, every videotape of every accident, all just put into an AI that says go and, you know, protect the

SPEAKER_04: driver at all costs. Or is this going to be more verticalized and narrow AI or narrow machine learning? You know, we're going to really work on this in sort of segments. And then how do you how is the model

SPEAKER_40: shaping up in that regard? So AI is a really big term. The way we approach it is like you're going

SPEAKER_17: to make a model and you make a model to answer a very particular question. Like, where are the things I'm going to collide with? Where are, where is paint on the road? What is the sky versus other things we should ignore? And then you can break that question down into what are all the ways we can analyze the data to figure out what are, you know, what are the axes along which that is important. So if I were to want to say all of the things I can collide with, I could start with a very simple thing to say, okay, well, I could have obstacles colliding with me from all different angles. Like, let's say I'm going to do a clock face. I have my 12 o'clock, my 1 o'clock, my 2 o'clock. I could have different colors,

SPEAKER_13: I could have different sizes. And out of that, you build axes for what you want the system to detect. And so, when we think about model building, it's very much an engineering process to say,

SPEAKER_17: well, first, you have to be very specific about what you want the model to, what question you want to answer. You have to be very specific about the domain, like what are all the types of inputs.

SPEAKER_13: And then out of that, we go into the world because we have cameras on cars and we have a fleet. And we look for those situations to occur naturally in the world. And then you can iterate on that and you can say, okay, well, we had two scenarios that we thought were the same that gave us a different answer. So now you put that back into a creative process that says, okay, what axes did we miss so that we can make a new training dataset? And out of that, you also get things like holes, like what combinations of the world haven't I seen? And then you can pull that in from your fleet. So when I think about it, it's like the AI or the neural network itself, yeah, that's how you execute it. But all the engineering goes into the process that actually got you that data and how much human knowledge can you encode into that system? And what is your process for detecting gaps?

SPEAKER_37: Yeah. And a gap I always think about because most of these companies are based in California is snow

SPEAKER_22: and ice. And are those problems solvable in the near to midterm? Because it seems to me when I drive back from Tahoe, you know, and it's just one sheet of white, I can't imagine when the highway is a

SPEAKER_04: giant sheet of white with a little bit of black ice under it, that an autonomous vehicle is going to be able to figure out a sheet of white and where the lanes are, because the lanes become, you know, up for interpretation during these kind of conditions, and people just do the best they can. And they kind

SPEAKER_51: of self form, right? If there was a three lane highway, it kind of becomes a two lane. And then

SPEAKER_32: you just you're following ruts, and you kind of divide up the space.

SPEAKER_51: Yeah. So what do you think about ice and snow conditions?

SPEAKER_13: And I think that goes to, you want to answer a very specific question. So this comes back to more narrow AI applications, right? Like, I think you just create a completely new driving mode for snow and ice. And you just say, Look, it's very easy to detect whether that's snow, you look at the road, you say, What is the albedo of the road? Or how much light is it reflecting? And you can come up with four different values. This is pretty well understood in the automotive engineering space. So to detect like, what is the material of the road surface? And then you can

SPEAKER_45: just create completely separate models from that. So you just switch between modes,

SPEAKER_37: your car just might say to you, it appears you're in a blizzard, your peers are in snow conditions. Please confirm you would like to go into blizzard mode. So it'd be kind of like,

SPEAKER_222: Or it should just do it. It should just do it. Just do it.

SPEAKER_04: I mean, in the midterm or the near term, just like a plane might, you know, during a wind shear, say wind shear, wind shear, or whatever, you know, or like, you know, it might give you some guidance and you kind of help it along. What about ice specifically? Can you detect ice easily? Or is ice just, you know, some, you know, we'll get to that 10 years from now? How do you think about ice?

SPEAKER_13: I think you detect ice the same way people detect ice, which is you can, you finally measure how

SPEAKER_54: much friction you have at any given moment, which you can do from the car right now. So you get the wheel rotation speeds.

SPEAKER_179: The oh shit indicator, like the car is floating, the car is not gripping. That's yeah, they basically,

SPEAKER_51: it's basically the, what do they call it, the all wheel drive, I guess, when it when it shifts the power to different wheels. So you actually know that because you're right into plugged into the computer's main system. Yeah, so you know that computer.

SPEAKER_40: Yeah, the computer tells you how fast all the wheels are rotating. And so detecting that you're on

SPEAKER_17: ice is actually a problem that's been pretty well solved in the automotive industry.

SPEAKER_20: Interesting. And when they solve that problem, did they hard code that? Or did they? I can't imagine the automotive industry used AI to figure out how to manage ice conditions. So there's probably

SPEAKER_22: an opportunity there. However good all wheel drive is today, an AI all wheel drive would probably be better because it would learn about the specific ice at that specific moment and adapt to it.

SPEAKER_13: Well, I think what you could have, and this goes to the development process. So when you're developing a car, I mean, you have fixed test scenarios and you put it through that fixed test suite. And what you have the opportunity to do once you put a computer in every single car that can communicate, you can now collect up all of the ice scenarios in the world. Wow. And so a lot of what I think about for building a system like this is what is your QA process? And that, how do you discover everything? Like what are all the possible ice scenarios so that we can bring it back to our data center? And so every time we change our control system, we can test them all simultaneously.

SPEAKER_236: Right.

SPEAKER_16: And that's how you build up that learning system.

SPEAKER_37: It's incredible. And what about car to car communication?

SPEAKER_04: So in the same ice scenario, you left, you know, Truckee at 11 AM and the road was, you know, light flurries. I left at 1 PM and it was, you know, icing up and it was kind of icy slushy. And then somebody else got pure ice at 3 PM. But we all had that data, you know, is car to car data or recent data, like in the ways we have traffic data, and that's been a game changer for routing. Is there an opportunity for the real time data to make it into the self-driving function or is that overkill? Like cause something could be on the road, like an object on the road is the obvious one. The other car should know there's an object on the road and avoid it, but there could also be things like, you know, paint or a construction area that changes, you know, or snow, you know, those kinds of conditions.

SPEAKER_13: I think that there's a whole industry around V to X, which is vehicle to vehicle or vehicle to

SPEAKER_17: infrastructure. Me personally, I'm not bullish on it because you have a huge security problem, which is how do you know that someone broadcasting information is a good actor?

SPEAKER_236: Right.

SPEAKER_13: And when you look at how those standards evolve, they sort of said publicly infrastructure, something, something, which is not a good answer for how you get a good actor or bad actor in there.

SPEAKER_17: Um, and I think that even if you had that system, you would still want to develop an observant, like a visual based system though, that was as accurate as you can make it, if not

SPEAKER_88: a hundred percent accurate. Yeah. It does seem rife with complications because the data gets old and

SPEAKER_04: it could be compromised. What about, um, changing the road itself in some way? I've seen demonstrations on the YouTube and other places where people specifically paint a middle lane in the highway for self-driving cars to kind of lock into. Um, and the Tesla kind of does that when you're in autopilot, it draws a middle lane that it's kind of dialed into when you navigate on autopilot. Are there things that could happen on roads that would help self-driving or is it, is it good enough now that that's not worth pursuing?

SPEAKER_12: I believe that it's good enough now. And again, we solve that kind of the same way

SPEAKER_17: that we would solve something like ice is we actually have libraries of videos of all the roads that we drive on. And so we don't have to make a heuristic and guess whether it works. We, we actually have original sensor data from every single road that we drive on. And so we can tell ahead

SPEAKER_13: of time, whether it's good enough. And something we don't do is we don't actually use map data. So a lot of companies use HD maps where they say, Hey, if I can just figure out where I am in the world through localization, then maybe instead of reading the road, I'm going to use what's stored in my map. So that's not something we do, but what we do store in a map is maybe some clues about how that road surface was constructed. And so, you know, some highways are like brand new, and it's like very bright white on black. Some of them like parts of 280 are old, where it's old concrete, it's very faded and giving the visual recognition system, those sorts of clues and help it perform, you know, well on each type of road. And so again, you can break down the problem in that you don't need to make a universal lane detector like a person would have. We want to, we want to simplify the problem for the computer and say, look, let's make a few stereotypes of road. And it's kind of, it kind of maps to the decade it was

SPEAKER_250: built. And, and just say, depending on that stereotype, we know how long ago that road was built.

SPEAKER_88: Black asphalt with white, you know, bright white line, lane dividers is going to be one problem

SPEAKER_22: set versus the faded concrete with faded white on the 280. Although that 280 is just beautiful driving with those concrete highways. Are concrete highways better for self? Is there, is there a difference

SPEAKER_41: between the highways? Is like New York asphalt, like just too random and like California highways are just perfect with that beautiful concrete?

SPEAKER_99: I, I, we have to work on everything.

SPEAKER_255: You have to work on everything so it doesn't matter. I'm not going to judge the roads. It's like, it's, it's our job to make sure that they all work.

SPEAKER_20: I'm curious as we wrap up here, is there a wild card that nobody's expecting

SPEAKER_22: that would advance self-driving quicker? Because most people seem to be thinking that,

SPEAKER_04: you know, that the steering wheel comes out in year 10. And, you know, to your point, like on highways, we might be able to get a little bit of our time back, you know, in, you know, whatever it is, 2022, 20, 28, somewhere in that timeframe, we'll be able to take our eyes off the road for a small amount of time or periods of time. But is there any technology that could be game changing either in chip technology or sensor technology? That's in the work that people are looking at going, well, if that goes right, boy, that could be a game changer. Are there any game

SPEAKER_144: changers you can think of?

SPEAKER_17: So, so the interesting sensor technology I'm looking at, and no one quite builds it. I've seen presentations, they all say 18 months away, which is like, never.

SPEAKER_258: Yeah, that's the standard technology.

SPEAKER_13: And that is, yeah, and that is 5G. And it's not, and not using it for data transport.

SPEAKER_16: One of the things that 5G has is it has a beam forming radar in it. So 5G at its high frequencies operates at the same frequencies as radar. But instead of just being a single pulse that you send out in

SPEAKER_13: the scene, you can actually scan the scene by controlling the direction of the beam. And so what's interesting to me about that is first, you can make a high resolution radar return, but based on solid state devices that are in every single phone. So you get high quantity, high quality, you know, that'll drive the power down, drive the cost down. So you have those forces. So there's something there in, you're going to have one side of it, the 5G people, you know, driving the cost on the intended design, driving the cost on the DSPs, or however you process the signals. Yeah. And that would be very interesting for us as a really, call it unintelligent collision

SPEAKER_16: avoider.

SPEAKER_266: Right.

SPEAKER_16: In that it doesn't have to understand anything. All it has to do is say, I want to avoid frontal collisions. And you could build a very, very simple system that would prevent all forward collisions. I think that would be a very interesting advance.

SPEAKER_268: Why doesn't that exist as a standard safety technology that every car is required to have,

SPEAKER_22: you know, under 35 miles an hour, slam on the brakes, if you're going to hit something within five feet of you, period. Like, shouldn't that be much easier than self-driving to create as a, you know, airbag type, you know, stepping stone along the way and just get rid of every single fender bender and then retrofit every old car. Like that could be a $300 product for you that insurance companies would pay for and say, you know what, no more fender benders.

SPEAKER_50: Technology in the auto industry evolved very slowly. I think that's what it comes down to. It's like,

SPEAKER_13: it already has to be proven. And some of that is because when you, when you build something into

SPEAKER_16: a car, as opposed to adding it on, cars have like a 20 year lifetime.

SPEAKER_13: Right. And so that's often why technology in the auto industry tends to lag the consumer industry by five to seven years, because it takes a bunch of extra time to prove that some piece of electronics is going to last 20 years into the future. And so I think it, it's something that they'll probably,

SPEAKER_32: it'll, it probably will be introduced into cars.

SPEAKER_22: Yeah. That'd be the ultimate aftermarket thing. Just no fender benders. You put this device like, you know, those in debt, those in grill radar detectors, you can get aftermarket, like an in grill device that just slams on the brakes. If there's something right in front of it, you just solve that one acute problem. That's gotta be half of all accidents or tiny 35 mile an hour or less fender benders, right? Yeah.

SPEAKER_130: All right, John, I know you're hiring, right? So how can people find out about the positions you're hiring for? And what are you looking for? What's the culture like?

SPEAKER_17: Well, the culture is a very intensive engineering culture, because we have to make a product that's safety critical. It's a very simple product, but in some ways, it's, you know, we cover every single area of computer science. So we cover compiler building data centers, and of course, machine learning infrastructure and math. And you can find out about the company, you could actually reach

SPEAKER_256: out to me directly on Twitter. Cool. And we could-

SPEAKER_20: Ghost haze. So just ghost haze, DMs open. If you're looking for a job and you want to save human life,

SPEAKER_04: there's 30,000 people die on the roads every year, and we're going to get that number below 30, I think. I think it's going to be like people intentionally crashing their cars is going to

SPEAKER_51: be the only way to crash. And even then you could, you can make cars impossible to crash. Like if somebody tried to run off the road, I could just not do it.

SPEAKER_120: Yeah. I think that's, that's further out. That puts a lot of faith in technology, but I think we're going to get there.

SPEAKER_281: This driver is trying to run this off the road. It's like, no, we're not going to let you do that. Sorry. I'm not opening the bay doors. All right. Listen, continued success, John. Thanks for coming

SPEAKER_51: on the pod. Stay safe during this crazy time here in California and everybody go ahead and check out

SPEAKER_04: Ghost Locomotion and follow Ghost Haze on the Twitter. We'll see you all next time. Bye-bye.