This Week in Startups - No hype, Just works: How Comma reached 100M miles in autonomous driving | E2011

Starting point is 00:00:00 I'm very excited about robotics, but I think we should be realistic. The big reason a lot of people went to self-driving 10 years ago, including me, is because it seemed like a great applied robotics problem that was easy. You have two-dimensional actuators, you have simple rules of the road. As far as robotics goes, that's relatively easy. And then what happened is all these companies were optimistic and ended up not reaching their goals, and now all of a sudden everyone's switching to humanoid robotics, which from the beginning, we always thought was harder.

Starting point is 00:00:27 So I think it's just another hype wave. I don't think there's going to be humanoid robots in your house and we should be somewhat cautious about everything that's just a demo and is not shipping. This week in startups is brought to you by dot tech domains. Don't miss our Jam With JCal contest. To apply and get more details, go to jam with jCal.com. Brought to you by dot tech domains. Vanta.

Starting point is 00:00:51 Compliance and security shouldn't be a deal breaker for startups to win new business. Vanta makes it easy. for companies to get a SOC2 report fast. Twist listeners can get $1,000 off for limited time at vanta.com slash twist. And Micro 1. Micro 1 is an AI recruitment engine to hire world-class engineers fast. Visit micro1.a.i slash twist to open a talent search and get a two-week free trial per hire. All right, everybody.

Starting point is 00:01:22 Welcome back to the program. I'm very excited today. To talk about self-driving. cars and autonomy feels to me like the autonomy endgame is upon us. We're seeing it in VTOLs, vertical takeoff and landing companies. We're seeing it with Waymo, Tesla, and today's gas comma AI's CTO. Harold Schaefer, welcome to the program. Harold, how are you? Thank you. Yeah, I'm doing great. All right. So I had George on the program. I'm trying to remember what episode that was, gosh, it was a while ago.

Starting point is 00:02:00 Eight years ago or something. Eight years ago. For the audience who doesn't know, explain what Hama. dot AI is doing in the autonomous space and how it's different than Waymo and Tesla, the other and Cruz, the other major players in the space. Right. So our goal is to solve robotics, general purpose robotics. And in the meantime, you know, ship useful products to people.

Starting point is 00:02:26 that we can sell them for money that add value to their lives. And so today what that means is we sell a kit that runs a software called OpenPilot, that's completely open source, and it's an ADAS upgrade for your car. So basically it will take over the internal messaging of your car, and it can send gas commands, brake commands, and steering commands, and it can make it drive itself on the highway. So it's a bit like Tesla autopilot slash FSD. Today, from our users, over 50% of the miles are driven by,

Starting point is 00:02:56 So it kind of gives you an idea of, you know, how much of the driving it does. It's a level two system. It's not fully autonomous. It just makes your drive more comfortable and it's kind of a value add. And, you know, as we progress with the technology, we want to, you know, increasingly make these things autonomous and, you know, make robotic products that we can sell. Got it. So the mission of the company is general robotics.

Starting point is 00:03:19 The first product is this level two autonomy. level two, I believe, correct me, if I'm wrong here, is taking over two functions, and I believe the two functions that you tackled first with comma AIs kit is staying in the lane and adaptive cruise control, am I correct? Yeah, exactly. But, I mean, it's a bit more general than that. Like, it'll work if there's no lane lines. It'll work if there's no lead, that sort of stuff.

Starting point is 00:03:46 It's just a gradual process to become more reliable, and eventually it will do everything and drive perfectly. but it's just an incremental game. But yeah, the goal is to do general purpose robotics. It's just that self-driving right now, especially partial autonomy, is a very sensible place to make a product. It's something people are willing to pay for

Starting point is 00:04:03 and something that's valuable even if it's not perfect. And so what does it cost to add this to your Toyota, your Honda, and how does that work for people who don't know about the different ports and how modern cars work in terms of controlling steering and speed? Maybe you could give us a little primer on that. Yeah, so it's $1,450.

Starting point is 00:04:25 We'll get you a kit that is a device that has cameras, compute, and sensors, and then a wiring harness that plugs into your car. So that wiring harness is specific to a certain brand or a certain model, and it basically connects to the Canbus, which is your car's internal network. And on that network, we can send the same messages that the car already accepts from the factory, and those messages can apply torque to the steering wheel, they can apply gas and they can apply brake. And when you can control those three axes,

Starting point is 00:04:53 you can basically, you know, fully control the car. So you can see in this video kind of how that works. There's, you know, connectors that connect to this can bus. We can just intercept it. And we can send the same messages that the car is designed to receive and that the stock A-DAS system of the car would have sent. It's just that the stock A-DAS systems generally suck and we can make one that, you know, is actually usable.

Starting point is 00:05:15 All modern Toyodas, even from 2017 onwards, ship with ADAS with the ability to control gas, brake, steering electronically, but they're just not very good and people tend not to use them. But when you use good software like OpenPilot, you can make actually a very enjoyable partial autonomy experience. And so I guess the, well, one question is, how does Toyota, Honda, you know, these, you know, car manufacturers, how do they look at what you're doing? Do they try to stop you? Are they excited about it? Or are they indifferent? Well, they haven't tried to stop us. People that work in kind of the ADAS development there tend to like us. You know, we know quite a few people that work in kind of their research

Starting point is 00:05:58 labs and their ADS. Generally speaking, they like us. They wish their companies would move a lot faster when it comes to this sort of stuff. You know, it's not lost on those engineers that their ADAS solutions are terrible and that compared to something like Tesla or OpenPilot, they're very, very far behind. So generally speaking, they like us. And so, 1500 bucks, you put this into your car, you can either install it yourself or there are

Starting point is 00:06:23 a third-party installers, I understand, who will do this for you? Oh, no, well, there might be, but not any that we're affiliated with. It's kind of a DIY thing. It's not that hard. It's like, kind of working on your own computer. It's not the hardest thing, but it is a project. Wow, this jam

Starting point is 00:06:39 with J-Cal contest has been a blast. So far, I've had the opportunity to meet with four great founders from companies like Corpod, Ullama, Ullama, Uptrans AI, and the ROMAP, all because they all use dot-tech domains. And we have room for one more. Do you want to come on the pod and tell me what you're building? Well, you only need two things to enter. You got to be a founder with under $2 million in funding, and you've got to have one of those

Starting point is 00:07:01 awesome dot-tech domains. So head to Jam with jacalc.com and tell me what you're building. And if you win, I will invite you onto this week in startups, and you'll get to share your vision with me and the world. I'm working with dot-tech domains because killer startups use them. You know, 1x.com, rabbit. Tech, so many others. And guess what? We use it too.

Starting point is 00:07:21 That's right. DotTech powers our founder Friday program. So tell me about your awesome. DotTech domain and startup. Apply for the Jam with J-Cal contest today at jam withjCal.com. We're picking the final winner soon. Okay, so the thing I really wanted to talk to you about is the open source approach that you've taken. It seems to me that open source has won so much of,

Starting point is 00:07:43 the problem space in computing, that it's odd to me that all of the self-driving companies and projects are closed, whether it's Tesla, Cruz, Waymo, or any number of them. So how is the open source project going? Are there many other people contributing to it? And do you see interest or engineers from these other major projects looking at the work you're doing, talk to me a little bit about how that space is shaping up. And if you believe open source is going to win the day here. Okay, so just to start off, I mean, I have dozens of great things to say about open source, but I think the biggest thing it does for us is it keeps us honest and it prevents us

Starting point is 00:08:26 from being able to rent seek. If we make hardware that's quite overpriced or worse than a previous version, someone can just come in, undercut us and run open pilot. If we make changes to the software that makes it just the shi experience, like we can run ads while you're stationary stuff like that, people will run a fork and they'll make changes to undo the things that we did. So it really forces us to make a good product, both in software and in hardware. So that's, I think, the biggest thing that's good about open source for that. And then are there a lot of people contributing? So because we support so many different cars,

Starting point is 00:09:00 it is useful for the community to be able to port new cars, to port their own car. There's definitely a lot of contribution there. When it comes to improving the core driving, experience, you know, it's not really feasible for external people to contribute there. It's mostly for this kind of small stuff. Some people make forks that have kind of different UIs or small changes that can kind of inspire us to look at maybe something we can change or changes we can take over. Talk to me about the approach. I saw in, you know, our notes

Starting point is 00:09:30 here for our discussion, the different approaches people are taking to autonomy and a lot of the the work that's being done in language models is supposedly becoming applicable here. Maybe you could just educate the audience on how this technology was working originally, and then how this is starting to evolve over time with the advances in AI and compute. I'll just add one final note about the open source discussion, which is that companies that are close source, it's most likely because they're trying to hide their lack of capabilities.

Starting point is 00:10:07 In Silicon Valley, it's pretty common that a lack of transport. Farsi means that, you know, they're not maybe not as great as they claim. But yeah, to answer your other question, we've been very big on end-to-end machine learning since the beginning, which means we can take data to see how humans drive. It's very easy to collect data on this. You can record their steering, gas, brake inputs. You can see the video of the road. And you can then learn a machine learning model.

Starting point is 00:10:31 You can teach a machine learning mold and teach it to drive like that. We've said that since the beginning, you know, when we started this company eight years ago, this wasn't really a big thing. Nobody was thinking this way. People had perception systems that would detect all sorts of things about the world with all sorts of sensors. They would then go into some planning logic that makes decisions based on that, based on some rules, and then take driving action. Whereas we've, in contrast, always said, just learn how to do everything like a human. Waymo is a good example of something that has this very classical stack where they detect things, then they have this classical planning algorithm, then they make decisions.

Starting point is 00:11:07 but now multiple companies are kind of coming around. Tesla talks a lot about doing end-to-end learning. There's companies like Wave and a few more that are, this is kind of gaining traction. I think your final point was about the generative AI models, how those become relevant. So to learn how to drive like a human, one of the best ways to do this is to learn in a simulator.

Starting point is 00:11:29 So what we do is you have a simulator that can simulate driving, and you can then let your student, that is learning how to drive, drive in it, and it will deviate from what a human would have done, and then you can tell it to recover to what the human was doing. So that's basically how our system works, but that requires a simulator of driving, and one really good way to make a simulator

Starting point is 00:11:53 is with these generative AI models that can generate arbitrary video, they can simulate the world, they can simulate physics. And I've got some clips. I don't know if we want to show those now. Yeah, let's take a look at this. I mean, I think this is sort of the fascinating, turn, so to speak, that this is taking, which is in a simulator here that we're seeing,

Starting point is 00:12:12 for those of you who are listening, we see a simulated road on top and an actual road on the bottom. Explain what we're seeing here. So those are both simulated road. They're completely generated by a machine learning model. They're just two different perspectives. One is kind of zoomed in, the other is kind of zoomed out. And then that machine learning model, that simulating the world is also telling you what it thinks a human would do over the next 10 seconds. And so what we can do is we can let this agent drive, and by giving action inputs like turn left, turn right, the world model will simulate that deviation and then try to get back to what the human would have done. So this is fully simulated, fully in the imagination of a machine learning model, and we can let a student kind of play and kind of drive around, make mistakes, and we can tell it to recover from those mistakes. And so this is what we're working on today.

Starting point is 00:13:01 We've been working on that for a little over a year now with these generative models, and we hope to ship that. that very soon. So this model has information on what roads look like, nighttime versus daytime, rain versus snow versus clear skies. And it will create simulations, let the driver attempt to do that. And then how do you know if it's making a mistake? And then how do you know to intervene? And how do you know it doesn't hallucinate, right?

Starting point is 00:13:31 Because that's like one of the things that we all experience using chat GPT is, hey, sometimes it's pulling information from maybe a website that has bad data. So how do you know, like, it's not producing, you know, something that is incongruous to the real world? Yeah, I mean, that makes sense. So basically, these models are seeded with some real video.

Starting point is 00:13:53 So we give them some context that is real video. And then we ask them to basically go from there and simulate and we can give it, you know, actions. And then you basically have a simulator. And yeah, I mean, the hallucinating, it's kind of the same thing as just general inaccuracy. These models, when they're very accurate, you know, they produce realistic looking rollouts. When they're inaccurate, they can deviate from the real world. And that's definitely a real failure mode. You know, you can divert from

Starting point is 00:14:18 what looks like a realistic road. Lane lines can cross in unrealistic ways. And that's kind of the project of making these models better. It's just bringing that error rate down and the videos look better and better. Waymo seems to be the one player that has full, autonomous vehicles on the road at scale. Their approach is the old school approach. It's taking all this input and it's saying, you know, if this, then that and, you know, it's got a rule set there that it's following. So maybe you could tell me why they've been so successful and, you know,

Starting point is 00:14:55 what you think of their rollout limitations on what they're doing and or things they're doing that are causing them to hit 100,000. it rides a week. Right. Well, let me start by saying, you know, what Waymo's done is incredibly cool. They're probably the coolest service that you can get today as a normal user.

Starting point is 00:15:15 That's like an actual robotic thing of something interacting with the real world that is actually to some degree autonomous. With that said, I think their strategy doesn't really make sense from a business perspective. I think, you know, they don't have unit economics at all. And a part of that is because of this strategy that they're using, which requires, you know, mapping all the areas that they drive in.

Starting point is 00:15:39 It requires a lot of remote supervision. Not sure how many remote supervisors they have now, but I'm guessing it's on the order of a half to one per car. You know, I just don't think this scales nearly as well as a strategy that we're using, which is far more end-to-end. What do you mean by remote supervisors? So it's hard to get exact numbers on this sort of stuff, but I would guess that they have interventions by remote operators that take some amount of action to fix mistakes at least once every 10 rides. And so it's not clear to me that their strategy that they're applying now, even though they do not have drivers in the car, necessarily scales that easily to actually having a,

Starting point is 00:16:18 you know, really, really autonomous fleet that doesn't require humans in the loop, essentially. So there are humans somewhere looking at the cars driving. in your mind, it might be one to one per vehicle or one to two vehicles. That would be my guess, yes. And they are not driving the cars, obviously. We actually recently had a startup on that is doing remote driving of cars, like a video game over 5G. Pretty clever. If you've got good connections and seems to work pretty well for dropping off a dropping off and also training, you know, like a Hertz car or something like that.

Starting point is 00:16:56 But with Waymo, you think there's a large number of people. Obviously, that would be very expensive to have, you know, a human being, you know, split watching two cars. That's just like having a driver essentially because he's probably well-paid people in an office somewhere. So you have that overhead. So, and then they use LiDAR as well, which adds a certain expense. What do you think the economics are in terms of running one of these Waymo vehicles? I mean, from my understanding, they've got over a billion dollars in burn rate and less than a thousand cars. So that's over a million dollars per car per year.

Starting point is 00:17:35 Now, revenue probably looks on the order of $100 to $150,000 a year. So it's very far off from something that makes sense. And I think some, you know, obviously they can get that down pretty quickly, but I think some of those things will be hard to remove, especially the remote operators, the costs in developing new mapping for all these new areas, I think there are just several issues that will come up that are costly. Like, I don't know what happens now. If someone leaves the door open, does the door auto close?

Starting point is 00:18:03 That sort of stuff, I think, will make the unit economics essentially not realistically come down to the, you know, 100,000 a year that is required anytime soon. I've been using Tesla's autopilot and FSD since inception and was using this morning. I get an intervention, I would say, in the backroads here in Texas or on the highway, once every, I don't know, 20 to 30 minutes. So it feels like it's doing a pretty great job on straightaways, easy turns, roundabouts. It feels like it's a little jittery. Left turns into traffic feels a little jittery like it's figuring some stuff out. But it does feel like it's getting more confident every year.

Starting point is 00:18:49 talk a little bit about Waymo's approach versus Tesla FSD versus what you're doing in a comma. I mean, so Tesla is definitely a lot more similar to us. And if I were to place a bet on anyone, it would be them. They're recently very focused on end-to-end machine learning just like we are. I think they've not quite rolled out as end-to-end of a strategy as we have. I think they've got some more classical stuff in there. But to be fair, they also have capabilities that our system does not have.

Starting point is 00:19:15 And I think it is harder to switch to end-to-end when you have these more capabilities. these like they can do, you know, left and right hand turns and stuff like that, new turns, stuff that we cannot yet do. But they're a little bit less end to end than us. Our system is completely end to end to end that we ship today. And Tesla is also working on generative AI simulation, presumably to one day train in. I don't think they do that yet. I think we would have heard about that if they did.

Starting point is 00:19:40 There is a lot of similarity to our approach there. You know, they also have a product. They have a very large fleet. You know, Tesla has the most miles collected on. any kind of autonomous system. We have the second, and then Waymo actually has quite a bit less than any of us. So, yeah, I think we're much more similar to Tesla in that sense, much more end-to-end. Waymo has really seemed to have pigeonholed themselves in this LiDAR sensor strategy.

Starting point is 00:20:05 You know, they don't seem to have any interest in moving away from LiDAR, which I think is a mistake. The world is made for eyes. Yeah. You know, the arguments they often use is that, you know, it's like more redundancy. it gives you information about the world that cameras could never do. But ultimately, the roads are designed for human eyes and good modern cameras can do everything human eyes can do,

Starting point is 00:20:27 if not more. And so there's absolutely no reason. You can't perfectly drive a car, at least safer than most humans, with cameras. It is all a software machine learning problem. And I think using things like LIDARs gives you short-term gains, but are long-term, essentially, bottlenecks. So I think it's a detour. I think they'll regret that.

Starting point is 00:20:52 What does it cost do you think for them to put LiDAR on these cars? I had heard in the early days $20,000 $30,000 per car. I don't know if that's still accurate. Last I heard they're paying $120,000 for their cars. And I think the cars themselves cost about half that. So I think the entire upgrade must be on the order of $50,000, I think. And then Tesla's, you think, is a couple of thousand dollars, and yours obviously is $1,500. So it can be done for a lot less.

Starting point is 00:21:26 Yes. I mean, also, 1500 is what we sell it for. You know, we build the devices for half that, and, you know, same for Tesla. Tell me about the cameras you use versus Tesla's, because when you say, like, hey, we should be able to be as good or better than a human driver. humans only see in one direction. They get tired. They have glasses. You know, there's blind spots.

Starting point is 00:21:52 If you have cameras all over the car, you're literally, could be behaving like maybe six, seven, eight human beings in terms of your field of view. And then in terms of accuracy, the fidelity of cameras is better than human eyes now. And I would think it's obviously more vigilant than humans. Maybe doesn't need a cup of coffee. It's late at night. Yeah, I mean, not being distracted, definitely, you know, when we start comparing safety, when we get actual competent self-driving systems, you know, that's where the advantage is going to be.

Starting point is 00:22:23 No distraction, no drunk driving, no sleeping. I think we're not even quite there yet. We need higher capabilities before we can really improve on that. And as to your comments on cameras, I think it's a distraction to talk about cameras. Even this webcam that I'm using now, which is not a great camera, you know, can let you drive a car pretty well if a competent human was operating behind the wheel with that camera view. There are some things that more cameras will help you with. And for a company like Tesla, I think it completely makes sense to install those cameras. For a company like us, the added hassle

Starting point is 00:22:57 of installing more cameras around the car is never going to give the upgrade in performance to make that work. How many do you use when you do it? Just the front facing one? So we have two cameras facing. Actually, I have a device here that I can show you maybe. So this is the different. device here. And so we've got a narrow camera and a wide camera to the front. So it's 180 degrees and 40 degrees. And then on the other side, we have a driver facing camera that makes sure that you're paying attention. So three cameras total. And what about like on the sides of the vehicle and the reverse cameras? Those could help with changing lanes, etc. So how do you think about lane changing and the next version of your software and hardware? So our device has, you know, with the two 180 degree lenses has 100, 360. the degree. So you can see the blind spots. But currently the lane changes are supervised. So you initiate the lane change. You're expected to check the blind spot. And most cars that we support have a blind spot sensor that we can also look at. And so when there's a car in your blind spot detected by the blind spot radar, it will prevent the lane change. But it is a supervisor.

Starting point is 00:24:02 You're expected to look as well. Listen, a strong sales team can make all the difference for a B2B startup. But if you're going to hire sharks, you need to let them hunt, and you can't slow them down with compliance hurdles like SOC2. What is SOC2? Well, any company that stores customer data in the cloud needs to be SOC2 compliant. If you don't have your sock too tight, your sales team can't close major deals. It's that simple. But thankfully, Vanta makes it really easy to get and renew your SOC2 compliance.

Starting point is 00:24:30 On average, Vanta customers are compliant in just two to four weeks. Without Vanta, it takes three to five months. Vanta can save you hundreds of hours of work and up to 85% on compliance. costs and Vanta does more than just SOC2. They also automate up to 90% compliance for GDPR, HIPAA, and more. So here's your call to action. Stop slowing your sales team down and use Vanta. Get $1,000 off at vanta.com slash twist. That's vanta.com slash twist for $1,000 off your sock two. What is your handicapping of the space? When will we see this rollout, you know, in a major way with multiple vendors in many cities?

Starting point is 00:25:09 you're talking about like a Waymo type taxi solution. Yeah, let's say no human in the in the driver's seat. We've established now that the majority of miles can be driven safely or safer with a human plus a level two, three system, four system, whatever it is. I guess the question for everybody is when do we remove the expense of the driver and have, you know, these fleets of cars everywhere, driving people, and burritos to their destinations without the expense of a driver. So I'm generally a lot more pessimistic than I would say the average. I think there's a lot of hype in the space. I think most of these things are generally overhyped.

Starting point is 00:25:52 I think the best thing to do is to look at the orders of magnitude of mistakes and kind of see how that's been trending and extrapolate that. I think exactly what you're talking about is relevant. You have a disengagement that maybe is safety critical, maybe not every few drives, let's say, I think the Waymos are similar. They have a bit of a different strategy, but they have remote supervision, remote intervention. Let's say a disengagement that is necessary every 10, 20 drives. You know, that's very far away from a system that can drive reliably day after day with absolutely no supervision. So I think you should look at the trends and kind of extrapolate

Starting point is 00:26:27 the orders of magnitudes of mistake. And we're still many years away, I think. Okay. So many years being three, four, five, six, seven, somewhere in that range? I think predicting past five years is so hard. It's not next year. It's not going to be the year after that. I predict within five years, there's going to be nothing that looks like a self-driving taxi solution in most cities. After that, I think predictions are so hard.

Starting point is 00:26:51 Why hasn't a major car manufacturer, the Toyota's Hondas of the world, looked at what you're doing in the open source project and just said, hey, let's go all in on open source here? That would seem to me to be a tipping point for the industry. If we really want to save lives, why not, you know, why hasn't Cruz open source what they're doing, or Waymo or Tesla or one of these? I had the co-CEO at the Olin Summit last week, and she said open source isn't a discussion at Waymo. So it does seem like open source tends to win in the long term because of the reasons you stated. But I'm just curious why there isn't a major open source project.

Starting point is 00:27:33 You do have open maps as a data repository, I believe. I'm not sure if you use it or if it's relevant here. But it would seem to be... Yeah, we use it open street maps. Yeah. And so maybe you could explain a little bit about that project and how that helps you. And then there's all this open hardware that exists in the world now and all kinds of libraries. Why hasn't an open source self-driving project kind of taken hold across many vendors yet?

Starting point is 00:28:01 So first of all, like I said, I think. think companies don't open source their stuff because they want to overhype what they have. Open sourcing means making clear what you have. And I think companies like Waymo aren't too excited about people finding out how many interventions they actually have about people finding out how much work it actually takes to do a lot of these things. Because, you know, that's their revenue stream, is investment. And if people have less of an opinion of where they really are, that is not good for, you know, their financial situation. Tesla, on the other hand, they're not open source. but they're relatively open and transparent about what they're doing and what the system does.

Starting point is 00:28:37 And you can use it at any time and you can test it in any conditions that you want. So it's not open source, but at least it's transparent. And then as to why these legacy car manufacturers don't take our system and just implement it and ship it, because it's a lot better than theirs. I mean, that I think is a great question, but it's a question for them. I think generally speaking, these companies are not interested in innovation. They run defensively and they act out of fear. if they see that their business model is under threat, they will respond and try to reduce that threat.

Starting point is 00:29:07 But when there is a system that is a clear upgrade to them available, that doesn't seem like an immediate threat, they just generally have no interest. I mean, it's the same thing with their infotainment systems. You know, you use infotainment system of even a modern car from a legacy car manufacturer, and it feels broken compared to your iPhone. There's no excuse for that.

Starting point is 00:29:27 They could fix that, but that's just not how these companies work. I think the bigger question is, why does companies like Lucid or Rivian perhaps not? You know, they're developing their own system in-house. They have hardware that can run OpenPilot and they're shipping solutions that are worse than Open Pilot. I think they'd be a great candidate to implement Open Pilot on their car, at least while they're developing their own solution. If they can make something better, sure, replace it. But in the meantime, why not just use our software? It's free.

Starting point is 00:29:51 It's MIT licensed. They can get something better running today. Yeah, it would seem to me that if you're behind, and classically, this is what we've seen, when a corporation is behind, they embrace open source, and when they're ahead, they embrace closed source. Google is a great microcosm of that. Android, they were far behind on the smartphone market, they open source it, search, they were far ahead, they kept it closed, Facebook, the social graph is closed because they're so far ahead and they have locked in, and then they just open source Lama and they're so far behind on AI that they decided to

Starting point is 00:30:27 open source. It's not key to the business. So, It would seem to me like a Rivian, a niche provider of vehicles would do so much better to partner with you. Have you talked to them or reached out to them? I mean, we're not really interested. Like, we have very limited resources. We don't want to invest resources into partnering with anyone. We do everything we can to make our stuff accessible, open source, and a company like Rivian, if they invest in the time, could easily port it to their hardware. I think it's like a stuff made here kind of thing.

Starting point is 00:30:55 They want stuff built in house. That's what I think. There's some limitations to our software too that they may not like. We don't do A, E, B, yet. They might not be interested in a solution that doesn't do A, B as well. That's something that we're working on.

Starting point is 00:31:09 But, you know, we've talked to some of these people. There is some interest. We've talked to legacy car manufacturers. There is interest. But it just doesn't align with our goals. We're really focused on solving robotics and we just want to make money and have a product in the meantime.

Starting point is 00:31:22 And anything that distracts from that is just not worth it for us. All right. Scaling product is one of the hard challenges. We as founders face, knowing what to build is obviously only half the battle. You also need the resources at exactly the right time. You might want a little help now. You might not need it later. And that's where micro one comes in. They're going to help you scale your product with a team of developers in days, not weeks. When you're handling the search process, when you're trying to get great developers, the vetting and the onboarding and the paperwork, that's going to take

Starting point is 00:31:56 weeks for each one. It's painful. It's one of the hardest parts about being an entrepreneur. Well, imagine if you just tell AI what you need on an engineering basis and then AI delivers with the team at Micro One exactly what you need, you're going to be in great shape. And Micro One built this incredible AI engine. This is why I invested. They interview 20,000 engineers every month. They pick the top 1%. And then they help onboard them into your company. It's simple and it's effective. They do all the legal. They do the compliance. They do the search. You don't have to do a mountain of resumes. You don't have to worry about somebody flaking out. You don't have to budget a huge recruiting bill. Micro 1's got you covered. And you'll find your next engineer in 48 hours or less.

Starting point is 00:32:35 That's the promise Micro 1 has for you. So here's the call to action, micro1.aI slash twist. M-I-C-R-O-1.a-I-slash twist to instantly ramp up your product team. And if you make a hire before October 1st, they're going to give you two weeks of free development per hire. Have you been tracking what's going on in China, there are some places in the world where they might look at self-driving and say, hey, even with one intervention every 10 or 20 rides, that's safe enough for how we look at safety in, let's say, Beijing or China, where like, you know, let's face it, people in factories there might have, you know, OSHA in America might be a very fine filter, you know, in terms of working conditions and in China

Starting point is 00:33:23 other places, maybe they're just like, you know, people are a little bit more expendable. We don't, we want to have a society move faster and rather than, you know, have very niche safety needs. I'm trying to be generous here, but they've got six or seven

Starting point is 00:33:40 different players on the road with technology that I think is similar to yours and Tesla's, yes? I'm not super familiar with what's going on in China in terms of self-driving. I think when you're talking about things like rolling out these systems. With the technology where it's at today, you cannot make a profitable taxi service.

Starting point is 00:33:58 That's just a fact today. And I think we're pretty far away from that. And so companies like Waymo, when they grow, they cost more money. And that's just not a, to me, seems like a terrible strategy. And so I'm not sure what the benefit would be of doing this in China, even if they're okay with the additional risk. It's not clear to me that that accelerates progress. I think these technical strides need to be made and they don't need.

Starting point is 00:34:22 that much data or they don't, like, we have more data than Waymo. Yeah. What do you think the Tesla announcement will be? You think Elon will show kind of a two-seater that people, there were some leaked photos on social of like, hey, maybe there's going to be a specific physical robotaxy debuted. That seems to be the case. And then do you think it will have a safety driver when they roll it out, or is their technology ready to do autonomous rides like Waymos does?

Starting point is 00:34:49 I think your experience with FSD is probably very reflective of how good Tesla's autonomous software is. I very much doubt that all of a sudden there's some secret project that is capable of doing actual taxi service. Elon said in 2016 that they were going to drive

Starting point is 00:35:04 coast to coast self-driving at the end of the year. That didn't happen. I think generally speaking, Elon's a bit optimistic. I think this is probably along those lines. So when do you think if you had to take a guess

Starting point is 00:35:19 they would be able to take the steering wheel out with FSA, just to make a while, I guess. I'd say again, over five years, under five years. Five plus years. Five plus years, okay. I would say, yes. As we wrap up here, tell me what your vision is for robotics. Obviously, you have Humane, you got Tesla doing Optimus.

Starting point is 00:35:38 We just had Sergei at the All-In Summit, you know, sort of, I think it was lamenting a little bit. They were early into robotics before AI was there. You know, cars. go very fast. It can cause a lot of damage, but a robot, you know, if it's weighs 50 pounds or 100 pounds, it's not going to do a lot of damage if it falls over. It's not going to certainly be going 65 or 75 miles an hour when it makes a mistake. So tell me a little bit about what you think the future of robotics is, given what you've learned in AI and what's your approach for that.

Starting point is 00:36:10 I mean, I'm very excited about robotics. That's, you know, having robots in your house that could do your laundry or anything like that, that is the coolest thing ever. And, you know, I think about that all the time. But I think we should be realistic. The big reason a lot of people went to self-driving 10 years ago, including me, is because it seemed like a great applied robotics problem that was easy. You have two-dimensional actuators, you have simple rules of the road. As far as robotics goes, that's relatively easy. And then what happened is all these companies were optimistic and ended up not reaching their goals. And now all of a sudden, everyone's switching to humanoid robotics, which from the beginning, we always thought was harder. So I think it's just another hype wave. I don't

Starting point is 00:36:49 think there's going to be humanoid robots in your house. I have a robot vacuum. I think that's kind of the state-of-the-art robots you can buy in your house today. And they get better every year, not super fast. But I think seeing that trend is what you should be thinking about, about realistically, what's going to happen. Those things will get better, but you're not going to have humanoid robotics in a couple of years. I think it's just another hype cycle. And we should be somewhat cautious about everything that's just a demo and is not shipping. Got it. Yeah, it does. And what do you think the first applications will be in robotics factories and doing very specific narrow

Starting point is 00:37:24 factory work versus hey, you know, this thing's walking around the ranch, you know, going and cleaning up, you know, horse poop and putting hay out for the horses. Yeah, I mean, I think it's just going to be along the lines of what we've been seeing, right? There have been factory robots for a long time. I think they'll become easier to program.

Starting point is 00:37:43 They'll be able to do more things without needing to make custom hardware. You know, we have robot vacuums, robot mops. I think someday they'll stop eating your cables and they'll stop eating your socks. You know, there's robot lawn mowers. I think, you know, we should look... I did see one of those in Texas. I was driving by and somebody, or I was walking by, rather, I just parked and somebody had one on their front lawn and it's at night, it's got the light on and it's out there at night running because I guess it's too hot here during the day in Texas. Yeah, exactly. I mean, it's great. I think, you know, I think those are the things we should

Starting point is 00:38:13 be excited about. We should be excited about the things that people are shipping, not the demos we're seeing. And those things are getting better, and I think they'll continue to get better. I'd be excited if in five years, my robot vacuum doesn't get stuck anymore. I have a robot mop and maybe, you know, something that can fold my laundry.

Starting point is 00:38:29 But I think we should dampen expectations from this human light stuff. Is it going to be an open source project as well when you start doing the robotic stuff? Yeah, so OpenPilot is, on the one hand, an open source robotics operating system. And another hand, it's a, you know, system. And we're kind of working on splitting those out and the self-driving part is going to be one

Starting point is 00:38:51 application. We imagine there's going to be models running that are kind of world models that have a general understanding of video and physics and how the world moves and more and more applications will work. I mean, we have a very impromptu robot that we built a while back that's in the background here, which we call the comma body. It's just a bunch of wheels or our device and we will be more interested in getting into that, if it's feasible with end-to-end machine learning to make something that navigates around your house or your office without getting stuck and without doing anything stupid. And today, that's actually not that easy. Yeah, there's a lot of detrius around most people's houses and things change pretty frequently. Exactly. Kind of the opposite of a highway

Starting point is 00:39:35 where you just have cars and nothing else. Well, listen, continue success. And where can people find out more about the self-driving project and also the open source project. Yeah, so I mean, our website, Commodore, AI, if you want to check out our device, you know, try it out. Don't listen to what other people are saying. If you don't like it, send it back. And, you know, our GitHub has all of our open source projects and open pilots on there. You can see what we're working on. We don't do anything in secret. If we're not publicly sharing it, it's probably not something we're doing. I mean, I love the idea of, yeah, I would love to see Waymos code base and,

Starting point is 00:40:12 understand how these mission control specialists actually interact. I know that was like a big controversy for them when I had mentioned it previously. They seem a little bit upset about like people even discussing that there could be interventions or crews. And then what are the interventions that are occurring? I think some transparency there would be good. And I think regulators now are, you know, very interested in double clicking maybe and seeing what's under the hood, right? Yeah, no, I think so. I mean, I would love to see more transparency. It's something we really strive for and, you know, that's the best way to do it, I think. Yeah, regulators, if you're listening, I think all interventions should be reported in public. I think that would be a good starting point, right? Like, if they had to keep a log of interventions, share the interventions, I think also sharing the videos of any intervention that occurs, you know, with regulators to review on some regular basis because

Starting point is 00:41:11 you know, it seems to be one of the great second order effects of what you're doing is you're going to be able to tell regulators and city planners hey, this is where stop signs need to be. This is where red lights need to be. This is where the speed limit could be higher. This is where the speed limit should be

Starting point is 00:41:29 lower. And they don't actually have a way of you know, in the real world getting tens of millions of miles of data and you know for this. Except I think they lay down like a little strip that counts the number of cars going by and the speed of those cars. It's not, it's pretty, pretty dumb information.

Starting point is 00:41:47 Yeah, no, they definitely don't have modern data gathering techniques. I've got a map open here of our cars that are driving of the last week. I don't know if you want to see that. Oh, wow, yeah, show me that, yeah. I love a good visualization. Yeah, so here you can see kind of this is, I think, last week or last 30 days. I'm not exactly sure. Oh, last 30 days.

Starting point is 00:42:03 Yeah, you can see. I mean, it's pretty global. In the U.S., we've got really quite good coverage, actually, of basically all the urban areas. and it's a bit more sporadic. The areas you don't have in the Midwest are simply because we don't have population there and there's a couple of mountain ranges there. Yeah, exactly.

Starting point is 00:42:19 Some of those arteries you're seeing are the ones that go through the Rocky Mountains and the Sierra. Exactly. It has much to do and population density. You know, when you see Florida and California and the northeast lit up, there's a reason and you see people driving to Tahoe.

Starting point is 00:42:34 That's really, you know, a powerful visualization. You got a couple of people in Alaska using it as well. And these people are... Hawaii. These people are hobbyists, yeah? And they're technologists who really are thinking

Starting point is 00:42:50 about the future of this technology and they want to contribute to the project, or are you finding, like, you have corporations using it for some reason? There are definitely some corporations, I think, generally using it out of interest to compare with their own system, you know, a lot of, like, people that are working on ADAS.

Starting point is 00:43:07 Yeah, most of this are just users, You buy the device. It takes 20 minutes to install in your car. And it makes your life bit easier if you're doing a lot of driving. Yeah. And yeah, it looks like you're popular down under as well. And when you see Australia, it's a very large landmass. People don't understand how big Australia is and how not populated it is.

Starting point is 00:43:26 When you go to the west of Australia, there are signs that just say, like, there's no coverage here. There's nobody coming to help you. make sure you have water, extra tires, extra food, extra jacks, a satellite phone because, man, those deserts out there are barren and they're barren for, you know, days and days. If you get caught out there, you're dead. You will, you will die. But the great Australian desert here, no data from there yet, unfortunately. I mean, if there's data from there, if you were to take a car there, I've watched some videos of people, you know, driving through that area in Australia. The key thing is, like, how do

Starting point is 00:44:06 much weight of extra fuel can you bring with you on your car? They're like adding, you know, half the car is filled with gas canisters basically when you're driving across because there's no gas stations, folks. You're going to die out there if you go and you run out of gas. So, yeah, it's a real adventure project. Oh, this is interesting. So here's your, on these metrics, I'm assuming, are public that you put out? They're not public, but I mean, we're not secretive about them. But yeah, Yeah, we can see here. There's just some general dashboard we have. You can see how much percentage of miles are engaged in the fleet right now. It's a bit over 50% how much time of the driving is engaged. People love to use it on the highway, I assume, right? That's like

Starting point is 00:44:47 super lower. I mean, my fatigue level goes way down when I was driving between San Francisco and Tahoe using my Tesla. I mean, when I would drive my suburban, which is my go car, if you know, batteries don't work out and it's the end of the world. And it's an apocalypse. I like to have one of each. Man, I mean, my fatigue level from one car versus the other, just staying in the lane. And then also people in the car prefer when I'm using FSD, I find because less motion, right? It's a better ride.

Starting point is 00:45:22 Yeah, no, I hear that a lot. My wife always says, well, did you disengage? It feels much worse now. Yeah. Well, I mean, that's very specific to you. And you're, I mean, you may be making a goal. great system for self-driving, but she has complained to me about your inability to stay in the central lane. More work to be done there, Harold.

Starting point is 00:45:39 Exactly. Listen, I appreciate you coming on the program, and we'll see you all next time on this week in startups.

This Week in Startups - No hype, Just works: How Comma reached 100M miles in autonomous driving | E2011

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.