comma ai | George Hotz | An OS for Autos | Self Driving Cars? Scam!

comma ai | George Hotz | An OS for Autos | Self Driving Cars? Scam!


Of the time and I’m pretty sure it’ll happen over the next two days That said I’d like to introduce our first speaker George. Do you want to start coming up? George Hotz Is a pretty interesting guy He had his first 15 minutes of fame as I understand it hacking the iPhone as a teenager. Is that right? George? and And it’s George will correct me if I’m wrong in any of this And his second 15 minutes of fame when he went into his garage and came out a couple of months later having hacked his car to be self-driving that set him off on a project that he’s still in the middle of creating an operating system for Self-driving both retrofitting existing cars and for new cars and and that’s what’s gonna tell us about today George Hey Oh Okay, I’m good When I talk I try to be really concrete with words, um, if I say anything that doesn’t sound concrete, please ask me a question Last couple months ago I started working with like formal languages like talk and Isabel and everything could be specified rigidly, and I think a lot of Definitions for a lot of things are just thrown around and there is no formal definition of them And then it’s a question what you’re really saying. A lot of people can have conversations where two people just talk past each other Because they have different definitions of a word But you know some disciplines like math have found beautiful ways to solve this by giving things formal Definitions so if I’m informal about anything, I mean it’s obviously an informal talk, but I’m happy to clarify So when you think of self-driving cars, what’s the first company you think of I? Hear a few Tesla’s I also heard way moe I think I think yeah way mo is way mo is probably the one that captures the Perhaps national consciousness of what self-driving cars are or the other one people mention is uber uber? Actually got the most press about self-driving cars. If you do a Google Trends search for them. It’s when uber killed somebody so but the traditional Maybe this is just because a lot of marketing dollars have been poured into it The traditional idea of self-driving cars is the way Moe. It’s a car with no people in it and You can call it from an app that it will drive down your idyllic suburban street and it will pull up on the curb next to you and it will flash its little lights and be like way Moe is Waiting for passengers to get on board You could see this in the Google ads and it does not reflect the reality of driving At all if you if you guys have seen a way Moe ad Yeah, but you know so you pull up oh look everyone’s so happy they’re riding around in a way Moe Way Moe is valued at a hundred and seventy five billion dollars despite Not only not making profit, but I believe never making any revenue So this this notion of self-driving cars may be distinct from Tesla is known as and we’ll get to Tesla But it’s known as level four Maybe this is the term that the levels don’t really mean anything, but that’s just the term people use What it technically means is a car that’s capable of full autonomy in restricted areas so full autonomy in restricted areas the Application people immediately think of is ride-sharing Ride-sharing is only a very small percent of rides people who? Google’s notorious for this if they don’t think about what Life is like for most people most people don’t live in the Bay Area Where they can call an uber and have an uber arrive in three minutes I’m from suburban, New Jersey very nice area, but you call an uber. I’ll be twenty minutes So ride-sharing is only a very small percent of rides most people still in cars and Level four self-driving is a strictly worse product So if ride-sharing Can’t even capture these markets. How do people expect self-driving cars to be able to so first and foremost Anything that’s level four. So also level four or says something about usually how it’s technically done way Moe uses High Definition maps a very high precision like precise to the centimeter maps of an area Uses lidar that can register every object up to three hundred meters away This is not how humans drive at all But because of these this HD map requirement they have a smaller service area They’re also slower So way most travel the speed limit they come to full stops at stop signs and They slow down at the slightest hint of hesitation Which may be if you’re trying to design your thing to make sure there’s never a PR disaster for your company like uber It’s the right call But if I’m sitting there, oh man, y’know I gotta get downtown or home way. Mo is 18 minutes uber saying 12 I’m gonna take an uber There also, more expensive So this might change this will probably change in the future everything eventually gets cheaper but right now when you look at away, mo those cars cost about 300,000 to 500,000 to retrofit Again, Google thinks oh, yeah. Well, we pay all our software engineers 200k a year salaries uber drivers ain’t makin 200k. Um, so They’re they’re they’re they’re more expensive the Autonomous vehicle vision wouldn’t even be a revolution at best It’s cheaper ride-sharing. A lot of times people get caught up in notions because they see Futuristic looking stuff like this and think that this is going to have some revolutionary impact But they don’t think about what the exact same product would be with a human so all those chatbot AI companies criticized like They’re just humans doing the chatting, but they put AI in the name and they’re there for some reason their valuation is way higher It’s a scam So who’s profiting off this game This is a company drive day I So they use all they’re all you read their medium posts aw They’re unbelievable with the well, so we want to make sure that our self-driving car is capable of operating in multiple areas So we hired engineers all over the world What does have anything to do with that? The very very kind of lofty language Then when you get down to the root of what they actually made they made like a neighborhood shuttle bus You know those like like like those buses that that ride around retirement homes that look like this It’s not self-driving. There’s a human in the driver’s seat. They made a 77 million dollar Retirement home shuttle bus. What even is the global market for retirement home shuttle buses I don’t know anybody Consulting with people about the ethics of autonomous vehicles all well So if I’m driving down the street and and there’s a baby in the way But there’s a tree do I swerve and kill the driver to save them has this ever happened to a human? Has this ever been a real scenario? Cars are incredibly dangerous and 90% of accidents are caused by drivers who are drunk distracted or asleep not because they swerve to not hit a baby yet this captures again a conversation about self-driving cars that’s completely disproportional to any Remote impact it may have but a lot of people are making money off this and they’re there branding themselves It’s like well, we’re gonna consult with the city of Beverly Hills about their autonomous vehicle strap. What? This stuff’s not a thing. I’m in this one So in the in the gold rush so you look at the the cruise acquisition Probably actually a good vote for GM but then it was all the follow-ons to the cruise acquisition this was the Delfy spent four hundred fifty million dollars for demo quality software that drives around a block with a human of a driver’s seat This was done by Stanford in in during 2011 and the DARPA urban challenge and even better They open sourced their code They paid four hundred fifty thousand forty fifty million dollars probably for a few engineers and some code that’s effectively open-source So what’s not a scam we gotta have the the positive atom actually sit here and rant about all the scammers for for an hour This is car sold in America Personal vehicle ownership is on the rise and not going anywhere for a long time Might convert to electric. I think it will Give it time for the infrastructure to proliferate but personal vehicle ownership is not going anywhere because How much cheaper would self-driving cars even be the nuber how much of the cost of the ride-sharing is the human driver? About half the other half is maintenance of the car and in fleet and fuel and lots of other things So it would even make goobers that are twice as cheap, but uber tries this we can already do twice as cheap uber today It’s called subsidies and they still, you know, can’t get more than a few percent of a lot of these markets This is US car ownership. This is Chinese car ownership Huge potential market car ownership will continue for a long time like we’re not going to Rethink and revolutionize the way that we have cars people like owning cars for the same reason people like owning houses So if people are going to continue to buy cars, what does a better car look like? You think of cell phones You go back people had cell phones in the early 2000s, but no one really had smart phones So what does there’s a benefit Kevin’s quote, right? What does a smart phone on wheels look like question down front on your right? Would would this change if it was Mixed use in the sense of if you want to get on an interstate. It will only be with the driverless car So you flip a switch or something and then when you get off it goes back to human driving. Well, so Yeah, I mean you can build that. We’re already seeing systems like that today with test auto pilot We’re seeing systems where they still require monitoring But they can do a lot of the work on the interstate and I do think this is the direction that we are headed When it comes to we’re going to rezone interstates as autonomous vehicle only zones. That’s I mean Some legislature person can might talk about that kind of stuff. But in reality, there’s no reason to do it Yeah, maybe you have you know, this is the autonomous Lane kind of like the HOV Lane, okay, I could see that happening But does that answer your question? So On a lot of new cars today in fact, it’s becoming standard on a lot of manufacturers you’re starting to see driver assistance technology So don’t view Autonomous vehicles as a binary like there’s either you know A human driver or there’s no human driver all of those way Mo’s driving around Chandler Arizona still have a human in the driver seat That still has to be there to intervene in case the car makes any mistakes So instead of thinking about it like a cliff think about it as a gradual as a gradual increase to full autonomy So we had cruise control cruise control came out in the first cars before in 70s. It’s automatic gas So when you’re on the highway instead of having to keep your foot of gas pedal you press the button and the car will will Handle the gas for you not so great if there’s another car in a way Because it will slam right into that car if you set the cruise control at 65, so that’s why we have adaptive cruise control How many people have cars with adaptive cruise control? Cool a lot of you It’s common in mice So it’s the car will the car will do the brake for you and will keep you a safe following distance from the car in front of you and then lane centering as distinct from Lane Keeping Assist Which will correct you if you’re drifting out of the lane Lane centering Actively keeps you in the land how many people have a car with Lane Center? Okay less. So if you guys didn’t see the hands it’s about 50% you with adaptive cruise control about 20% with Lane centering All of these combined gas brake and steering that’s all of driving a Lot of Manufacturers were selling their cruise control systems and their dapping cruise control systems in the lane keeping system All kind of on separate pages if you’ve ever tried to navigate The OEMs website. I think Tesla did a great job with the branding on this when they came out to test the autopilot Tesla autopilot is just these same features but packaged into one Nice Oh Test autopilot. Okay, cool. What is autopilot? Oh, it drives the car for you. Well, that’s not really what Auto powered on the plane does But interesting interesting pyaare around it, they know how to be provocative. So, this is a Paul Graham tweet The Tesla Model S was like the iPhone And that all the manufacturers now have to copy it. The difference is many don’t realize it yet So when these things come out Even blackberry is scared. There’s there’s initial articles about all the iPhone will never take off because people want a physical keyboard Blackberry doesn’t think this blackberry starts working or if you guys remember the blackberry storm they tried they tried to they tried to stay around but they didn’t I Spoke with someone I believe it was from from Mitsubishi and they said there’s no way that we can now build electric cars Because because Tesla is Tesla is so far had on the battery technology. They have a 5-2 lead in the battery technology There’s no supplier. You can go to to buy Tesla’s battery from but even if you built a car with the same range and specs as a Tesla nobody would buy it because they’re also five years ahead in a Huge number of other areas like the operating system like the supercharger network So, I don’t know what’s gonna happen with it with what the car industry but This is kind of where we come in I was approached by a friend To build a new to build the vision system for Tesla autopilot before it launched Tesla was using a mobile eye chip. So mobile eye is a really the original a task company and I’ve come to respect them a lot more throughout this journey because they’ve been saying a lot of His self-driving cars a scam stuff for a long time And they are they were acquired for fifteen billion dollars by Intel Not a bad price. Sorry Intel Botham name an Intel acquisition that’s gone on to be a successful company You get bought by Facebook. You’re successful you get bought by Intel you’re dead Uh-huh Thanks for everything though whatsapp Instagram oculus Let’s go sold a million of those things so I’m working I met with Elon a few times who talked about a deal for building the vision system But eventually the deal kind of fell through I’ll chalk it up to a miscommunication And I come up with this idea for a company. Okay build better test the autopilot sell to automakers be rich Just pitch that I get this is the pitch that I gave to VCS and it sounds like a good pitch. I talked about the smartphone stuff and I’m like look, I mean Tesla’s running iOS We’re gonna build the Android So we did we did step one. It’s called open pilot Um open pilot. I have a little explainer video made by one of our Community members after this but it’s effectively autopilot for cars people actually have So this is this is a little what is open pilot open pilot is an advanced driver assistance System that turns your car’s lane keep assist function from this To this as well as provides smoother and safer radar adaptive cruise control Open pilot allows you to be driving hands-free and feet free for long periods of time without intervention even more so on interstates and highways Everything is controlled by open pilot, but it’s easy to intervene at any time. But how does this all work? You need three pieces of hardware all bought from coma ai slash shop first the giraffe This device converts your car’s lane keep assist camera connector into a more Universal port for the next piece of hardware Which is the gray panda the Panda receives information from your car and can also send information as well This is what sends the car commands such as speed up slow down and turn the wheel in either direction It also has a high-precision GPS antenna used to create HD maps as you drive Finally the e on this is the main device that drives your car first plugged into the Panda then mounted on your windshield it sees your car’s existing radar information which shows exactly how far away the car in front of you is With this it instructs the car to speed up and slow down as needed it also has a high quality camera that looks for lane lines in the road and tells the wheel to stay exactly centered in the lane lines at all times As an added bonus. It is also a fantastic dashcam with free unlimited cloud storage Open pilot runs on Honda’s Toyota’s Acuras Lexus GM in Hyundais with even more car support being worked on every day and the best part Altogether this setup costs less than $1,000 check if your existing car is compatible then go buy a neon panda enter after day and upgrade your car to the future That’s our slogan So Back to the the the the the business plan build better Tesla autopilot. Well, I really think you’re not really gonna build better test autopilot because Tesla autopilot is Not a fixed target It’s much better right now than Tesla autopilot when it came out Is it better than Tesla autopilot now? No Because Tesla’s good engineers when I saw what Tesla was doing with autopilot like When I compared myself to the way that like Toyota and auto manufacturers and even the tier ones Delphian Bosch are thinking about this They’re thinking in old ways. They’re coding in. Well, we have to code everything in Misra C and run it on 32-bit freescale trike or micro processors that have as ‘old these safety ratings and This doesn’t really make sense You’re not gonna get Functional safety, you’re not gonna be able to get a safety spec of these things because there’s no formal definition of a car You’re just gonna have to use statistical safety You’re not gonna want to code all of this in You know this this very restricted C because you’re never gonna get any sort of modern performance You’re never gonna get all the niceties OBE how it’s gotten in Python, by the way You’re never gonna get you’re not gonna be able to be competitive Tesla autopilot is coded in C++, but it’s a very Nice high-end abstracted C++ not the kind of low end stuff that the car manufacturers are doing so Tesla is using neural networks. They’re using GPUs. We’re using neural networks. We’re using GPUs We’re using all the same tricks. So we’re not actually going to pass autopilot. We’re gonna kind of tie him So we do this right and I hadn’t thought that the reason that other cars didn’t have autopilot Was that the car manufacturers did not know how to build it? This is not true So I met with the car makers I met with several CEOs Drove down to their offices. I met with a lot of business development people. I’m still kind of confused about what business development is Oh, I was I was I was appalled You couldn’t you couldn’t talk to these people? There is no I guess it makes sense when you think about it Ford is a four hundred thousand person organization Who is Ford if I had mind control over the CEO of Ford? I still couldn’t make this stuff ship It’s like a big it’s like a big boat and no one’s really skewing. Everyone’s got like a little paddle, right? Question on your right in the back Yeah, so I think I’m your worst nightmare because I’m both a business development person and a PhD in psychology. You’re not gonna take So if you’re appalled now, you should be a little worried. Um my question Listening to the way you’re thinking about going at the problem what you’re trying to solve Makes me think back to your slide with that big red X on it yeah, and in the way I shorthand of what you did was you put a big red X through the category or the the Topics having to do with the ethics and the Interaction between the technology you’re working on in people and I know the baby case could be put off as an edge case I taught probability I get it Yeah, but that’s a placeholder for a whole bunch of issues that aren’t about whether pythons gonna cut it for me or not And so I’m wondering how you think about that as you develop your technology I mean I get look I Understand big companies and challenges moving through but I think there’s a pretty fundamental one when we’re looking at some of this stuff around Assisted driving autonomous driving, whatever part of it you think you want to address? So when it comes to what I put an extra was people talking specifically about the the ethics of autonomous vehicles The handoff problem is absolutely real on the problem of making sure Of just remove from a from like an HCI perspective Understanding whether the human is in control of the car whether the machine is in control of the car making sure expectations are correct Oh, these things are extremely important. I Don’t go into it in this slide deck, but we have some medium posts about we have a camera facing the driver I mean we make sure that the driver is paying attention at all times that we expect them to be paying attention We make sure our our transition system When we how we alert how we do the handoffs has been praised by a bunch of people Auto makers it like beeps and flashes like a little icon. You have no idea what it means We have consistent messaging about whether the car is capable of doing this whether it’s capable of engaging here So I think things like that are extremely important What I don’t think is important is the kind of people who give far reaching twenty years from now We’re going to rethink cities around autonomous vehicles Are we what’s the difference to the autonomous vehicle and a non autonomous vehicle at most its cost? In these cases. It’s also building a product that is more convenient for a user but from an external perspective The car looks identical you can’t tell If you see a test on the highway with their cars on autopilot or not on autopilot, I’m sorry I don’t mean to group those two things together. I think one is extremely important I think one is People self-promoting and stuff that like, I don’t see the relevant stuff, but does that answer cool? So, yeah, you you try to work with these uh with these automakers and You can’t It’s a structural problem Mobileye has more business development people than they do R&D engineers. That’s why their text not that great, but it’s in all the cars So, yeah, I mean to go back to the Paul Graham tweet You don’t realize it. The car manufacturers don’t realize when you get in some of these cars and you use the their adaptive cruise control the Quality varies wildly from manufacturer to manufacturer. It’s almost like these people have never gotten in a Tesla and tried it So yeah, what do you do? So is this hard as hard at the beginning of the company this was like my plan I’m gonna sell to these people. I’m gonna talk with them. They’re gonna be reasonable So I realized that who really cares if Tesla is the iPhone We just have to build the Android and the phones already exist It’s not like we have to get manufacturer by n we just have to get Manufacturer compatibility, I would much rather work with the product than the company Open pilot already supports this whole list of cars you can go right now and you can put open pilot in these cars and it will take over the steering the gas and the brakes and it will do a Significantly better job than the manufacturer system in all of those cars open pilot is the Android and Just like Android It’s open source Why make it open source well Nobody has any excuse anymore You’ve seen it in the last two years how many companies try to raise money off of a demo video of self-driving? nobody Because if you can’t build something better than what’s available and open source and market of it nobody knows that Stanford codes open source You have nothing right? There is no reason for somebody now to start from scratch on this stuff Just take open pilot modify. It build it extend it and hopefully upstream your stuff change the conversation from Intellectual property and business development deals and exclusivity to do a pull request on an open pilot Well, upstream it. It’s it’s everybody’s MIT license So somebody else could take this rebrand it call it their own great. I just want to see the stuff you better Have you tested Open pilot and all of those vehicles that you had previously you have um So not not personally but so we have a we have we have a safety model I can go into after after I run through the slide deck I can go through I can go to the website and show what our safety model is so our safety model involves It’s two things It’s one Make sure that the system can never do anything quickly make sure it can never like jerk the wheel or slam on the brakes or do something faster than a driver could react and to Make sure the driver is paying attention at all times And as soon as they touch either the gas or the brake the car immediately reverts to manual mode So as long as you ensure those two things you have a pretty good level to safety model Um and yet be explicit about this. So we say like it cannot do anything bad within one second So as long as you’re paying attention, you know just reach out and it’s not like it’ll ever like turn the wheel like this It’s really designed for highway driving kind of stuff So if you just reach out put a little bit of pressure on the steering wheel, it’ll just stop doing anything So yeah, yeah making sure this stuff is well done and also the safety code is not a Python on the safety code isn’t see the safety code is is on an as will be microcontroller The stuff that’s in Python is like the machine learning model part of things Which you’re never going to be able to formally validate. Anyway, the safety code is formally validated It has tests it is you got to do it stuff correctly I am Ashley McCorkle from Intel and Just let you know we haven’t killed mobile yet. We’re trying really hard not to that was kind of an expensive acquisition but and You actually have a lot in common in your approach and am nons approach with mobile I and one is trying to scale up from a TAS into more sophisticated Levels of autonomous driving. I think the other is assuming that this will never work if you have to put you know multiple xeon supercomputing resources in a car You know he’s trying to scale it up from basically like a a In SOC like a very light computing and addling thousands rather than tens of thousands of dollars the price of a car Right. So one of his assumptions though, and I think it’s relevant for the Lead up with that we had previously the the warm up on trends in AI is Rather than trying to use a brute-force approach Where you were I think you like you were saying, you know Having 3d lidar mapping of the entire terrain and having these massive crunching to be able to map You know in 3d, you know everything entire surroundings the car Using a very light compute method that is really based not on, you know these deep learning based number crunching but on a rules-based approach and I think he’s also published that as you know, basically open sourcing this responsibility sensitive safety where you’re programming a set of rules into the AI that that Basically encapsulate what society’s idea of safe driving is in by society It can be a little bit flexible as well. Like an Israeli car can drive differently than a French car than an American car So what do you think of that approach and and you do you agree with that particular? assessment of AI and brute force computation in cars so Multi parts here. I’ve watched I’ve watched all of his recent talks Especially especially after auto pilot comes out. I really I’m going to respect His perspective so a lot more give a very good cvpr talk where he went into Actually, like what it takes to make these computer vision things reliable enough to drive cars. I thought the formalization of that Stuff he goes into one where he talks about like formally specifying what it means to be a safe driver. I Think that stuff was pretty weak. The problem is What he’s really trying to do there the subtext is we want to make sure we are never liable for any accidents we caused I mean you can view the world that way and sell like lawyers and accountants view the world I don’t I want to figure out an engineer. I want to drive accidents down to zero. Let’s root cause it let’s figure it out NTSB not nishta So I also don’t think it really works from a technical perspective because take a double yellow line The rules of driving say you cannot cross the double yellow line But if you do not cross double yellow lines and you’re driving around San Francisco, you will never get anywhere There’s an uber with their blinkers on right there. Okay, so then when is it? Okay, do we formalize this or do we look at big data? My argument is that the definition of driving is what people do when they drive Now you say not everybody’s a safe driver is the anna karenina quote Every good driver is good in the same way Every bad driver is bad in a different way because of that the statistical algorithm earning algorithms Kind of ignore the bad data because they can’t find much signal in it Instead of thinking of the average driver think of a committee driver Think of 100 people sitting in a room voting 100 times a second on what the car should do Some of them may be drunk Some of them may be distracted and some of them may be asleep, but the committee will probably still make the correct decision So I think that’s more the approach and when you try to formalize this stuff even trying to formalize what a car is becomes extremely hard So, I don’t know. I don’t think it’s gonna Gonna go anywhere. I think the way to actually do this is with big data and statistics Jenna has a question Yeah, I work for a big insurance company So I’m curious in your safety stats when you talk about it if you’re gonna cover it great I can’t find any data as to when open pilot actually came about so I’m curious how long it’s existed and then to When it’s been in an accident How is fault determined and how is that? Dealt with so I have some stats on a later slide, but it’s a dash cam as well So we’ve had no At fault accidents so far but the system engaged Which is better than the average you’ll see how many miles we have but How is fault determined I mean the exact same way it’s determined the human is still legally in control car open pilot is a level 2 system So just like Tesla autopilot or just like imagine you put your car on cruise control, right? You can’t say well, I understand. I rear-ended you but the car was on cruise control and it wasn’t supposed to Great, you’re liable. So liability is determined in the exact same way that a An accident between two humans would be determined right In fact people think that the problem with these kind of systems is you can’t introspect and ask them You can’t say why did you make that decision? You can’t ask humans either if you go after a car accident and ask humans What happened there recall is extremely bad if you have dashcam footage from both cars Now you start to figure out what happened um So I think devices like this which also function as cloud dash cams will be an extreme boon to the insurance industry Here you go right away instant fault. I like insurance. I like the whole businesses Well, the whole business is great its statistics. It’s we’re going to drive down accident So we’re gonna charge and we’re gonna make art we’re gonna make our profits Great So yeah, how is fault determined in the exact same awareness with humans So yeah open pilot is not it’s not some oh Hi, so I am a human machine teaming engineer so You talked earlier about automation as a gradual rise rather than as a cliff There’s a lot of work on levels of automation like you were talking about do you expect to open pilot to eventually? You know climb that gentle slope to full automation. And if so, What do you expect to need to change or grow and capabilities to achieve that if not, why not? Yes III do The so we’re gonna need a few more okay right now it only has one forward-facing camera, that’s clearly not enough You’ll need 360 degrees of vision. I think you’ll need a bit more compute than what you currently have If you really want to say the human doesn’t have to pay attention You’ll need to talk about like systems reliability engineering. You’ll need to put two Processors and making sure that failure cases are not just right now A lot of failure cases we have or safety model is based around having a human driver in the driver seat so as long as we don’t kind of have False positives it’s okay, but you start to get into true negative is becoming really bad as well if you’re trying to remove the human from the system, so Do we eventually get there? Yeah. When are we ready for it when the statistics show we are when we start to see in the statistics, oh Wow, open pilot has driven this stretch of road 1 million times and has never made a mistake. Is that good enough? Figure out those statistics and let’s see maybe ask the question. Uh Yeah, but kind of an add-on question. So you were talking earlier as well about Restrictions in the utility of full automation in terms of the you know, the larger system that the car is driving in what are your thoughts about that going forward that just the Ecosystem of consumer desires will change and legislation will fall over what? design of traffic systems It’ll be gradual it’ll be what makes sense Like I don’t know I don’t think it’s gonna change that much because I don’t think machines and humans are that different I think that the things that Have been developed over hundreds of years for human drivers will actually work out to be Oh, well, those things are actually pretty useful for machine drivers, too Like you’re like well machines won’t need stop signs. Oh, I forgot about that whole scam v2v communication Machines won’t need stop signs. The two cars will just communicate with each other. This is horrific This is I don’t know if anybody knows anything about safety who comes up with ideas like this because now you need two Active components to work correctly for a stop sign to work The beauty of stop signs is a decentralized as the logic Either party most accident there would be so many more accidents on the road today If every driver wasn’t actively avoiding accidents. Think about every time you hear a horn honk that’s an accident That would have happened if I didn’t actively swerve you need both parties to be paying attention so yeah with respect to like changing these things radically I Don’t know. I think that what’s good for you most to be good for machines as well some things are already becoming less relevant like road signs since everybody doesn’t have Like like signs saying oh this exit here. Okay, so yeah open pilot is not Some fancy research stuff. This was actually my drive to the conference. I guess. I’m on my way to the conference Oh piles driving right now It’s got the wheel. It’s got the gas at the brake It’s got the bridge traffic I have the car in the parking lot. Maybe there’s driver I Know that’s in the back. So I noticed there’s no rearview mirror So we don’t encourage people to do that You do not you you you do not have to remove your rear view mirror for driving but this is still a road-legal car on swivel law in California is you need to have two mirrors on your car that can see The back if you have to side members, you don’t actually need a regular that’s more of just an aesthetic taste on my part But we do not encourage people to do that and the kit does not require it because that would violate FMVSS But you can do it personally Doesn’t it doesn’t violate anything? So yeah business was simple profit off hardware. I Mean, you know, it’s a business model. I can respect we make the hardware for less than we sell it for We don’t license IP Ya on the website you buy the three components separately I think it was 600 mm 600 Don’t you need all three you do? Yeah. It’s about it’s about $650 for all of them. Yeah. Should we make this easier? Yeah, we should we should make it easier. I don’t know I’m the head of research now. I used to be CEO. I’m not anymore We’ll get a real CEO who who knows things about execution and business right now. We got a new CEO. He’s great, but Uh Yeah, so it could just about stats. We’ve driven a point five million miles So yeah to not have seen to not have seen the system calls accident an eight point five million miles It’s pretty good. We’ve seen a bunch of accidents with the system engaged There is one that’s particularly hilarious There’s a the system comes to stop behind a car, but the car overshot the stop line And the driver just puts it reverse and just Rams right into it Good thing you had the dash cam footage. All right So first some comparison whammo just reached ten million miles and cruise probably has about 1 million to have significantly more than them I’m the one that I did not put there is Tesla which has billions of miles on your left in the back again real quick Over what time span is this? This is oversold. Well, it’s been around now for almost three years Times this is the life timing system This question will be shorter When you have a deep learning based approach, yeah and you know, it’s really a game of getting a bunch of training datum so And what competitive disadvantage does that put a smaller company like yourself? I mean if you back to your Tesla example, right? Yeah, these are V. Say Wemo or Tesla which are just going to be piling on the training models We train our models on only about 100,000 miles of data Um, the models are not even trained on all the data and we have almost the same number of miles way now so so to pull out my amnon Well, we don’t have any small blows, right But they’re not we send all the video back. He would say that, you know if you’re trying to Specifically avoid crash situations that you know that you it’s a very low probability event Yeah, so therefore you would need billions of miles okay in order to properly train This system against these sorts of events And also every time you update the system you need another billion you know billions of miles in order to retrain the model for to avoid these events, so Is deep learning really the Scalable approach particularly for a small company. Yes Um, so the former part of that about needing billions of miles to get to level 4 I completely agree if you 1 billion miles is 10 fatalities, maybe five fatalities You have to have that kind of stuff in your dataset. If you’re trying to say to your systems level 4 Do you need to gather new billion miles? No, you need some good statistical way to Show that your system can function on replace Um, so we can just rerun the new system over the billions of miles and say it’s still good. Um does deep learning scale? Yes You have to start to do you’re not going to train on all billion miles and you don’t have to When you want to deal with crash scenarios, we’re gonna have like a data set where it’s like here’s 1 million near crash scenario These are the good ones. This is what you should do. We can do automatic recovery on that stuff, too So I got two minutes let me just get through the last few slides here this is daily miles driven And this is daily percent of miles engaged More than 50 percent of the miles are being driven by the system and this is crushing Tesla Tesla’s about 30 Their reason are is is significantly higher is because we don’t disengage on the steering wheel and everything is just more Gradual and gentle we view it as much more of a system than an autopilot But our goal is to get that number up to 100 So it’s kind of growing over time. We got some new users who were less power users It’s also engaged last around the winter because there’s snow on the roads So we expect to see a hump in the middle of every year Yeah, so how to scale and win We already have a product people pay for and use daily continue to improve driving quality and car support Lower the barrier to entry you don’t have to buy three reared animal named things and then partner dealerships for mass distribution Don’t try to talk to don’t try to talk to the OEMs. I think it’s hopeless If you partner with dealerships, it’s like partnering with you you got your phone and the AT&T store in the t-mobile store I’ve talked with some dealerships They’re super hungry for this kind of stuff and they’re the kind of people who you can talk to you’re like look we make this thing we sell it to you for $600 you sell it to people for $2,000 you give $1400 That’s a good deal Since you just did when Twitter, I call me and Instagram George Hotz Let’s talk When I’m back in your left, please. All right. There was a question over here first, and then we’ll come back Yes, so How do you deal with and you answered it a little bit or at least hinted at it because of the the winter comment? All the sensors on these cars are in different places. They have different limitations or blind spots How do you deal with the fact that your hundred thousand miles of training data? Probably doesn’t accommodate Degraded conditions and so how do you build models and understand how they vary and these in these degraded capacities because Your system. I understand that the sort of the humans always engage, but at the same time The AI is going to be generating some faulty data and how do you detect that? Yeah, it’s not generating faulty data And we do have a huge diversity of weather conditions and locations in our data set the data set just comes from all the users So we have whatever the users are actually driving on So that’s not really that big of a problem When it comes to the sensors being different This is why we have the eland the EON has a camera on the back of it that cameras the same across all the cars The radars are similar enough and the actuators are similar enough that we built a hardware abstraction layer on that abstracts them to this uniform API that gives it to all cars and going a lot more into technical details of that but uh Yeah. So again, it’s level 2 if the car like traction control fires. It’s like the ESC fires we disengage We’re like you got to be driving right now. Just like you should not use cruise control on the snow You should not use these systems if there’s snow on the road if there’s snow on the side of the road. Yes sure, it’s fine Robert Did you still have a question? My question was around you mentioned the cloud storage of the dashcam video Is that a subscription model or how does that work right now? It’s free, but we’re moving towards Yeah, we’re looking into that kind of stuff It’s expensive storage is expensive Ok Zee, and then we’ll pop back up this You’re tracking all the location data What is your how do you manage the privacy around that you by using our system you consent to giving us full access to all? the data If you don’t if you’re not ok with that, don’t use it. I’m very open about that. I’m not like Facebook I’m not wishy-washy you are giving us all the data and we will use it for whatever we want if you’re not ok with that Don’t use the system Right side so I have two quick questions one is have you found any issues with contextualizing how people drive in different regions I know in Buenos Aires for example as soon as you see a yellow on the other direction You start going in see see where I’m from. You kind of let rides Lauren out I always imagine taking half DC drivers and half Buenos Aires drivers and watching what would happen So when you have these voting’s you have differing dimensions. The other one is is Are you looking at the attention data as people are going more confident in the system? Are they more willing or more likely to pull out their phone or start losing their attention because they don’t have to again here We’re starting to get in a situation where people are feeling pretty safe on this up on the highway And they’re not paying attention anymore This is why after the second Tesla accident where the guy hit the guardrail we’re like we have to ship driver monitoring immediately So right now if you’re driving and you look at your phone if you look at your phone for more than four seconds it will beep and Then it will continue to steer but it will stop gasps. It will stop doing the gas So if you are distracted, it won’t accelerate anymore And then people learn eventually if this becomes a problem, we’re also gonna put in lock outs But that hasn’t really become a problem yet. Tesla should do driver monitoring as well So yeah, as people have become more common That’s the danger as these things get better people pay attention less and that’s you just solved with driver monitoring make sure the driver actually Pays attention and punish them for not And then the contextualization of regions, oh regionalization. Um, so we just have a country filter right now Most of the data comes from America it works. The best in America driving is similar enough in America You will get into weird edge cases, maybe like the Pittsburgh left I think the way to solve that problem long term is to train different models for different regions It’s like the iPhone has different stuff. He’s loading the region model Okay, maybe time for one more You have a Security aspect of it. Sorry. Could you raise your hand wherever you are? Oh right in the middle Yes security aspect against hacking for Oakland pilot. Um, so Car hacking is one of these things that also people make a ton of money talking about in conferences Can you give an example of what it actually happened? But at least you have seen situations where people have You know had a gang, you know a regular laptop or something talking to the canvas Oh, yeah, I’ve seen I’ve seen Charlie Miller do a PR stunt But I’ve never seen car hacking does not appear to be an actual threat by malicious actors in the real world When it does I’ll consider it much more seriously, but like until it does like what are people really going to do. They’re not It’s not really a thing but use industry standard security. Don’t do stupid things Thank you speaking from experience though. If you want to sell your system in certain countries like China, yeah Oh, we’ve given up on that mark. Yeah, they have a lot of restrictions and They actually want to be able to monitor the AI because they’re very Paranoid about about cyber security our stuff gets into China through github. I’m not I’m not touching that market I’m not working with no people like that Okay, well I personally think this was a great way to stop the conference cuz this is one of those This is a talk where I’m like one minute I’m like, yeah I’m really on board with that and then the next no I hate what you just said and then the next minute I’m like, yeah, I really and then I’m back to But talk about a provoking way to stop the conference George. Thanks very much

100 thoughts on “comma ai | George Hotz | An OS for Autos | Self Driving Cars? Scam!

  1. We express gratitude to Bryan Fox (Director of Marketing) at TTI/Vanguard and George Hotz (President of comma.ai) for allowing us to upload this video.
    We appreciate it. Follow TTI/Vanguard on social media for more cool events!
    -> https://www.ttivanguard.com
    -> https://twitter.com/ttivanguard
    -> https://www.youtube.com/channel/UCr0zqeqNY6O_pRepkY4Dcpg
    All material displayed in this video belongs to their respectable owners. We uploaded this video in good faith to share the work and progress of George Hotz and comma.ai.

  2. Excellent talk! I like when you don't give a shit about political correctness and people weren't expecting straight answers. 👍

  3. Great idea, not an easy market when people want better best, and they are lazy.
    Not to be rude, but I’ve got a lucrative idea. If you’d like to have a look. I wouldn’t mind sharing. You were honest enough to admit that Tesla’s system is currently better even when it hurt your case. You even gave their team credit, which shows a lot of respect for people you may not even like…

    So, if you see this, and are interested, let me know.

    ✌️

  4. That's a great talk. You can be sure that someone really believes in his idea when he allows anyone in the audience ask any question.

  5. Is near crash data scenario data enough to build the system or do we need data on real human fatality crashes to have a system that protects against it?

  6. Was that lower car ownership graph of the US market? The left and right units weren't labeled and the tri-color labels were using non-english characters which made me wonder what it was actually graphing.

  7. roads are paid for by taxes which means you would never have a freeway restricted to only self driving cars

  8. L4 is a worse product, right now, right? Is that still the goal? I'd rather be able to do that… I treat Uber like an autonomous driving service.

  9. I'm from the russian province. We don't have self driving cars here, really we don't have any cars here expect some old ladas. Instead of this we have horse carts. Self driving horse carts. Police once stopped me when i was sleeping drunk when driving back home from local gopnik party. Policeman said he can't let me go because he thought I did not even know where I was going. I said that my horse knows where is going and he is taking me back home. The policeman didn't have any other questions and he let me go. You see americans you think that you are better with all your fancy technology, but we are ahead of you thousands of years living simpler live. All hail Mother Russia! Amerika gay and go fuck yourself!

  10. I love how the business development bitch got her ass handed to her.. Ms.BD/Pysc: "now you should be really worried" George: "hold my beer"

  11. dude gives a great talk.
    could have answered the cyber security question at the end a bit more seriously, even though it might not be a thing.
    there are people who would say that if you make hackers' job easier i'm not buying, and that was never really addressed.

    also, i would ask if i can get a ticket for not holding the steering wheel and if they protect the customer.

  12. I was in the IRC chat you were dabbling in when you cracked the iphone, i even recall you finding the expliot somewhere everyone had overlooked/already checked, defying everyone, it was great !

  13. Especially in the first half, I felt like most people asking questions just wanted to hear themselves talk.

  14. Fantastic talk, wish I had a compatible car now. I really think I would actually buy a system. The least I can do is start talking about your product so everyone becomes aware of it. Great talk and I thought you handled the questions brilliantly.

  15. Nobody gives a shit that you have a "PHD in psychology" explain your ideas without having to try to get respect upfront by quoting your silly credentials.

  16. 14:37 That guy with the quote on the screen is wrong. There were many electric cars out before Musk the loser came out with his electric vehicle that murders people.

  17. 16:02 Yeaaah, right, loserbook is the stupidest s*it ever. They're constantly losing money, the only thing keeping it afloat is the loser wall street with it's money hungry losers that live off of other people's work plus all those acquisitions were already successful before that monster loserbook acquired them.

  18. I hope he got some more vc funding from that one. I'm surprised he didn't talk about comma prime, but maybe the talk was before that was finalized.

  19. LOL, kids getting driver license dropped over the last decade. If one does not have a driver license, what good is it to own a car? The trend is against car ownership. People like the convenience of on demand ride service. I know you are arguing for your company's product but it is going to be a ride service future.

  20. George understands vastly deeper levels of intricacy surrounding the subjects he talks on, yet he's still capable of talking in everyday terms that are at least somewhat comprehensible by the majority of people.. this is a rare and invaluable trait.

    He is clearly guided by a strong moral compass and is brave enough to publicly stick 2 fingers up at the big money grabbing scams of today. He will make mistakes but that's just part of the process, get behind him and help push him forward, the momentum you give him will speed up his mistake making, and thus speed up OUR acquisition of even better times for humanity moving forward.

  21. Would just like to thank you. George’s videos and your diligent archiving along with an immense amount of dedication have helped my boyfriend net a 250k job working on conv nets for an autonomous vehicles company. That’s not even including RSUs! Thanks so much for all you do

  22. people say Lidar is not the way people perceive while driving the car….. aren't we making self-driving cars that can drive better than human?

  23. Correction. Uber killed no one. The hopo druged lady jumped in front of the car as seen in the clip released. Killed assume guilt or intention. This guy is selling a product, and I suspect he is a fraud more the scams he is calling

  24. he says we need more cameras a better chip with a failsafe option, more data. This is everything that Tesla has or working on. This is GG for Tesla boyz

  25. Your oratory skills have greatly improved, I love your work, vision, and this presentation was excellent.

  26. It's natural for us to root for a small underdog company that says "F-U" to the establishment. And he is interesting to listen too. I wish him success. Unfortunately, in my opinion his talk is filled with so many misleading statements that I have to question if he is guilty of the same scam he accuses the industry of or is just plain naive.

  27. That Son of a bitch from Intel. Fuck him and Intel, he thinks he is soo superior and better than everyone else because he works in a huge shitstained company that has lost track of time.
    Fuck Intel this has only made me hate them even more.

  28. You think he intentionally confused real miles driven with simulated miles? George said 8.58M miles in 3 years with comma.ai (7800 miles a day)… He means simulated miles, for which Waymo has described that they have driven 10B (with a B)

  29. Have the cars link together in a close distributed networks and linked used for autonomous swarming like pods/trains for commuting the main car would be priority with optimal teck like a rig with long-distance Higher sensors/network links. That would transition big rig drivers in to different areas like near field charging too. 🌊〰️

  30. There will be a different story when automakers starts building cars without steering wheel. Then I will believe there is autonomy in the cars.

  31. Hes being cynic without giving a specific reason why. So what if Waymo still have a driver? ITS THE LAW for now.
    Title is misleading too, he acknowledges autonomous cars are coming just disagreeing on what's available now.
    Shortsighted.

  32. I didn't like that they allowed Q&A throughout the lecture. They would randomly cut off the speaker, and they wouldn't even be on the same topic that he was discussing at that time. I believe this to be rude. Structure is important in a presentation. Save the questions for the end.

  33. That's a good point of comparing comma.ai as the Android. But I remember when there were some compatibility issues between phone manufacturers vs App makers. This could also happen on comma.ai, given that they support different models at the moment, there will always be a difference for every supported car models. I can say, surely it will be minimal, however this will accumulate over time.

  34. Half that audience doesn't work with engineers and dont know what to do when someone tells them the truth. Great presentation Geo.

  35. "i think im your worst nightmare because im both a business development person and a phd in psychology"
    god just shut the fuck up

  36. How amazing is it that I can be spared the hassle of driving and, in exchange, be given the freedom to focus 100% of my attention on the performance of my open pilot system! I'm surely not gaining any time to be productive, but at least I've found yet another area of life in which I may become a passive observer rather than an active participant!

    Long live the 21st century! Families? No! Porn and tinder! History? No! Sociopath politicians and talking heads on the passive idiot box! Philosophy and tradition? No! Prole herd ideology! Driving an automobile? No! Watching my machine drive for me! The flourishing of my nation? No! Multikulti 3rd world ruins! Hooray!

  37. If you think self driving cars are a scam you need to pull your head out of your ass as soon as freaking possible and wake the hell up. Every form of transport is going autonomous even if your flat earth believing brain doesn't believe it.

  38. "you're never going to get any sort of modern performance in c" clearly this guy doesn't understand programming

  39. I love how most people said Tesla is the self driving car in their mind and he just pretended like they all said Waymo XD

  40. George's speech was excellent. Thank you for doing this George. Question should be left till the end. Mosy questions seemed to be of self-interested bureaucrats and bean counters who have no interest in forwarding the progress of self-driving cars.

  41. I want to tell you guys how I knew autonomous cars wasn't going to be a thing. Me – im a VERY good driver. Im really not just saying that. My driving techniques are from formula one race drivers. Ive saved some people a big insurance cost because instead of taken the hit, I dodged their car and put my car in danger instead. But the real reason why I knew was because I read a lot, and I roam my library a lot. In my library – in large print – is a picture of the human creative brain cell. The info card on that picture said that human creativity can never be imitated. And when you drive, you use creativity to solve driving obstacles.

  42. It sounds to me that this guy is a bit of a narcissist who wants to make himself important by insulting the overall auto industry. I don’t think he realizes once fully autonomous vehicles get the cultural thumbs up lots of people will stop buying a car. My wife and I are already counting on not needing a car in our retirement in 10 years. If I can order a self driving car using an app to go to the store… wait there’s Amazon, okay to go out and eat, when I’m not having food delivered, or to go to the doctor when my issue is beyond an AI online diagnosis. If I don’t need to go to work daily the reasons to own a car are few, if there are autonomous vehicles.

  43. Man these reporters are THE WORST. So stupid, it's a glorified adaptive cruise control..(which is awesome).. but, That's. It.

  44. I think it would be good to partner with the oem because that thing looks terribly ugly. If you work with the manufacturer they can incorporate the product into the car seamlessly.

  45. George Hotz : Self driving car is a scam.
    Also George Hotz : Hey, check up my low quality "self driving" gadgets !

  46. How does autopilot work in countries where you often run into roads with no lane lines at all ? 🙂
    if self driving works based only on lane lines and cars in front of you etc it will never get advanced…

  47. I just put this eon system in 4 days
    ago. Taking a trip from Redding ca to uvalde Texas. 1850 miles.
    So far have gone 880 miles. 99 percent of time your Eon was doing all the driving works great.
    Little things are hard to figure out. Right now the Eon is 74 percent full. I emailed comna.ai support and they said it will upload when hook up to a cell phone or wifi.
    I can not get that to work..

Leave a Reply

Your email address will not be published. Required fields are marked *