Self-driving vehicles

George Hotz: From Self-Driving Cars to Robots That Cook and Clean

The visionary hacker on how he plans to "solve A.I." and why he thinks this will be a "decade of decentralization."

|

HD Download

George Hotz, the hacker and entrepreneur who as a teenager became famous as the first person to jailbreak the iPhone and then the PlayStation 3, is determined to build a human-level artificial general intelligence. As a first step, he's racing to build the world's first fully self-driving car system.

Hotz's company, comma.ai, which has raised $8.1 million, is competing against giants like Google offshoot Waymo, Amazon-owned Zoox, and General Motors project Cruise. Those efforts rely on custom vehicles with expensive sensor arrays such as LIDAR—light detection and ranging—to identify surrounding objects and people, navigating a pre-mapped path. Hotz says his better-financed competitors have misled the public with hype about what their products can do and on what time scale.

But comma.ai takes a radically different approach, using vision alone to analyze the road ahead and drive as a human would. Its latest product, the comma three, runs on a smartphone processor and plugs into most new cars, taking over the built-in steering, gas, and brake systems. The company's open-source software, openpilot, uses artificial intelligence (A.I.) to ingest the comma three's video feeds and predict where to drive in real time.

While Tesla is also moving toward a vision-only system, Hotz says comma.ai's approach is unique: The latest version of openpilot features the self-driving industry's first "end-to-end" architecture for lane keeping, which means that it won't explicitly identify objects and people or rely on road features such as lane lines. Instead, the software analyzes a scene holistically and determines the most likely path a human would drive.

Hotz says only this approach to computer vision and self-driving can lead to the development of robots that can cook, clean, and do manual labor—and ultimately to an A.I. system that can do everything a human can.

According to Hotz, comma.ai's commitment to open-source software also represents a philosophical split with companies like Waymo that are building closed systems with centralized control. In the self-driving space and the larger tech industry, Hotz is optimistic this will be a "decade of decentralization," and predicts that the 2020s will see distributed systems powered by blockchain technology—including cryptocurrencies like bitcoin—overtake corporate giants like Visa in both trust and value to consumers.

Produced, written, and edited by Justin Monticello. Director of photography Dillon Mortensen. Camera operator Elvis Leon. Audio production by Ian Keyser.

Music: The City of Hope by Borrtex.

NEXT: Why It Took So Long To Leave Afghanistan: Antiwar.com’s Scott Horton

HD Download

Editor's Note: We invite comments and request that they be civil and on-topic. We do not moderate or assume any responsibility for comments, which are owned by the readers who post them. Comments do not represent the views of Reason.com or Reason Foundation. We reserve the right to delete any comment for any reason at any time. Report abuses.

  1. “ultimately to an A.I. system that can do everything a human can”

    Yeah sure and air cars and fusion are just on the horizon too.

    1. Since I started with my online business, I earn $25 every 15 minutes. It sounds unbelievable but SQa you won’t forgive yourself if you don’t check it out.

      Learn more about it here……. VISIT HERE

      1. After leaving my previous job 12 months ago, i’ve had some good luck to learn about this website which was a life-saver for me… They offer jobs for which people can work online from their house. My latest paycheck after working for them for 4 months was for $4500… Amazing thing about is that the only thing required is simple typing skills and access to internet…Read all about it here… READ MORE

        1. Fantastic work-from-home opportunity for everyone… Work for three to eight a day and start getting paid inSd the range of 17,000-19,000 dollars a month… Weekly payments Learn More details Good luck…

          See……………VISIT HERE

    2. Since I started with my online business, I earn $25 every 15 minutes. It sounds unbelievable aaq but you won’t forgive yourself if you don’t check it out. Learn more about it here… Visit Here

    3. I made over $700 per day using my mobile in part time. I recently got my 5th paycheck of $19632 and all i was doing is to copy and paste work online. this home work makes me able to generate more cash daily easily.FHk simple to do work and regular income from this are just superb. Here what i am doing. Try now………

      GOOD LUCK………. VISIT HERE

  2. As a motorcycle guy, it isn’t enough to avoid your own mistakes. Because the insurance company says it wasn’t your fault is of little consolation if you’re in a wheelchair for the rest of your life. They can put, “It Wasn’t His Fault” on your tombstone, but you’ll still be dead. Driving safely requires a motorcycle guy to anticipate and avoid the mistakes of other people.

    And some of that has to do with intuitive factors about the kinds of people who are driving and the kinds of cars they drive. I might pass a truck pulling a plaster rig on the left, where a young person yapping on the phone or a small car with great handling that might change lanes to zig zag through traffic, I wouldn’t pass on the left.

    I do not believe AI will ever approach the “intelligence” necessary to outperform my own driving when it comes to safely navigating the freeways of southern California for 15,000 miles on a motorcycle year after year after year after year. And even if they could, I know I wouldn’t give up that much fun for such a smaller increase in safety.

    My safety record is perfect.

    1. Driving safely requires a motorcycle guy to anticipate and avoid the mistakes of other people. And some of that has to do with intuitive factors about the kinds of people who are driving and the kinds of cars they drive.

      Everything you said also applies to car drivers. And it’ spot on.

      I’ve never been a motorcycle guy, but it annoys me how little people deviate from their routines in their presence. It just demonstrates a total lack of spatial/situational awareness.

    2. Well, they aren’t trying to build self-driving motorcycles, they are trying to build self-driving cars.

      Some of the issues you mention may be addressed by AI-to-AI communications and by increasingly excluding human drivers from roads. (Don’t shoot the messenger.)

      1. Some of the issues you mention may be addressed by AI-to-AI communications and by increasingly excluding human drivers from roads. (Don’t shoot the messenger.)

        And pedestrians… oh and inclement weather.

      2. Everybody’s got a line that other people shouldn’t cross, and if they come after my ability to ride my motorcycle, the way I want, there will be very few other issues that matter to me.

        From an AI perspective, I think they’ll have trouble initiating a lean on a motorcycle to start a turn automatically, which is how motorcycles steer. That’s why the tires are round on the bottom instead of flat. You’re turning on the side of a motorcycle tire and you initiate the turn by leaning and then counter-steer.

        Make me go electric on a motorcycle, and that wouldn’t be worse than taking away my ability to lean, steer, shift, accelerate, and break as I please with AI. And if they try to force us to go AI on that, they might have a really hard time enforcing their stupid law.

        Check this out.

        https://www.youtube.com/watch?v=urgQKNMdTQM

        The white car is an unmarked police car. As libertarians, we know that the law is just a fantasy when it doesn’t reflect reality. And sometimes reality is that the law is unenforceable. The other thing propeller head tech people have is accounting for people’s qualitative preferences.

        My qualitative preference for fun on motorcycles is such that they probably don’t want to know what I care about less than having fun on my motorcycle. And I am not alone–not even in California.

        1. What about your qualitative preference to live in a place that gets snow? Or… doesn’t have lines on the road because it’s the English countryside.

          “So you guys run that self driving test track in Las Vegas?”

          “Yep”

          “How do those cars do in the snow”?

          “Dude, there IS no snow in Vegas…”

          “So none of your prospective customers will ever take one of those cars up to Tahoe?”

          *silence*

        2. They’re more likely to just take your bike away. Not safe they’ll say. Like guns, drugs and vaping.

        3. Everybody’s got a line that other people shouldn’t cross, and if they come after my ability to ride my motorcycle, the way I want, there will be very few other issues that matter to me.

          I’m simply saying that if self driving cars become popular, the way conflicts between them and human driven vehicles may simply be resolved by banning human driven vehicles from the road. Along with it, I expect motorcycles of any form will also be banned.

    3. You are comparing self-driven cars to the wrong baseline. AI doesn’t have to be better than the best 10% of drivers, it only has to be better than the worst 10% to be a net social good.

      Remember that while some drivers are in denial about their (in)ability, many are self-aware and will opt for safer options when given a practical alternative. I, for example, would be much more willing to take away my elderly mother’s keys if she could still have independent mobility. Shoot, I’d give up my own steering wheel on those days when I’ve put in too many hours but still have a long drive home.

      1. “You are comparing self-driven cars to the wrong baseline. AI doesn’t have to be better than the best 10% of drivers, it only has to be better than the worst 10% to be a net social good.”

        Caring about the top 10% of drivers is too large a group for me. I care about one driver primarily, and his name is Ken Shultz. The point is to preserve Ken Shultz’s ability to make choices for himself–and to cut off the scumbags who would violate their ethical obligation to respect my agency with some bullshit about the greater good for the most people.

        I do not exist for your benefit and neither does my right to make choices for myself. If you need to universalize that to make it okay for your sweet sensibilities, no one’s else’s agency exists for you to exploit when you calculate the best interests of other people either.

        Moreover, your belief in your ability to make qualitative judgements for 90% of the people out there is based on nothing. You have no ability to make qualitative choices for other people whatsoever. You don’t know how much they value their liberty and their ability to make choices for themselves vs. their safety, and any claim you make to be able to do so is based on a logical fallacy–at best.

        On the one hand, you have 330 million Americans who can optimize our collective qualitative preferences when each individuals is free to make conflicting choices for themselves, balancing their individual qualitative preference for freedom against their individual qualitative preference for safety. On the other hand, you claim to be capable of making choices that reflect the qualitative choices of 90% of Americans–with no facts to support such a claim and no basis in logic.

        How do you know 90% of the people who aren’t great drivers would rather forgo their freedom in the name of safety?

        There are people in this world who value safety more than anything else on a qualitative basis. They’re diagnosed with psychiatric disorders like agoraphobia and paranoia. Maybe you could claim to make better qualitative choices than insane people do for themselves, but the rest of us, who choose to compromise our safety to different degrees every time we leave the house, make qualitative judgements all the time for ourselves, and the belief that you can make better choices for them on a qualitative basis than they can for themselves is simply ludicrous.

        1. Roads are not a part of nature. They are created and maintained by the state and voters will determine what they are used for. If voters determine they don’t want any human-driven vehicles on taxpayer-funded infrastructure anymore, that’s what’s going to happen. If self-driving cars start to work reliably in a few years, that’s likely what’s going to happen because it’s hard for self-driving vehicles and human-driven vehicles to co-exist.

          Now, while I’m perfectly sympathetic to the argument that roads should be privatized, the outcome would likely be the same or even more restrictive.

          What vehicles are permitted on taxpayer funded roads is simply not a matter of people “balancing their individual qualitative preference for freedom against their individual qualitative preference for safety”, it’s a decision about how to utilize and maintain a shared resource, not very different from the pool in an HOA.

  3. How do you convert a floor sweeper into a snow remover?

    Hand her a shovel.

    1. Exactly. And why would I need to buy a robot to cook and clean? That’s what girlfriends and wives are for.

  4. To financial kings with armies of killer robots.

    There’s nothing money won’t buy. It’s a free market

    Don’t give up your guns just yet.

  5. ultimately to an A.I. system that can do everything a human can.

    Yep, eventually humans will be that stupid.

    1. There is much in what you say.

  6. So there are only two things he’s wrong about in the headline…

  7. >>is determined to build a human-level artificial general intelligence.

    yes yes, I’ve also been watching Humans on amazonprime

    1. Are White Mike and Jeff considered human? If so then…

  8. The latest version of openpilot features the self-driving industry’s first “end-to-end” architecture for lane keeping, which means that it won’t explicitly identify objects and people or rely on road features such as lane lines. Instead, the software analyzes a scene holistically and determines the most likely path a human would drive.

    I appreciate him. I appreciate he’s thinking outside the box, but I don’t have a lot of hope that he’ll be more successful than they are. Or even get a reliable self-driving AI on the road that doesn’t kill its passenger or pedestrians.

    1. I appreciate him. I appreciate he’s thinking outside the box

      What “box” that would be? End-to-end training is one common approach to self-driving cars. The reason many people don’t like it is because it’s hard to diagnose failures and because regulators may object.

      1. The “box” that the other self-driving behemoths are currently failing at.

        1. And Hotz isn’t succeeding either. Nor has he succeeded at anything AI-related as far as I can tell.

      2. And also, Hotz seems to be the only person who thinks that approach, which is literally the first thing you learn when taking an AI class, is sufficient on its own.

        Sure, humans don’t have access to sensors and lidar and advanced communications and everything else, and we drive alright, so it is theoretically possible.

        Although he seems to ignore the fact that to date, we still have not developed computers with spatial reasoning anywhere close to that of humans (who use an entirely separate section of the brain to do it), and saying, but AI, doesn’t substitute for that (which is precisely why everyone else is using augmented technology as a substitute like lidar or other technologies)

        And even if you could brute force past that with a large enough dataset, his dataset is tiny. 20 years is nothing. Yes, humans make do with less data. But that is because humans do have reasoning capabilities to adapt smaller amounts of data, which Tesla tries to do by reconstructing real life data in simulation and changing it … which Hotz disparaged.

        Also, I dont really understand why he is limiting himself to company data, which can be biased, over simply … buying a whole lot of dashcam footage. Or calling literally any insurance company. Like if the tech is as great as he says it is any insurance company can just run their data through a simple AI model and boom.

        His answer that humans use the same amount of data to account for edge cases is just not true. Computers need way more. Again there are standard ways around that issue, but that is precisely what other companies are doing!

        My problem with this is just, he doesn’t seem to have insight that Waymo and others don’t have. Everyone else tried this, it didn’t work, so they developed workarounds and better tools. What exactly is being done here that gets around the issues everyone else ran into?

        The one thing I do admire is his willingness to break rules to get this done. But I’m not sure what else is great here.

        1. Human level AI would need 16 years to get ready to drive a car safely (and really more like 20 if we’re being honest). Because humans need that long.

          1. Humans learned to drive cars over 60 million years ago. The fact that so many computer science AI nerds haven’t internalized that is why this process has been so messy.

        2. Sure, humans don’t have access to sensors and lidar and advanced communications and everything else, and we drive alright, so it is theoretically possible.

          The point of a self-driving car is its supposed to be better than a human. So if it can see in no or zero light conditions, see through fog, snow, do high-speed detections of closing objects– is what we’re SUPPOSED to get. If we think something with the field of view of a cell phone camera can make a judgement of where the Prius driver in front of me is going because I’m watching how his head is swiveling as he slows down… I think we’re lying to ourselves.

          And even if you could brute force past that with a large enough dataset, his dataset is tiny. 20 years is nothing. Yes, humans make do with less data. But that is because humans do have reasoning capabilities to adapt smaller amounts of data, which Tesla tries to do by reconstructing real life data in simulation and changing it … which Hotz disparaged.

          You might find a lex friedman podcast with Sir Roger Penrose very interesting. I like Penrose because his view of AI more closely matches mine– and as a lay person, it makes me feel warm and fuzzy to see someone with his scientific insight come to similar conclusions (and be able to back them up with science, research and a profoundly deep understanding that I simply can’t).

        3. The one thing I do admire is his willingness to break rules to get this done. But I’m not sure what else is great here.

          I agree, and that’s my position. Guys like Hotz can find innovative things that the institutional researches missed, and can incorporate into their own models. That doesn’t mean Hotz is going to win this particular space race, or even succeed at it.

          1. Particularly because so much of this is a legal issue. When his systems fucks up and causes a crash, and it will if it ever goes at scale, who is responsible? That’s a non-trivial amount of Uber and Tesla and Waymo’s dealing.
            If they can do a safe job of interstate driving that’s a gain in of itself. If I can set my car to take me to Kansas City for some barbecue and just nap or read the whole way, that’s a boon.

            1. Driving on the interstate is way easier than driving in town. The self-driving trucks are going to do the interstate portion first, then switch to human drivers for pick-ups and deliveries in the cities.

  9. As a first step, he’s racing to build the world’s first fully self-driving car system.

    Has Hotz actually done anything of note with AI?

      1. So, no track record at all.

  10. “ultimately to an A.I. system that can do everything a human can.”

    Then they won’t need to give us a vaccine for the next plandemic.

    1. When better self driving cars are built Buick will build them and a hacker will hijack them and steer them into a tree.

    2. “plandemic” lol

      1. Freudian slip; Misek isn’t that bright.

        1. He also wants to exterminate Jews.

  11. “…takes a radically different approach, using vision alone to analyze the road ahead and drive as a human would.”

    Somehow that doesn’t make me feel any safer. Have you seen how other people drive?

  12. I live in an area of SF where the autonomous driving cars are in nearly constant ‘testing’, and have been for at least ten years.
    They are still not capable of identifying an issue and driving around a double-parked car, and if double-parked cars are an issue, you’ve got a LOOOONG ways to go.
    If he truly believes this, as a hacker, he’s going to be unemployed.

Please to post comments