4 ms·

This is one of those "you need to be an expert to evaluate the claim" problems. Talk to anyone with research experience who are building world scale ML systems

This is one of those "you need to be an expert to evaluate the claim" problems. Talk to anyone with research experience who are building world scale ML systems that are used by real people, and it's unambiguous that Elon Musk is a con man. It comes down to small details, like the nature of the training data, the nature of the customizations, the nature of the business problems, the nature of the regulatory problems, the difference between highway driving and city driving. The enormous amount of special casing, the nature of the QA infrastructure, the capital on hand, the routing and mapping problems... What Tesla has is a decent lane finding system. What Google has is an end to end system. It won't take them long to expand to all the major cities; solving the geofencing problem is easy, they already made street view. And they were doing L4 successfully years ago. They're keeping drivers in the cars right now mostly because something going wrong is a PR risk, not because they need them. And L5 isn't a particularly important goal. My guess is I'd have to write up 20 pages of detailed analysis at minimum to convince folks. In general if you don't have 5 to 10 years in depth experience in this space, you probably shouldn't consider qualified to have an opinion.

valine 7y ago
I'm not saying that Musk has solved fsd or that he will in the near future, but if you asked anyone with experience building cell phones in 2006 if the iPhone was possible they would have said absolutely not. Technological progress is inherently unpredictable even for experts. Musk has some really good people working for him and their datasets get bigger every day, I wouldn't rule Tesla out just yet.
- serioussecurity 7y ago
  Their datasets aren't useful for the problems that make driving hard. Musk is making claims which are paper thin false. If he weren't making these kinds of claims I'd take him more seriously. I promise you that the reasons this is ludicrous are extremely solid. Edit: fwiw I've actually taken a look at a number of waymos internal systems, and sat down with people who worked on autopilot.
  - valine 7y ago
    If Tesla doesn’t have the right training data then no one does. Tesla has both simulations and vast amounts of real world data from cars all over the world. Extracting depth maps from 2d images is a solved problem, which negates the need for Lidar. Best I can tell Tesla is the closest out of everyone to solving self driving. I don’t think they’ll solve it by next year, but If I had to guess I’d say 2.5 to 3 years from now Tesla will have relatively safe self driving cars.
freerobby 7y ago
Appeal-to-authority arguments are always uncompelling, but they are especially uncompelling against Musk. Authorities said hydrogen fuel cells were the future. Authorities said EVs will never outperform gas cars. Authorities said batteries will always cost too much. Authorities said reusing rockets will never work. With each of these breakthroughs, Musk presented a well-reasoned view that was highly contrarian against the authority consensus; and in each case, his view turned out to be not only correct, but so correct that it became the new authority consensus. Maybe this time he is wrong and you are right. But if you are wondering why the market and the masses believe him, it's because he continues to provide reasonable explanations for what he's proposing, and he has a track record to back them up. If history predicts the future, then in a few years, we'll add "authorities said you needed LIDAR for self-driving" to the list above. This is already starting to play out in the SDC space, with Hotz of Comma.ai sharing Musk's views on vision, and Levandowski sharing his more generalized concern about LIDAR (the latter is more recent if you're not familiar): > Levandowski contends that this “race” to deploy autonomous vehicles has yet to start in earnest largely due to shortcomings from “crutch technologies,” a descriptor he uses for hardware like LiDAR and HD maps. His position is that while these provide sensing and localization for the vehicle in the present moment, they have serious compromises and don’t produce the level of predictive ability required to commercially deploy autonomous vehicles. https://techcrunch.com/2018/12/18/anthony-levandowski-prontoai-self-driving-trucks/ https://techcrunch.com/2018/12/18/anthony-levandowski-pronto... Karpathy was able to explain in very simple and understandable terms this week why Tesla's approach is advantageous. If the critics understand this domain so well, how come they can't provide a straightforward retort? If Feinman could do it with quantum physics, surely a machine learning expert can do it with neural networks. I would love to read/hear/watch such a retort if anybody is interested in presenting one.

valine 7y ago

I'm not saying that Musk has solved fsd or that he will in the near future, but if you asked anyone with experience building cell phones in 2006 if the iPhone was possible they would have said absolutely not. Technological progress is inherently unpredictable even for experts. Musk has some really good people working for him and their datasets get bigger every day, I wouldn't rule Tesla out just yet.

serioussecurity 7y ago
Their datasets aren't useful for the problems that make driving hard. Musk is making claims which are paper thin false. If he weren't making these kinds of claims I'd take him more seriously. I promise you that the reasons this is ludicrous are extremely solid. Edit: fwiw I've actually taken a look at a number of waymos internal systems, and sat down with people who worked on autopilot.
- valine 7y ago
  If Tesla doesn’t have the right training data then no one does. Tesla has both simulations and vast amounts of real world data from cars all over the world. Extracting depth maps from 2d images is a solved problem, which negates the need for Lidar. Best I can tell Tesla is the closest out of everyone to solving self driving. I don’t think they’ll solve it by next year, but If I had to guess I’d say 2.5 to 3 years from now Tesla will have relatively safe self driving cars.

freerobby 7y ago

Appeal-to-authority arguments are always uncompelling, but they are especially uncompelling against Musk. Authorities said hydrogen fuel cells were the future. Authorities said EVs will never outperform gas cars. Authorities said batteries will always cost too much. Authorities said reusing rockets will never work. With each of these breakthroughs, Musk presented a well-reasoned view that was highly contrarian against the authority consensus; and in each case, his view turned out to be not only correct, but so correct that it became the new authority consensus. Maybe this time he is wrong and you are right. But if you are wondering why the market and the masses believe him, it's because he continues to provide reasonable explanations for what he's proposing, and he has a track record to back them up. If history predicts the future, then in a few years, we'll add "authorities said you needed LIDAR for self-driving" to the list above. This is already starting to play out in the SDC space, with Hotz of Comma.ai sharing Musk's views on vision, and Levandowski sharing his more generalized concern about LIDAR (the latter is more recent if you're not familiar): > Levandowski contends that this “race” to deploy autonomous vehicles has yet to start in earnest largely due to shortcomings from “crutch technologies,” a descriptor he uses for hardware like LiDAR and HD maps. His position is that while these provide sensing and localization for the vehicle in the present moment, they have serious compromises and don’t produce the level of predictive ability required to commercially deploy autonomous vehicles. https://techcrunch.com/2018/12/18/anthony-levandowski-prontoai-self-driving-trucks/ https://techcrunch.com/2018/12/18/anthony-levandowski-pronto... Karpathy was able to explain in very simple and understandable terms this week why Tesla's approach is advantageous. If the critics understand this domain so well, how come they can't provide a straightforward retort? If Feinman could do it with quantum physics, surely a machine learning expert can do it with neural networks. I would love to read/hear/watch such a retort if anybody is interested in presenting one.