4 ms·

I'm the opposite. I don't think I'll buy another car (current is 3 years old) and I'm even surer that my son (aged 2) is not going to learn how to drive. Driv

• hntrader 6y ago
  Starcraft was solved in the sense that the winrate against top players was high, but the production model often found itself in pathological absorbing states that it couldn't escape from, which no human would ever get itself into. That's important, since if we are using it as an analogy for driving, it's not just the win rate we're interested in, it's the ability to consistently avoid pathological states when presented with edge cases (e.g. not braking when there's a strange looking object in front). Some would call this "common sense". Also, games like SC2 are distinct from the driving problem as far as RL goes, since you have unlimited training data which allows you to explore all edge cases given enough compute. In self-driving, the data is scarce and you can't explore all edge cases by just throwing compute at the problem.
  • treis 6y ago
    >avoid pathological states These just need to be relatively rare and not dangerous when they happen. If my car slams on its brakes every 100,000 miles for no good reason that's not really a big deal. At least it's not big enough to stop them from being used. >unlimited training data Unlimited is not possible, but we can get tens of millions of miles pretty easily. Then many many more as they roll out into production.
• macintux 6y ago
  Edge cases can crop up while you're en route. My go-to example: you're driving at night, two-lane road, and there's a minor accident in your lane ahead. You crawl forward to the point where you can pass the accident, and then what? Someone involved in the accident is directing traffic. In the dark, with no safety gear. On the other side the traffic is backed up, so you see headlights facing you in the lane you have to use to pass the accident. To me, that seems like an insanely difficult problem for a computer to deal with. A human driver, that's easy, but a computer? (Having said that, I'm historically terrible at predicting just how fast technology is going to advance, so it's safest to assume I'm wrong and this will turn out to be easy.)
  • dane-pgp 6y ago
    For edge cases this rare (and brief), there can be a small team of remote operators that take over, using the streamed camera data to decide when to tell the car to venture into the other lane. The remote driver could also call the passenger's phone to double check that the sensor data is correct, and to reassure them that the AI isn't just guessing.
• TulliusCicero 6y ago
  > Chess, Go, Starcraft, DOTA all were called unsolvable at one point and all were eventually solved. Depends on what you consider 'solved'. AlphaStar was never able to demonstrate the level of superhuman performance in Starcraft that AlphaGo did. It got to grandmaster level, but even that was with significant mechanical advantages over humans, when what they wanted to test was the tactical/strategic thinking, not raw ability to issue basic commands (which an AI will easily win at, obviously). That said, AlphaStar was vastly superior in performance and well-roundedness compared to any previous Starcraft AI, so it was still very impressive. It just wasn't "beating the best in the world at strategic thinking" impressive.
  • treis 6y ago
    Solved meaning it was way better than the average Starcraft player. If we can get driving to even average levels that will be a huge win because it won't get distracted or drunk.
    • TulliusCicero 6y ago
      Yeah fair enough. Though I'd like to see AlphaStar perform with comparable mechanical limitations to humans as well. That's a complete non-issue for Go or Chess, but for Starcraft it matters a lot.
      • treis 6y ago
        Self driving cars will have a mechanical advantage too. They can go to maximum braking essentially immediately. It takes time for humans to move their legs and then some time for the pedal to move. Probably a 200-500ms advantage for the AI.
• lm28469 6y ago
  > Your edge cases are just that. Edge cases. Nobody is going to care if they can't drive in extremely heavy snow. The problem when you have 1.4B vehicles in circulation is that your edge cases happen dozens of time per second. > Nobody is going to care if they can't drive in extremely heavy snow. A loooot of people will care. Snow isn't an edge case in a very large part of the world, it comes every year and stays for weeks or months. What we're going to have is a system that's slightly better than existing "smart" systems [0] but much worse than "fully autonomous". People who will learn on these systems will suck at driving and be in complete panic mode at every single of the "edge cases" they'll face very often People have to understand that the PCH from SF to LA and quiet Mountain View streets are the edge case to the rest of the world. [0] ESP/ABS, collision avoidance, blind spot warnings, auto emergency braking