3 ms·

Level 4 doesn't actually include the long-tail though, that's Level 5. It's really a terrible benchmark system. Tesla could be level 4 tomorrow if they said (li

Level 4 doesn't actually include the long-tail though, that's Level 5. It's really a terrible benchmark system. Tesla could be level 4 tomorrow if they said (like Waymo) "we're gonna enable FSD only in the greater Phoenix area", where the roads are well-maintained, they could map them all out, etc. The issue here is that Tesla actually wants a generalized system that will work anywhere a human driver can drive, and does not seem interested in geofencing at all until they've reached that (level 5) point.

debrice 6y ago
I find this strategy very close to the lean startup model (MVP). Go to market with something you believe is better than what’s out there early, have rapid feedback and iterate. Feels like a risky bet for Waymo and others to reach level 4 or 5 only in a limited environment, you might end up with so many untested assumptions
- fastball 6y ago
  Yes, to me this is also another reason Tesla decided to not go with LIDAR (besides cost): if your self-driving car uses a different sensing system than a human, there are no guarantees that this car will be able to drive anywhere a human can drive. Meanwhile, if your system is relying on entirely the same thing humans rely on, the likelihood seems higher you'll end up with a more generalized solution, since all roads / etc on earth are built around vision, not LIDAR.
kmonsen 6y ago
Do you have any evidence Tesla could do this in Phoenix? I think that is highly unlikely and at this stage might not be possible without lidar.
- fastball 6y ago
  With a geofenced area you can hard-code behaviors in any area where FSD struggles. I see no reason to believe they couldn't do this – FSD can already drive quite a few places with zero assistance, and most of the hiccups seem to be around edge-cases, which are easily pre-programmed.
  - wpietri 6y ago
    Do you have any evidence that's what Waymo is doing, hardcoding behaviors in specific spots/
    - fastball 6y ago
      When I say "hardcoding" I don't actually mean hardcoding. I mean forcing the ML system to keep iterating on specific problems that it did poorly at, with human supervision. e.g. the car did the wrong thing at the intersection of 9th and Elm? Give it 100 similar situations (at 9th and Elm), let it generate lots of possible behaviors, and then have humans tell it which are correct. Then whenever a car is at 9th and Elm, it uses that training as a priority. For a relatively small geofenced area that is a reasonable approach and should work well. Unfortunately it doesn't scale to the whole country/world.
      - wpietri 6y ago
        Why doesn't it scale? That's basically how humans do it. We learn general-case rules and then we learn how to handle specific situations specifically. You can observe that just by driving through an unfamiliar city with a local. They'll feed you a stream of tricks for optimizing the drive.

debrice 6y ago

I find this strategy very close to the lean startup model (MVP). Go to market with something you believe is better than what’s out there early, have rapid feedback and iterate. Feels like a risky bet for Waymo and others to reach level 4 or 5 only in a limited environment, you might end up with so many untested assumptions

fastball 6y ago
Yes, to me this is also another reason Tesla decided to not go with LIDAR (besides cost): if your self-driving car uses a different sensing system than a human, there are no guarantees that this car will be able to drive anywhere a human can drive. Meanwhile, if your system is relying on entirely the same thing humans rely on, the likelihood seems higher you'll end up with a more generalized solution, since all roads / etc on earth are built around vision, not LIDAR.

kmonsen 6y ago

Do you have any evidence Tesla could do this in Phoenix? I think that is highly unlikely and at this stage might not be possible without lidar.

fastball 6y ago
With a geofenced area you can hard-code behaviors in any area where FSD struggles. I see no reason to believe they couldn't do this – FSD can already drive quite a few places with zero assistance, and most of the hiccups seem to be around edge-cases, which are easily pre-programmed.
- wpietri 6y ago
  Do you have any evidence that's what Waymo is doing, hardcoding behaviors in specific spots/
  - fastball 6y ago
    When I say "hardcoding" I don't actually mean hardcoding. I mean forcing the ML system to keep iterating on specific problems that it did poorly at, with human supervision. e.g. the car did the wrong thing at the intersection of 9th and Elm? Give it 100 similar situations (at 9th and Elm), let it generate lots of possible behaviors, and then have humans tell it which are correct. Then whenever a car is at 9th and Elm, it uses that training as a priority. For a relatively small geofenced area that is a reasonable approach and should work well. Unfortunately it doesn't scale to the whole country/world.
    - wpietri 6y ago
      Why doesn't it scale? That's basically how humans do it. We learn general-case rules and then we learn how to handle specific situations specifically. You can observe that just by driving through an unfamiliar city with a local. They'll feed you a stream of tricks for optimizing the drive.