4 ms·

I totally understand where you are coming from. However, I don’t think it is possible to prove this type of technology is safe in the real world without lots o

by slg 4y ago

I totally understand where you are coming from. However, I don’t think it is possible to prove this type of technology is safe in the real world without lots of data of it being used in the real world. Requiring proof first like you suggest creates a catch-22 situation in which self driving technology can never be released.

rootusrootus 4y ago
Automakers already use proving grounds. We can set a minimum standard for FSD before turning it loose with amateur drivers.
- slg 4y ago
  You can’t test self driving tech the same way you can test most safety technology. It is easy to test something like an airbag. You throw dummies in the seats and run the car into a wall. There is no equivalent for self driving tech. The tech needs to be able to handle any weird situation that humans and nature throw at it. You therefore need a huge number or interactions, billions of miles worth, to be confident in its safety. You can’t just put a car through a test track to see if it is safe. The range of situations the technology needs to be able to respond to is too large to fully test like this. It is a perfectly valid policy to want that proof ahead of time. I’m just pointing out that this requirement would effectively kill development of this technology for the foreseeable future.
  - rootusrootus 4y ago
    Is Tesla using customer cars for training data? That should be very noticeable, it is a lot of data that would be getting fed back. If they're not, then the billions of miles driven by customers isn't relevant, and the same "can't just put a car through a test track" problem exists at the training level.
    - jjoonathan 4y ago
      Yes they are, yes it is very noticeable, and yes it is a lot of very important data that could not have been collected on a test track.
      - rootusrootus 4y ago
        I'd like to see something to back that claim. The amounts of data we're talking about are quite significantly more than any I've seen people experience in real life. As far as I can tell, customer cars run inference only, collect very little data unless something prompts them to (and even then, store it locally), and contribute almost zero to the training set.
        jjoonathan 4y ago
        Here's a statement from 2022. They had a labeling team of 1500 people and they laid off about 200 after the labeling infrastructure team finished an auto-labeler, so it's not just a bloated team. Do you think they are keeping 1300 people busy correcting labels on purely internal data? https://teslamotorsclub.com/tmc/threads/tesla-lays-off-over-200-autopilot-data-labelling-employees.271163/ https://teslamotorsclub.com/tmc/threads/tesla-lays-off-over-...
  - UncleEntity 4y ago
    Google’s robocars seem to be taking a very conservative approach to integration with the motoring public. Don’t get me wrong, I irrationally hate the robots, used to see them all the time around Phoenix, was actively gunning for them and only had one time I could have rammed one[0] with it not being my fault. Which says a lot because it was (well, still is) a life goal to take out a robocar so I was really trying. I used to see them doing a bunch of really stupid stuff but never anything that was outright dangerous, relatively speaking. One did stop in the fast lane when everyone was zipping by at 70+ because it couldn’t figure out how to merge into the left hand exit ramp but plenty of snow birds do that too so it’s not completely unheard of. [0] (un)fortunately I was behind on my rent and couldn’t afford to spend who knows how long dealing with an accident investigation so the robot got to go about its day unmolested.
    - rootusrootus 4y ago
      Google's approach is more defensible. They are careful, default to conservative behavior and stop when confused, and had professional drivers for years. And even then they're a road hazard. At least they hold liability if anything happens, though, which is an important step.
  - kaba0 4y ago
    Also, since parts of it is a blackbox, doesn’t any update literally clears out any former result and should be revalidated from scratch?
ndsipa_pomu 4y ago
The sensible method is for the manufacturer to perform lots of testing and make available the data. Legislators can then allow limited real world usage to determine if there's problems and then expand that usage as appropriate. e.g. allow autonomous vehicles to take control on motorways (sorry, that's a UK term - is it freeways in the U.S.?).
- jjoonathan 4y ago
  That's exactly what happened. There was a back and forth with NHTSA for years before the beta started going public, and the rollout was carefully staged. At first, you needed a high safety score and it was aggressively geofenced. Those restrictions were gradually lifted as they collected data, and it still has attentiveness monitoring and a strike system. Tesla and NHTSA go back and forth on details all the time -- just recently NHTSA demanded some adjustments for stop signs and yellow lights.

by slg 4y ago

rootusrootus 4y ago

Automakers already use proving grounds. We can set a minimum standard for FSD before turning it loose with amateur drivers.

slg 4y ago
You can’t test self driving tech the same way you can test most safety technology. It is easy to test something like an airbag. You throw dummies in the seats and run the car into a wall. There is no equivalent for self driving tech. The tech needs to be able to handle any weird situation that humans and nature throw at it. You therefore need a huge number or interactions, billions of miles worth, to be confident in its safety. You can’t just put a car through a test track to see if it is safe. The range of situations the technology needs to be able to respond to is too large to fully test like this. It is a perfectly valid policy to want that proof ahead of time. I’m just pointing out that this requirement would effectively kill development of this technology for the foreseeable future.
- rootusrootus 4y ago
  Is Tesla using customer cars for training data? That should be very noticeable, it is a lot of data that would be getting fed back. If they're not, then the billions of miles driven by customers isn't relevant, and the same "can't just put a car through a test track" problem exists at the training level.
  - jjoonathan 4y ago
    Yes they are, yes it is very noticeable, and yes it is a lot of very important data that could not have been collected on a test track.
    - rootusrootus 4y ago
      I'd like to see something to back that claim. The amounts of data we're talking about are quite significantly more than any I've seen people experience in real life. As far as I can tell, customer cars run inference only, collect very little data unless something prompts them to (and even then, store it locally), and contribute almost zero to the training set.
      - jjoonathan 4y ago
        Here's a statement from 2022. They had a labeling team of 1500 people and they laid off about 200 after the labeling infrastructure team finished an auto-labeler, so it's not just a bloated team. Do you think they are keeping 1300 people busy correcting labels on purely internal data? https://teslamotorsclub.com/tmc/threads/tesla-lays-off-over-200-autopilot-data-labelling-employees.271163/ https://teslamotorsclub.com/tmc/threads/tesla-lays-off-over-...
- UncleEntity 4y ago
  Google’s robocars seem to be taking a very conservative approach to integration with the motoring public. Don’t get me wrong, I irrationally hate the robots, used to see them all the time around Phoenix, was actively gunning for them and only had one time I could have rammed one[0] with it not being my fault. Which says a lot because it was (well, still is) a life goal to take out a robocar so I was really trying. I used to see them doing a bunch of really stupid stuff but never anything that was outright dangerous, relatively speaking. One did stop in the fast lane when everyone was zipping by at 70+ because it couldn’t figure out how to merge into the left hand exit ramp but plenty of snow birds do that too so it’s not completely unheard of. [0] (un)fortunately I was behind on my rent and couldn’t afford to spend who knows how long dealing with an accident investigation so the robot got to go about its day unmolested.
  - rootusrootus 4y ago
    Google's approach is more defensible. They are careful, default to conservative behavior and stop when confused, and had professional drivers for years. And even then they're a road hazard. At least they hold liability if anything happens, though, which is an important step.
- kaba0 4y ago
  Also, since parts of it is a blackbox, doesn’t any update literally clears out any former result and should be revalidated from scratch?

ndsipa_pomu 4y ago

The sensible method is for the manufacturer to perform lots of testing and make available the data. Legislators can then allow limited real world usage to determine if there's problems and then expand that usage as appropriate. e.g. allow autonomous vehicles to take control on motorways (sorry, that's a UK term - is it freeways in the U.S.?).

jjoonathan 4y ago
That's exactly what happened. There was a back and forth with NHTSA for years before the beta started going public, and the rollout was carefully staged. At first, you needed a high safety score and it was aggressively geofenced. Those restrictions were gradually lifted as they collected data, and it still has attentiveness monitoring and a strike system. Tesla and NHTSA go back and forth on details all the time -- just recently NHTSA demanded some adjustments for stop signs and yellow lights.