4 ms·

I've had FSD for about 3-4 months and it can do the vast majority of my drives without disengagement. There are some very obvious things it still struggles wit

I've had FSD for about 3-4 months and it can do the vast majority of my drives without disengagement. There are some very obvious things it still struggles with like construction and roads without clearly defined lines. But I just don't use it on those routes. I think Tesla made a mistake trying to take a "boil the ocean" approach that will work literally anywhere instead of focusing on key areas first. It ensures you need to 100% nail it before it can be considered "released" as opposed to releasing with a specific set of criteria to meet.

misiti3780 4y ago
I agree, I have been testing it for a year or so. 1. I have many many multi mile 0 intervention drives 2. it is getting a noticeably better every release Their video model will be very good in a the future if the current release pattern keeps up. Not sure what it will take to be certified by the USG as level 5, but it seems like it will be at the top of this chart in the future.
- croes 4y ago
  Or they will hit a wall with no further progress.
  - misiti3780 4y ago
    not likely for a long time, given they are getting more data every day.
    - jcranmer 4y ago
      Getting more data is useless if it's not the right data.
      - rafaelero 4y ago
        They are getting the right data because their users are reporting exactly the instances where it fails.
        croes 4y ago
        Knowing where it fails isn't the same as knowing what is right.
    - croes 4y ago
      What if it's not a data problem? That's the same reasoning as thinking a faster CPU or more RAM would solve problem. More data could even lead to worse models.
- oliveshell 4y ago
  >Their video model will be very good in a the future if the current release pattern keeps up It appears they’re abandoning the video-only approach: https://techcrunch.com/2022/12/07/tesla-appears-to-be-turning-back-to-radar-for-its-vehicles/amp/ https://techcrunch.com/2022/12/07/tesla-appears-to-be-turnin...
  - misiti3780 4y ago
    sure, but it still going to use a video model, with additional sensors.

misiti3780 4y ago

I agree, I have been testing it for a year or so. 1. I have many many multi mile 0 intervention drives 2. it is getting a noticeably better every release Their video model will be very good in a the future if the current release pattern keeps up. Not sure what it will take to be certified by the USG as level 5, but it seems like it will be at the top of this chart in the future.

croes 4y ago
Or they will hit a wall with no further progress.
- misiti3780 4y ago
  not likely for a long time, given they are getting more data every day.
  - jcranmer 4y ago
    Getting more data is useless if it's not the right data.
    - rafaelero 4y ago
      They are getting the right data because their users are reporting exactly the instances where it fails.
      - croes 4y ago
        Knowing where it fails isn't the same as knowing what is right.
  - croes 4y ago
    What if it's not a data problem? That's the same reasoning as thinking a faster CPU or more RAM would solve problem. More data could even lead to worse models.
oliveshell 4y ago
>Their video model will be very good in a the future if the current release pattern keeps up It appears they’re abandoning the video-only approach: https://techcrunch.com/2022/12/07/tesla-appears-to-be-turning-back-to-radar-for-its-vehicles/amp/ https://techcrunch.com/2022/12/07/tesla-appears-to-be-turnin...
- misiti3780 4y ago
  sure, but it still going to use a video model, with additional sensors.