FSD Beta Videos (and questions for FSD Beta drivers)

[heltok]

What a weird intersection. Supposedly the car doesn’t learn between attempts on its own, but it definitely did look like it handled it progressively better each time.

It doesn’t learn, but it has some long/short term memory memory.
Long short-term memory - Wikipedia
Exactly what it will learn to remember we don’t know, but it will be things that are useful to improve the loss function at training time. And some road layout from the past seems reasonable that would be useful to learn to remember.

 

[Frenchie]

Enjoy!

 

[Todd Burch]

wouldn’t you be running the red light right in front of you though?

I'm pretty sure the red light is for the controlled intersection right after the 90 degree turn...but it is facing the intersection where the stop sign is so that drivers who haven't yet made the turn can see it.

There is also a red light AFTER the 90 degree turn for the 1 or two cars who have already turned.

 

[mhan00]

Nice little drive in downtown LA from Vincent and friend. Nice to be able to use my fading mandarin. Passenger was impressed at the end when Vincent asked him his impressions and said he thought that it was so good they needed to find harder stuff for the car to do.

extra little bit at the end where they found a more complicated maneuver with a road with unclear lane markings, a metro line train, and some barriers and construction and decided to record the car handling it again.

 

[Mardak]

Can anyone tell if the gore point is visualized differently? Kinda looks like just red road edges on both sides. Although a little bit concerning the predicted path is to go straight towards it:

 

[n.one.one]

FSD Beta 48.35.1 (build 10) 9 minute drive with zero disengagements - 9:29 - Kim Paquette

 

[n.one.one]

Awesome FSD Beta Night Drive, Going Back Home to my Local Supermarket. Version 2020.48.35.1 - 32:58 - TesLatino

 

[mhan00]

Enjoy!

Not sure why the car decided to drive straight for a barrier on a closed road or why it’s having such trouble with some curbs.

FSD must be leveraging a lot from map data because it’s doing a horrible job in Chicago right now. Trying to go into opposite lanes with oncoming cars in them should be at the top of things FSD should not be doing.

 

[n.one.one]

FSD Beta 10 - No Disengagements Here! (2020.48.35.1) - 15:22 - teslatrey

 

[mhan00]

FSD Beta 48.35.1 (build 10) 9 minute drive with zero disengagements - 9:29 - Kim Paquette

Nice little pallete cleanser from Frenchie’s video. FSD in Chicago scares me right now with its attempts to drive into medians and towards oncoming cars. Still has some impressive maneuvers even still, but the screwups are very bad.

 

[aronth5]

Very disappointing. I'm starting to wonder if FSD is even possible.

FSD seems to work reasonably well in residential and low density city driving but it really struggles in high density urban settings especially older cities. There is a reason why Waymo selected a warm climate "newer city". Reduces many of the complex variables found in older cities. (i.e. Chicago, NY, Boston).

I guess that is to be expected but still discouraging as you've noted.

 

[Battpower]

Enjoy!

It's crazy that there are such positives and such negatives! If this video was from 2 years ago, I feel that it would have absolutely been universally mind blowing. I think false dawns and elapsed time since first promises made for FSD have numbed us and also make the dissapointment over some of the 'fails' more poignant.

I'm suprised there are not (obvious) frequent phantom braking events driving under The El.

It seems that some of the 'decisions' that don't work well in mainstream FSD NoA (on and of ramps, merging under difficult conditions) are still problematic. I don't think I would have been quite so excited at 'missing my turn' and having played it safe. Of course better than an accident, but not really a 'positive' result.

Based on progress so far, I would personally regard this as a driver aid for a long time to come, especially if I was on unfamiliar roads. Handling of temporary signage / diversions is difficult at best, but signs can be very varied and not necessarily correctly located. Sometimes the wrong signs are used. Not the car's fault of course, but still needs to be handled.

 

[aronth5]

If FSD isn't possible, Tesla's stock price will take a huge beating. Will have to be a bit more patient. Maybe dojo will speed things up, but this isn't looking very good. Anyone who thinks Tesla can keep an $800 billion marketcap by just selling cars is being delusional.

And yet beta 10 is getting some favorable reviews by some testers.
"First drive with 2020.48.35.1. This is the biggest leap forward so far. Much better in a lot of ways!"
https://www.youtube.com/results?search_query=2020.48.35.1

For now just assume complex cities will take a lot longer but that should be expected. As long as Tesla can expand the beta and continue to improve on FSD the stock price will stay ok with China's expansion and the Texas and Germany factories coming on line later this year.

 

[n.one.one]

Beta 10: First Impressions (2020.48.35.1) - 25:01 - Trevor Mahlmann
** First FSD beta video from Trevor Mahlmann but he has a very active YouTube channel geared to SpaceX and NASA. **

 

[Mardak]

The thing that annoys me is that I was working on a FSD Beta 9 that @Mardak suggested. I'm yet to decide if I finish that test or pause that and do testing on FSD Beta 10.

Thanks for trying it out! It's definitely a lot of effort to test out so many of the potential 30 paths for a 6 corner intersection, so it's fine to spread it over multiple versions. Also quite impressive that FSD beta is even able to make predictions for the 6 ways even when occluded:

 

[n.one.one]

Beta 10: First Impressions (2020.48.35.1) - 25:01 - Trevor Mahlmann
** First FSD beta video from Trevor Mahlmann but he has a very active YouTube channel geared to SpaceX and NASA. **

A few observations after watching this video.
I don't remember seeing projected turn paths for other vehicles before this (1st image). Has it been there before this version?
In the 2nd image the parking lot visualization seems more detailed/robust than I remember from previous versions.
Any thoughts on either of these items?

 

[n.one.one]

FSD Beta 48.35.1 (Build 10) unedited 30 minute drive - ONE disengagement. - 30:38 - Kim Paquette
** I think this is the longest video we've seen from Kim Paquette. **

 

[n.one.one]

v10 in action - 11:23 - Frunk Puppy

 

[ZeApelido]

If FSD isn't possible, Tesla's stock price will take a huge beating. Will have to be a bit more patient. Maybe dojo will speed things up, but this isn't looking very good. Anyone who thinks Tesla can keep an $800 billion marketcap by just selling cars is being delusional.

I don't think FSD has been priced in that much. A high confidence in FSD robotaxi scenario for Tesla would be worth another trillion dollars at least.

FSD as a driver assistance still has value, as if you can sell a $10,000 FSD to a bunch of your cars, you may say in a few years you could generate 1 million cars x $10,000 = 10 billion dollars of mostly profit per year.

 

[Iain]

I don't think FSD has been priced in that much. A high confidence in FSD robotaxi scenario for Tesla would be worth another trillion dollars at least.

FSD as a driver assistance still has value, as if you can sell a $10,000 FSD to a bunch of your cars, you may say in a few years you could generate 1 million cars x $10,000 = 10 billion dollars of mostly profit per year.

I think you are right, it's not priced in yet. I think we are in the early days and it could be worth a whole lot looking at the progression of the beta videos.

Elon is looking at FSD and it's value to the consumer. Tesla bull asks Elon Musk for Full Self-Driving value to be added with vehicle trade-in