Tesla FSD V12.3 is HERE: The Good and the Bad

Thanks for listening! Subscribe to the channel.

Show your support and get access to perks:

[MrClaudehammond]

The one safety intervention could be from the car switching from highway stack to city stack right at that moment. Would love to hear from Tesla on this.

[zombietech]

Awesome to see some good progress! That failure with the flashing red light is shocking though. That is super basic stuff, not an edge case. And trying to drive into the barrier was almost another Walter Huang moment

[brianfromireland]

Great to see balanced and measured Tesla news, thanks

[garretthoneycutt]

The red lanterns is an amazing example of an edge case. Incredible!

[eyesuckle]

I'm a bit surprised that in the part of the video where the car overtakes a cyclist, neither of the car's occupants comments that not only did the Tesla move into the oncoming lane where visibility was limited by the curvature of the road, but it crossed a double line to do it. That's not only poor judgment, it's illegal.

Despite such errors, it looks to me like we're getting very, very close. It's handling edge cases that I never would have imagined it overcoming with ease.

[carl-Sp]

Musk said 12.3 was a big update. The bigger the update the more problems to expect. Things improve via many SMALL updates.

[gerrycaragayan]

Thank you. It is not perfect yet but looks impressive already.

[richardhamilton-gibbs]

Love your channel, Jacob.

To me, it seems that Omar is only having problems at night. Tell him that. I think Tesla's night vision may need work.

I've done a lot of R&D in my life. I had a tech team that did the work, but I did the architecture and inevitably, I had to solve the problems. Trouble-shooting is something that commonly got me this response from the tech team: surprise, and the comment "I never thought of that". 😁

I figured Omar must have immediately realized that the FSD problem was caused by the change in camera perception and image processing between day and night. But he might not have thought of that immediately.

Partial explanation: We can't use a human parallel because we already know that Tesla's vision processing doesn't convert the image into a picture that makes sense to our eyes. On a Lex Fridman interview, Elon explained how Tesla is using raw feed from the camera. Tesla started by using the camera output as human-meaningful pictures, but that increased the jitter. Frames with more detail required more computer analysis by the FSD board, which messed with the processing time. 'Jitter' is the variation in processing time, and the more variance, the slower the response: you can't move until you're sure you've got ALL the data processed. Tesla discovered that using raw data from the camera made the analysis and decision process much faster and much more consistent.

So we don't have a clue what the AI is doing with that data. But I'm sure if we give it data from night driving in SF it will figure it out much faster than any human programmers.

I'm expecting great things before May 30, 2024. 9 1/2 weeks? That should ring a bell with some old dudes.

[bjornklotzki]

I think in the last case FSD wanted to turn Right!?( take the right lane) Is that possible? The Navigation showed a different way?

[baldbikeboy]

Yep, it’s not yet perfect. I just got it last night and find that in city driving, 12.3 drives about 5 mph lower than the limit. Step at a time…

[Gargamel-n-Rudmilla]

You are thinking Roboxi problem. Robotaxi is many yeara away.
For an adanced level ADAS v12 is there even with 2.1 in my view.
For me new capbilities must never undermine saftey and the benchmark is a 17-year-old girl driver.
The benchmarket can not be a msn aged 50 with decades of driving experience.
Also there indeed things that FSD now can do that no human can do.
The incident with the Chinese lanterns must be very site specific but also one can see that the hight of the lanerns was not at the same level as a normal trafic light. The key escence of a trafic light needs more work. The lantern's high was differnt than a trafic light anf there are unusually a lot of them and they way too many all slosely tied together. So for a human this would not have mattered.

[importon]

I have FSD. it's not worth this level of hype

[edhill]

Frankly unless you own FSD and have experience with past and the latest version you shouldn't critique it.

[dr_frankenmiller]

Picked up another 10 shares TSLA for payday today Jacob have missed your channel for few months but will be picking up more shares on the cheap. ChatGPT predicting share price $300 by Christmas