Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

Wiki MASTER THREAD: Actual FSD Beta downloads and experiences

This site may earn commission on affiliate links.
I'm getting quite used to stopping autopilot when I see other cars behind me.

If you already have your route set, doesn't FSD still gather data regardless of whether you use it?
Its complicated. Depends on what triggers they have set.

But the way to give specific feedback would be to use FSD and disengage or press the "report" button. This is the most valuable feedback since we are telling them directly what FSD did at that moment was wrong.
 
  • Like
Reactions: Phlier
Question about lane changing. I'd swear that the prior release allowed you to set the lane change to confirm. Is that not true? I was surprised that mine just changed lanes today without warning. When I checked the lane change options are grayed out while in FSD. At least when driving.
Change lane confirmation is only for NOA (freeways) not FSD (city).

But - when FSD puts the blinker on to change, you can cancel it.
 
Last edited:
So there is a reason I am not seeing many FSD Beta videos in LA. Honestly I was hoping 10.4 was going to be better but it's not. It is so dangerous in a city like LA. Within 2 minutes of turning it on I had to disengage. There was a light way ahead with a pedestrian on the corner. It phantom braked to a stop about 300 feet ahead of the crosswalk. If there was a car behind me they would have crashed into me. In LA people ride your tail. Between that and purposely trying to crash into other vehicles without warning, I do not see how this can work in a busy city with aggressive drivers like LA. When I look at videos of people smoothly driving, I do not understand. My experience is not like that at all. Maybe it is the new refresh S, or maybe LA, but I gotta turn it off and wait for a few updates before I try again. I understand it is "Beta", but autopilot is also "beta" and it's mostly trustable. This feels more alpha to me. Very saddened by this because I wanted to participate and help make it better.
I’ve had good drives, tolerable drives, and bad drives in LA. My bad ones were mostly where they were doing a lot of construction and had road shut downs. But even the good ones had me intervening a fair amount by hitting the go pedal when appropriate.
 
  • Like
Reactions: JeremyJ
I understand it is "Beta", but autopilot is also "beta" and it's mostly trustable. This feels more alpha to me. Very saddened by this because I wanted to participate and help make it better.
Definitely fsd beta is not beta in the sense of AP. Its a real "beta" which requires constant care and supervision. If AP is a 10 year old, FSD Beta is like a 2 year old.

But I do see people try to use FSD in various cities with tough traffic. I don't think LA is particularly worse than others (I say the same thing about NY, NJ, Chicago, SF or Boston).
 
Interesting moment on an unprotected left from PCH onto a minor street in Redondo Beach. The car felt like it was going to turn into a stream of oncoming cars, or at least the steering wheel started to turn to the left. Maybe the car would have hit the brakes and stopped with the wheels already tilted to the left for a quick response for a turn when a gap appeared, but no way was I risking that so I took over immediately by taking control of the wheel to disengage, flipped the gear stalk slightly up to disengage TACC, and then hit the snap shot button to boot. That is completely unacceptable behavior. One, obviously the car can’t just barrel into an oncoming stream of cars. Two, even if it was going to wait, the car should not be turning the wheel until it is time to go. That reassures both oncoming drivers who won’t get nervous seeing your wheels tilted like you’re getting ready to turn in front of them, it reassures the Tesla driver who doesn’t have to worry about the car creeping straight into oncoming traffic, and it protects against a rear end scenario so the car won’t be pushed forward into oncoming traffic if some idiot hits it from behind.

I’ve never understood why human drivers pre-turn their wheels for turns. Power steering makes turning the wheel quickly for even a narrow gap ridiculously easy so you’re not saving any time, and I’ve seen some human drivers accidentally creep too far forward carelessly forcing cars to swerve around them. Tesla really needs to fix the logic on the turns so Teslas don’t do the same stupid thing.
 
  • Like
Reactions: JeremyJ
Interesting moment on an unprotected left from PCH onto a minor street in Redondo Beach. The car felt like it was going to turn into a stream of oncoming cars, or at least the steering wheel started to turn to the left. Maybe the car would have hit the brakes and stopped with the wheels already tilted to the left for a quick response for a turn when a gap appeared, but no way was I risking that so I took over immediately by taking control of the wheel to disengage, flipped the gear stalk slightly up to disengage TACC, and then hit the snap shot button to boot. That is completely unacceptable behavior. One, obviously the car can’t just barrel into an oncoming stream of cars. Two, even if it was going to wait, the car should not be turning the wheel until it is time to go. That reassures both oncoming drivers who won’t get nervous seeing your wheels tilted like you’re getting ready to turn in front of them, it reassures the Tesla driver who doesn’t have to worry about the car creeping straight into oncoming traffic, and it protects against a rear end scenario so the car won’t be pushed forward into oncoming traffic if some idiot hits it from behind.

I’ve never understood why human drivers pre-turn their wheels for turns. Power steering makes turning the wheel quickly for even a narrow gap ridiculously easy so you’re not saving any time, and I’ve seen some human drivers accidentally creep too far forward carelessly forcing cars to swerve around them. Tesla really needs to fix the logic on the turns so Teslas don’t do the same stupid thing.
Not to mention the wear and tear on the steering column moving the wheel while stopped. A new future revenue source for FSD cars. ;)
 
Anyone had their brake pads changed yet?:eek::eek::eek: Thought about it for 10.4 but looks like 10.4 is going to eat them up too so might as well wait.🤣🤣🤣

Also the Beta is SOOOOOOO inefficient that Tesla may need to offer 2 EPA milage ratings 🧐 in the future. One regular and one (WAAAAAAY lower) on FSD. Here is last weeks.
Definitely inefficient. I drove up to a local mountain a couple days ago and normally on the way down I gain 20 miles of range due to the regen. This time I lost 5 miles of range. My car rode the brakes most of the way down.
 
Definitely inefficient. I drove up to a local mountain a couple days ago and normally on the way down I gain 20 miles of range due to the regen. This time I lost 5 miles of range. My car rode the brakes most of the way down.
I’ve taken to disengaging Beta as I see imminent slow downs, stops, and down hill slopes ahead to maximize regen. Basically exactly what I was doing with AP before I got the Beta.
 
Never see a compilation like this before, where a driver attempts the same obstacle from FSD Beta version 8.2 up through 10.4, so I thought this was worth a share:


Definitely shows both overall progress and occasional regressions.
I bet you could see the same breadth of results out of the same FSD version, the path jumps around all over randomly, it's running off a coin flip.
 
  • Like
Reactions: AlanSubie4Life
Never see a compilation like this before, where a driver attempts the same obstacle from FSD Beta version 8.2 up through 10.4, so I thought this was worth a share:


Definitely shows both overall progress and occasional regressions.
I think 20 tests from the same version, compared, would be a bit more informative. Also removes the variable of changes in the environment over time.
 
10.4 seems about the same, as fully expected. Veering & ping-ponging abruptly between lanes on merges (hadn't done that before, new behavior on this trivial merge, but entirely possible it was just luck), slowing suddenly for flashing yellow overhead lights (for an old and now non-existent school zone), and phantom braking for shadows, as normal. Not enough stops to judge smoothness there (I expect the same - “butter smooth,” but think frozen butter here). Haven’t been able to have it successfully do my unprotected left yet (0 for 1 on attempts; disengaged because it did not advance crisply to the point of visibility, immediately, as a human would - traffic behind was present so I had to use a very strict disengagement policy). Cutting off corners in neighborhood as usual. Only 5 miles so far but a sufficient sample size to say it's about the same.

Overall exactly as expected. We probably have at least a year (at least!) to go before it shows some semblance of smooth driving style. Lots of tricky features to add, and lots of refinement. Maybe much longer, given the current state of NoA! (Especially since Elon has said they put a lot of work into NoA jerk minimization, yet it still can't minimize jerk!) Maybe there will be some sort of perception overhaul at some point which leads to much better results. Probably not though.
When I read your drive summaries they are very similar to what I experience yet others seem to have reasonably smooth drives. Why is that? In the beginning I was definitely causing some engagements by preventing the wheel from sliding thru my hands but that problem is ancient history. I just came back from 2 more rides and 10.4 is just plain awful compared to 10.3, With 10.3 I was getting some zero disengagment drives and many with just 1 or two. 10.3 for the most part was predictable. For me 10.4 is pretty much unpredictable chaos. Just wish I knew why. The only positive today was less phantom breaking.
 
Last edited:
Individual cars don't learn. It is a neural net and all "learning" is centralized and for the fleet.
This is a another reason why it will be years before FSD surpasses the comfort of human drivers IMO. It is a locally trained NN with higher processing speed and memory than any vehicle, albeit with fewer cameras. Also local context and experience. I hope they add locally trained components since most of us drive 80% of the same routes weekly. Even if I think I won’t see a self driving car drive better than me, it is still amazing if it gets close and I can trust it enough to close my eyes.
 
Last edited:
anyone else experienced an issue with gradually reducing speed limits - where the car correctly spots the speed limit changes 75-70-55 for instance - but doesn't actually slow down. This has had me doing 75 in a 55 several times.
It reduces the set speed but maintains its current traveling speed, only slowing down if its blocked by a car in front.

Also, the auto high beam setting that is re-enabled each time it flips from old stack AP (speed limits above 65) and new stack AP (speed limits below 65)
I haven't had it not slow down at all, mine just slows down stupidly slow. It takes about 10 seconds for it to reduce its speed by about 5mph so when I come into town and the speed drops I end up doing 50 in a 25 if I don't press the brake. I have noticed this whether I use the scroll wheel or let it detect the speed limit signs. I will say it only seems to do this at some speed limit signs and at others it drops the speed like it's suppose to. I am unsure if it's a difference in mapping data or what. I kinda get the feeling we are relying a lot more on map data than Elon wants to make it sound just based on how it acts with mapped vs unmapped stop signs, out of date road nav data, and then with this speed limit thing, although it's just a guess with the speed limit thing.