-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
FSD v12.x (end to end AI)
- Thread starter Buckminster
- Start date
-
- Tags
- #fixphantombraking autopilot chill is not chill elon is an ass fix fsd extreme late deceleration fsd does not commit to go fsd does not maintain proper speed fsd hard and late braking fsd jerky driving fsd late deceleration fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
AlanSubie4Life
Efficiency Obsessed Member
Seems unlikely given current behavior! The ability to choose the right gap and perceive traffic is quite good at the moment.
I firmly believe one 9 is possible.
Just needs more rocket-like behavior. Looks to me like it is encountering some sort of decision paralysis. Not sure why. It should go extremely briskly if the path is clear.
Knightshade
Well-Known Member
.
Do we KNOW that?
Maybe a single 720p B-pillar cam isn't awesome enough at judging speed and and distance to not sometimes make the NNs doubt itself during a maneuver?
Do we KNOW that?
Maybe a single 720p B-pillar cam isn't awesome enough at judging speed and and distance to not sometimes make the NNs doubt itself during a maneuver?
Accidents per vehicle or total accidents? Sticking with the theme of qualitative anecdotes, if there are 100 accidents at the intersection but 100k vehicles/d pass through, versus 10 accidents on unprotected left but only 1000/d vehicles, then the intersection is 10x safer despite an order of magnitude more accidents. I guess my main point being, it’s hard to make a determination without actual quantitative data.
Last edited:
Iain
Member
That's why they added cement islands in many places. To keep people from being able to turn left out of parking lots.
AlanSubie4Life
Efficiency Obsessed Member
Yes I think we do. I haven’t seen any evidence to the contrary lately.
As I said before I would not be surprised at all if it sometimes screws that estimation up entirely, but the likelihood of that impacting evaluation of the first 9 is low.
Remember the goal is the first 9.
Look at the decisions it is making. Generally solid go/no-go decisions and really does wait for openings - haven’t seen any going when not clear recently (would expect to see this occasionally if there were truly a frequently occurring detection issue). Seems to have these problems:
1) Stops too early.
2) Poorly executed inconsistent creep.
3) Doesn’t just wait for traffic to clear in both directions (simplest algorithm)…though there seems to be evidence they are actually more biased to this than v11.
4) Stops in middle of the road.
5) Uses the median (inappropriate here in most cases, even if humans do it).
Don’t do these things.
Last edited:
Found this in the investment thread. For those who think FSD might be better to skip Chuck's turn. Interesting.still struggling with folks calling Chuck's turn an "UnProtected Left turn"
"I stopped watching his videos when I realized it is far faster to go to the light and turn left with all the other cars seen in Google street view instead of trying to be quicker but more dangerous even as a human driver. It's a whopping .2 miles. During rush hour probably actually faster than waiting,"
See map and post below
Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable
Here is the update from my normal ~13 mile drive on FSDS v12.3.6:
Notes:
Notes:
- Acceleration curve from stops seems to be toned down a little. (Will have to confirm in future drives, as I wasn't the lead car at a lot of the stop lights.)
- Braking seems a little more aggressive.
- I didn't experience the "FSD Wiggle" on any of the lane changes.
- I didn't have to press the accelerator to get it to go a reasonable speed. (In prior versions I had to push it all the time to get it up to speed.)
- I didn't come across any yellow lights to see how it performs on them.
Last edited:
FSDtester#1
"Circle Jerk Old Guard" per cyborgLIS
It looks like the different versions are for cars with different Chips and uss vs no uss
I think the difference now is they've tried to speed things up, throw caution to the wind, give the impression the system is more capable than it is, and rely more on very understanding supervisors versus taking more time to process the scenery, achieve higher confidence, and do less harm.
As long as there isn't a demanding scenario FSD tends to work. And they relay on crutches for some demanding scenarios like stop signs.
Hard to say. There's a 9/2023, non-USS build of a M3 sitting in the garage at the moment, getting its update, and it's going from 2024.3.15 to 2024.3.20 (i.e., FSD 12.3.4 to 12.3.5), when it's clear from Teslafi that people getting 12.3.6 are coming from 12.3, 12.3.3, 12.3.4, and 12.3.5.
On the other hand, a check of a 2021 MY, also in the garage, says, "You've burned your check for the day." That car has USS.
Got a coin? I'll flip it for you
.
Haven't we basically accepted the fact we'll have to wait for 12.4 before we see any sexy improvements? And hopefully many more of those than regressions?
This has been discussed many times. In heavy traffic, I would definitely opt for the right turn followed by u-turn. And this makes even more sense now that FSD is starting to make u-turns, albeit not super smooth ones yet.
Another factor is that in some states or locales, it is illegal to use the technique of pulling into the median, and waiting for a gap in the the right-to-left traffic lanes. It's illegal here but people do it all the time and it really clogs things up for people turning left out of the high-speed traffic into the residential streets- you have cars coming out blocking the way, and nobody can see anything safely.
On the other side of the argument, Chuck has mentioned several times that there have been numerous accidents at the traffic light intersection, but none he's aware of at his tricky left turn.
I recognize that whatever my views on this, I can stand back and consider the interesting progress of FSD cameras and AI trying to conquer something that humans are doing everyday, advisable or not.
I don't know how you measure the arrival of "the first 9". I'd settle for "Wow, this does so many things better than 12.3, like being able to consistently handle my funky intersection down the street."Any beer bet that 12.4 will be the one that finally gets the first 9?
Knightshade
Well-Known Member
Curious if you can cite the law in question? A relatively quick search found a lot of folks in various states saying they heard this was illegal and asking if it was, and then lots of folks finding they can't find such a law on the books.
I do that myself all the time. Not only is it safer, but keeps me and the traffic moving. I don’t know what ID10T AV engineer thought it is best to cross both ways of traffic in one shot. S/He should be fired, and if it was Elon who pushed for it, he too should be fired
Iain
Member
Similar threads
