I have a Model Y, so don't use uber, but I do loose sleep after a cab ride in NY city. However I did learn some interesting driving style tips from those cabbies: In NY, you can not wait for a clear spot to lane change, so the rule is that whoever's front bumper is ahead has the right-of-way. You get a bit ahead, then just start to change lanes, and let the other guy back off or crash into you, which they don't want to do. This works because it is too crowded to ever get a clear spot to change into. In California this in considered rude or dangerous. In NY it is just how driving is done.
-
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 doesn't read signs properly fsd hard and late braking fsd has decision wobble fsd jerky driving fsd late deceleration fsd sometimes may not detect train crossings fsd stops too early too quickly fsd stops too far behind line fsd tailgating in city streets
The not-exactly-off-topic issue was artificial intelligence, neural network technology, aka FSD end to end AI, to quote the thread title.
Driving style is central. Uber drivers, unlike FSD cars, do learn from their own experiences. Uber drivers who don't respect customer comfort soon become former uber drivers. FSD subscribers who do not like how it drives soon become former subscribers. As many already have done.
Fair point. And my post, after second read, came across rather snarky so I edited it.
FSDB and now SFSD is a newish product and constantly changing. When I first started using it, it was quite raw and scary. In two years it has come a long way, and I respect everyone who has decided it’s not their cup of tea or worth the $. But it has progressed far enough that now we can start questioning nuances of it fitting our individual driving styles. I think that’s overboard. We have 3 settings for aggressiveness. Those can be refined to actually make a difference. For example, turning on Chill mode and the least aggressive setting the car will still accelerate at a rate that is too fast for my wife as a passenger at times.
I also think conflating people’s past behavior with subscribing and / or unsubscribing from FSD doesn’t validate your argument that it needs to learn each families specific driving style to reach mass adoption. It just needs to work better. Just like an Uber driver.
Last edited:
I had neither AP or FSD engagement. The map on the right had changed, showing much more detail/resolution than normal.
In my Intel car at least, the visualization in the left panel is quite different when FSD is engaged vs. standard visualization. I thought that was a normal thing.
Anyway, it was the map change that amazed me.
KArnold
Active Member
You are right, perhaps technically not a requirement for a Robo Taxi. But maybe for a profitable RT as peeps won't use transport they are uncomfortable with. Or perhaps the better question would be "Why not personalize?". Should be cheap and easy for any competent AI down the road to learn and apply automatically, learning like calibrating cameras if not ongoing.
That's more an issue of social graces than anything else. Perhaps the better question is: do you request the same Uber driver when they don't drive in a way that you prefer? Configuring a robotaxi for how it drives may be one of the best unique selling points in the business. As with Uber drivers (or GPS voices), services may offer a variety of driving types by name. Rocket Ralph. Mellow Morris. Scenic Sally. Then there are the silly possibilities with LLMs, where each driving style interacts with the passengers (when requested) in a way that is consistent with the driving style.
Doesn't seem all that silly for the passenger to be able to articulate "slow down" or "drive faster" and have it obeyed.
fasteddie7
Active Member
I would imagine everyone on 14.9 would move since it’s still under the software version for 12.4.x.
There are still ~675 pending updates to 2024.14.9, so it is likely most of the ~600 left on 2024.3.25 have pending updates that just haven't been installed yet.
My guess is that most of the FSD V12.4.2 testers will come from the 2014.14.9 version.
stopcrazypp
Well-Known Member
Per vehicle learning introduces too many variables and unpredictability. I suspect it will never happen. Rather, pre-trained parameters will be used just like it is used today (assertive, chill, minimal lane changes, etc).
Are you ready to sell your car?
www.notateslaapp.com
Tesla Officially Announces FSD Hardware 5.0 and How It Compares to Hardware 4.0 [VIDEO]
Tesla confirms hardware 4 training divergence and announces upcoming hardware 5.0 to hit production lines in 18 months.
www.notateslaapp.com
Last edited:
zoomer0056
Active Member
Well it's too late to sell now that the cat is out of the bag. It's amazing that Tesla really has a hold more than any product I can remember. New hinges (HW5), golden hinges (AI5) just around the corner always causing such a buz ... impressive.Are you ready to sell your car?
Tesla Officially Announces FSD Hardware 5.0 and How It Compares to Hardware 4.0 [VIDEO]
Tesla confirms hardware 4 training divergence and announces upcoming hardware 5.0 to hit production lines in 18 months.
My 2022 yoked USS HW3 MSLR will be for sale in 2027. I figure it will still have some value with 3 years left on the battery waranty. I wonder what the buz will be at that time.
Last edited:
HW5 power won't be fully utilized until 2028-2029. First HW5 cars won't be available until end of 2026 (they said end of 2025 but I think there will be a delay). Then HW5 will run HW4 emulation mode for at least a year. Then they will release HW5 real mode and it will take a year to fix the bugs. I won't buy until 2029.
Last edited:
I will Stop/slow down at the line but only when the line is at or very close to the intersection. When the Stop line is not near the intersection I always Stop/roll to the imaginary line in the road where I can make that decision. Police where I live don't care if creep slowly past the Stop line when it's 5-20 feet from the intersection and then do a very slow roll. I mean that is how the Police drive too so why be different.
dachshund
Member
It's alarming how much energy is needed to train these models. It's an eye opening video, for me anyway. These models don't learn - it's just another hard coded system in a way. It kind of makes me wonder what we've gotten ourselves into, as a society. But I digress.
Driving styles? Remember Elon wants the robotaxi. Driving styles don't figure into that idea.
AlanSubie4Life
Efficiency Obsessed Member
While of course this can also be ticketed, it’s less problematic in terms of communicating with other road users and in terms of the other issues outlined by the court of appeals.
It’s not complying with the letter of the law, but it satisfies the core purpose of the law in many respects. Whereas stopping 5-10 feet before the line of course does not - which is why it is so problematic, as discussed at great length in the earlier linked post.
Not to mention it serves zero purpose and there’s no good reason for a partially autonomous system to be doing this. Legally of course FSD will always need to stop at the marked line, which is going to lead to some non-optimal stop behavior at some intersections (stopping can be highly optimal but the position may be largely useless). But it is what it is, and FSD should start doing so, and then quickly proceed.
Last edited:
Are you ready to sell your car?
Tesla Officially Announces FSD Hardware 5.0 and How It Compares to Hardware 4.0 [VIDEO]
Tesla confirms hardware 4 training divergence and announces upcoming hardware 5.0 to hit production lines in 18 months.
No. With Tesla's knack for over promising and under delivering, I won't hold my breath. Remember Elon said HW3 was 'a game changer' and over built with redundancy. There's been too much trial, error, and obsolescence. In the end HW5's increased capabilties will help but only time will tell if TSLA can ever make a refined consumer grade FSD product.
