No doubt there's much more profit potential. In the past he's mentioned a large price tag and subscription. But we're still struggling with L2. Maybe HW5 will pull out all the stops.
-
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
No doubt there's much more profit potential. In the past he's mentioned a large price tag and subscription. But we're still struggling with L2. Maybe HW5 will pull out all the stops.
It's not inconceivable (sorry, we just attended a screening of The Princess Bride with Cary Elwes appearing in person afterwards lol) that by the time robotaxis begin rolling off the production line, Tesla snags a license for autonomous L4 taxis that can at least work while geofenced (which would be fine for urban transport) with perhaps extra training for that dedicated location?
OTOH, it's becoming harder to imagine our own currenet cars ever being competent enough to be totally driverless, but if it does happen, I hope I'm still alive to see it.
OTOH, it's becoming harder to imagine our own currenet cars ever being competent enough to be totally driverless, but if it does happen, I hope I'm still alive to see it.
Compared to the other FSD cumulative mileage charts shared in quarterly updates with data points for each month, this one seems to be more granular reflecting a partial month of April and sudden increase in slope mid-March for 12.x rolling out to existing customers and new free trials.
Roughly comparing the slope with very early data (and few data points) from before to after 12.x seems to be at least 4x steeper. Previously the monthly usage of FSD Beta was around 75M mi/month, so 4x would mean it's around 300M mi/month? For reference, introducing 11.x made the slope 6x steeper (~March 2023 -- this newest chart X axis seems off).
It'll be hard to say from these charts what is driving the increased mileage rate as there's a mix of existing and new users along with 12.x's significantly better behavior as others have reported here as well as future increased adoption/subscription of FSD Capability. In any case, it'll be interesting if FSD Supervised maintains this higher monthly mileage rate or even accelerates with more improvements.
Roughly comparing the slope with very early data (and few data points) from before to after 12.x seems to be at least 4x steeper. Previously the monthly usage of FSD Beta was around 75M mi/month, so 4x would mean it's around 300M mi/month? For reference, introducing 11.x made the slope 6x steeper (~March 2023 -- this newest chart X axis seems off).
It'll be hard to say from these charts what is driving the increased mileage rate as there's a mix of existing and new users along with 12.x's significantly better behavior as others have reported here as well as future increased adoption/subscription of FSD Capability. In any case, it'll be interesting if FSD Supervised maintains this higher monthly mileage rate or even accelerates with more improvements.
I’m not sure about the need for thousands of good examples. I have no machine learning training, but I read a paper referenced here a while ago that might be relevant. The paper was about correcting errors in LLMs ([2401.07526] Editing Arbitrary Propositions in LLMs without Subject Labels). The idea was that there are already inexpensive techniques that allow you to interrogate the hundreds of millions of weights in your model to see which ones are most significant in producing a wrong answer. You can then tweak those up and down until the model produces the right answer with minimal effect on its responses to other inputs. This tweaking is also inexpensive and can be parallelized.
So, rather than having to curate a massive dataset that produces a perfect model, maybe Tesla could create an initially good-enough model, and then use the described “Locate and Edit” techniques to deal with edge cases. When a new edge case is discovered, you can adjust for it without having to start all over.
I also speculate that you could use the results of this tweaking as the initial values for the next round of training using the full dataset.
Therealm80
Member
Has anyone noticed with the version 11 Highway stack that it is kind of weaving in the lane a little bit like trying to re-center or something like that?
AlanSubie4Life
Efficiency Obsessed Member
I think they are one and the same. What is being described is something like this. Reaction time was 16 frames, about 1/2 second. If it’s going to do this, at least respond at superhuman speed. But good that reaction time is down to a half second, I guess. Nearly to below-average human performance! (Question: I wonder if the NN takes into account its known reaction time in the decision making?)
About a 3-second light. No reason for this, at all.
Also hit the dots near the curb (@MP3Mike - this is too close. About 1 foot, whatever the concrete width is; no reason to do so here).
And cut the corner and drove in the obviously wrong lane (could have been due to accelerator input - tends to actually tighten the corner sometimes now, rather than slaloming wide).
Last edited:
willow_hiller
Well-Known Member
For the first time, we've actually got a short official tutorial for FSD:
A couple of weeks late, but better now than never!
zoomer0056
Active Member
This is called ping ponging, and it’s come and gone but mostly been around since ~ Oct 2020? when did we start this rodeo?
AlanSubie4Life
Efficiency Obsessed Member
Fake news. Autopark doesn’t let you select spaces on USS vehicles and lane change notifications don’t work (on city streets) on any vehicles (booo!).
Last edited:
I noticed the same improvement today with auto max speed. This is the second time on V12.3.3 that auto max changed. Initially auto max was extremely slow for me causing accelerator pushes. Then whatever Tesla did dramatically increased the speed and many complained that FSD was too fast. Now it's better and more useable. Having a lead car didn't seem to matter much.
Two biggest reasons for disengagement now are 1) potholes 2) lane selection especially in Boston where I drove today. If I'm not in an urban area and don't have to avoid potholes my drives are zero disengagements.
AlanSubie4Life
Efficiency Obsessed Member
I missed a “lower” in that post. Anyway as I am sure you figured out, I think manual mode and ASSO mode speeds more there because the limits are 10-15mph lower than elsewhere all else being equal. I am used to seeing 35mph speed limits in Oregon and Washington where they would be marked as 50mph in California with people traveling 55-60mph.Ah ok, so you're saying you need to move to Washington then
I like all three states equally.
I'm starting to wonder if our cars are being micro-adjusted or 'learning' because I expected problems with potholes in my neighbourhood (because of V12 driving reviews) and it was really careful.
I'm noticing more and more people saying it is getting better but we haven't had a new release in something like 7 days, so is it that people are getting used to the quirks or is the behaviour maturing on the fly without waiting for a new release? If this really is due to seachange of using the NN, this time it might really be different, not just incremental lipstick on a pig. (I'll forgive anyone who rolls their eyes at the "this time it is different" reference because I've lived through that sentiment disappointing people so many times in my life, I tend to mock people who say it, and, in this case, hesitate to use it myself.)
What is different this time is I made it as far as I did on FSDS today and am willing to take it out tomorrow and give it a second test (perhaps we'll ask it to drive 'home' to the SC, an easy 3 road drive except for the U-turn to get into the SC so we can finish the update I have that it thinks I still need). That has never happened before; every attempt to use FSD on city streets (or leaving it on at the top of highway ramps to make a right hand turn onto a road leading to a supercharger) has failed on previous versions. The first version's fail was the most spectacular, again just 3 streets to get us home and there were disengagements at every single intersection as it panicked, stuttered, signaled lane changes even though it was a single lane in our direction and we were going straight, and basically scared the crap out of us. Today I can't get the image out of my mind of the car waiting patiently for the traffic to clear enough to make a left turn, then making the turn confidently and the immediate next turn confidently (although, in that case it had a stop sign at which to catch its breath!) If that had been a teen driver on their first drive off our street, I would have been effusive in my praise.
First drive with 12.3.3 on my regular camp run.
Some good bits, some scarily bad.
First really bad bit -car decided the it really wanted to be one lane to the left, so it crossed double yellow lines into the oncoming traffic lane. WTF. Immediately took over before it got too far and reengaged only for it to do it again. I stopped it much sooner because this time there was oncoming traffic. Ridiculous that it chose that option on a two lane road with obvious double yellows.
On certain roads it will not see anything faster than a 50mph speed limit sign. It lets me dial up to whatever speed I need but ignores 60, 65, 70 mph signs. So when the limit changes from 65 to 70 then to 60, it just barrels along at the manual set speed.
Not inspiring at all.
Sure it’s more natural but I’d prefer robotic 11.4.9 that at least avoids oncoming traffic thanks.
Some good bits, some scarily bad.
First really bad bit -car decided the it really wanted to be one lane to the left, so it crossed double yellow lines into the oncoming traffic lane. WTF. Immediately took over before it got too far and reengaged only for it to do it again. I stopped it much sooner because this time there was oncoming traffic. Ridiculous that it chose that option on a two lane road with obvious double yellows.
On certain roads it will not see anything faster than a 50mph speed limit sign. It lets me dial up to whatever speed I need but ignores 60, 65, 70 mph signs. So when the limit changes from 65 to 70 then to 60, it just barrels along at the manual set speed.
Not inspiring at all.
Sure it’s more natural but I’d prefer robotic 11.4.9 that at least avoids oncoming traffic thanks.
In instances like that, I would try and get dashcam to capture… you can set it up to record with a. HORN honk, or just tap the screen.. that sort of video would be very useful for FSDb Engineering.
Iain
Member
Agreed. I have yet to see any video of the car doing anything that dangerous. The only real issues I have had are clearly seen on the route the car is trying to take. I feel the car is driving fantastically but is let down with the route planner
It's been my experience since 2018 that when AP capabilities are up graded some ping pong often appears. In subsequent updates it goes away. I think I've probably seen a dozen cycles of that.
I've been driving with FSD supervised 12.3.3 and all I can say is that it is amazing. It feels like a human is driving. Drives are relaxed and enjoyable. I am very glad I own a Tesla and bought FSD.
Todd Burch
14-Year Member
Past versions have shown some curious form of improvement over time witnessed by *many* people.
What we know for a fact is that the car is not itself "learning". The car does not have the compute horsepower to do any neural net training and upgrade its network weights. However, there are some possible reasons for this:
1. Tesla of course has the capability of programming in some biases, offsets, or limits--various parameters outside the neural net that they can tweak remotely without a version change.
For example, if the neural net says "I want to go speed v", they could start a rollout conservatively by subtracting 3 MPH from the requested neural net speed and feed that lowered speed into the vehicle controller. After a week, they could remotely remove that offset so that every car now actually goes X. I feel that Tesla addressed the curbing issue by introducing a lateral bias to the vehicle in certain scenarios, which is how they addressed it without a new network.
2. They could have clamping around various parameters that they widen over time. For example, they might limit max steering angle left or right of centerline to X degrees at speed Y, then open that up as the system proves itself to allow for more varied maneuvers.
3. Theoretically, data could be crowd-sourced from the fleet, fed back into Tesla's mapping database, then dished back out to the cars when they download map metadata as they drive through a region. So while it felt 42mph was the right speed the other day, 3 different vehicles were observed with the driver "encouraging" the car to go 48mph on that road instead, so they updated the map metadata, and the next time I drove down that road, it used the updated "preferred speed".
We don't really know what's happening for a fact, but any of these options are possible. (Although in a past AI day I believe they did talk about how, at least in the olden days of FSD, they limited maximum steering angle at a given speed for safety--and they almost certainly still do this).
Similar threads
- Article
