JulianW, I looked a bit into the "liquid" NN model. It looks interesting, but I didn't get a clear idea about it learning from experience after it's training. Perhaps it is purely pre-trained, like our cars.
-
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
stopcrazypp
Well-Known Member
Why the focus on learning after training? One of the major advantages of LNN (liquid neural networks) is it actually can reduce compute requirements on the end product, so it can be used in less powerful hardware. Adding post-training learning will increase compute requirements, which is contrary to that goal.
There is nothing that indicates learning after training is required for the driving task. As posted above, your desire for different driving styles can be satisfied by parameters that the NN is pre-trained to handle. In fact, post-training learning will make it so the end model used by the end user is no longer the same as the trained model, which is not desirable in a safety critical application, as that is too unpredictable.
I think the testers are already mostly moved to 2024.14.9, myself among them. The number still pending for 14.9 is larger than the number still on 3.25, but it always takes a while for every car to update. There are still a small pending to move onto 3.25, weeks after it became available.
I think that on TeslFi there were around 5,000 testers. We are now appear to be held on 2024.14.9, perhaps awaiting release of FSD 12.4.x.
I gleaned this by looking at individual cars in the list for each version. The recent converts from 3.25 to 14.9 had their first FSD version in 2022. Those in 2024.20.2 got their first FSD more recently. I don't know exactly when they stopped adding testers, but it was a while back. I also don't know if the tester group will be dissolved, but I assume so, once FSD development if farther along. It makes sense for Tesla to try our new releases of a significant number of cars so that edge case problems can be caught before releasing to the entire fleet . Many of us have tried to get out of the tester group for various reasons, but have not found a way. Like the Hotel California, we checked in but can never leave...
I'm still on 3.25, so I'm one of the special 3%'ers
I'm driving a 2023 MYLR bought in October 2023. Originally on EAP and now subscribing to FSD on month to month basis since the free 12.x trial expired at the end of April. I had got one month subscription in December '23 on FSD 11.x and had dropped it afterwards.
With all that background above, I am also moved from 3.25 to 14.9 so I have a feeling they may be keeping the earlier testers + newer paying customers like myself on the same pool going forward. We can validate this once they release next FSD drop and see if we all get it at the same time or not.
Comparing with that Uber driver or the remotely monitored robotaxies, if they drive crazy (or try to molest you) you can tell them to stop and let you out. Complaints on the app also get the drivers who smell bad fired. Realtime feedback is present in those systems.
But you are right that driving "style" for FSD will just have to be a lowest common denominator type, think limo driver. Smooth, safe, courteous, not too fast and not too slow. Avoiding speed bumps and pot holes and funky neighborhoods is the sort of learning that limo drivers do, but FSD not so much. Yet. Traffic is an example of rapidly updated map data which may provide a model for some of these other details.
As an example, I did a couple FSD drives from Oakland to and from north Berkeley. Routing was right across the top of the UC campus where some event had pedestrians everywhere this weekend. The car traffic was fine, but the peds weren't counted in the route planning. On the way home I had to manually drive a different route to avoid that congested but low car traffic area. "Hey Tesla, take a different route to avoid the campus." This is not in FSD's vocabulary. Yet. I think the Berkeley locals know to avoid that area, so there is not a lot of car traffic to make the routing algorithms avoid it.
A decent Uber driver would do the same thing. Google could map pedestrian congestion, but they don't yet, so FSD is blind to it. So far. Cars with all these cameras could talk with each other, but they don't, yet.
A little OT but here is a video from......well I'm not sure and a little ambiguous but made in Shanghai. Also it is not clear if it is L2 or driverless L4. I'm guessing L2. Maybe it is a software compony named DeepRoute AI. But damn if it doesn't look VERY Tesla like including the UI. It is sped up most of the time so appears smother that it actually is. There is a good bit of Tesla lane wobble and the steering wheel seems to jerk more like V11.x.
Also it is narrated in first person by an AI bot but is in English. First person?????? AI bot talking as a car is actually NO person
.At the end it asks for Alan to personally rate it 1 to 100.
I'm guessing he will give it a 42.
Last edited:
stopcrazypp
Well-Known Member
Waiving all responsibilities isn't necessarily required anyways in China, given China tort law doesn't support the multi-million dollar payouts that is commonly seen in the US (even in cases of fatal accidents).
That said, unlike the US, there are already other non-Tesla door-to-door L2 systems operating in China, and they also waive all responsibilities (just like FSDS is doing), so it's not unprecedented.
The only negative for Tesla is they are a foreign make that is completely independent (not in a JV with a domestic make), so they are not given the same treatment as domestic makes affiliated with the government, in terms of squashing negative news in the media. This is how for example the Tesla brake failure scare back in 2021 was able to spread wide on the Chinese internet, whereas similar incidents that happened with domestic makes are immediately quashed before news an spread.
And it's hitting curbs too. And it's stopping up to 10 feet before the line. NHTSA needs to get off their asses and order an immediate recall of v12!
Having never traveled to China I was wondering if the driving laws and road layout throughout China are more or less standardized as compared to the US where there are differences from state to state? Especially in China's metropolitan areas? Are the roads well marked and painted? Just wondering how this will affect FSD's capability in China?
FSDtester#1
Large Member
stopcrazypp
Well-Known Member
My point is no it doesn't have to be lowest common denominator! You can simply pre-train different driving styles and have it in the NN (just like there are chill / assertive profiles right now). None of that requires learning on the fly.
You are talking about route planning which also does not require learning after training. The route planning is currently done by the standard navigation system that is hard coded, but you can certainly add a hard coded or NN function that takes more factors into account (I have advocated a macro lane planning function, which plans out the exact lanes the car will plan to use prior to even setting off), and that does not require learning on the fly either.
This is how a good human driver does. plan ahead and take the correct lane at your exit so you are well positioned for your next turn. Completely agree with you that FSD need to be trained in this regard to be more efficient and reduce unnecessary lane changes.
I think that's not the fundamental problem. In the areas I've been, drivers use every bit of the road and virtually ignore lanes, so I'm wondering how FSD will work there. Note that the drivers are not aggressive (much more of a problem in the US) but simply try to get to their destination as soon as possible.
It sounds plausible, so I offer this tiny datapoint.
My 2018 3 got FSD 10.3 way back when, this year I kept the 3 and transferred FSD to a new Y.
Soon after, the 3 ended up on 2024.14.7 and the Y on 2024.3.25
About a week ago, the 3 got 2024.14.9 and yesterday the Y also got it.
AlanSubie4Life
Efficiency Obsessed Member
willow_hiller
Well-Known Member
The Tesla ADAS drivers are back in Chuck's neighborhood:
Chuck first spotted them testing what might have been 12.4.1 on May 23 and got it himself on June 7. So maybe 2 weeks from today we'll see 12.4.2.
Chuck first spotted them testing what might have been 12.4.1 on May 23 and got it himself on June 7. So maybe 2 weeks from today we'll see 12.4.2.
