FSD v12.x (end to end AI)

[Twiglett]

@JulienW I spent a bit more time with the speed limit blindspot weirdness and found something "interesting".
FSD is using the type of road to figure out which stack to use and is using V12 stack on 65-70 two lane roads, it only goes to V11 when its divided.
At least in this local Texas area, I can see it happen on roads that flip from two lane to divided and back. Its really rather annoying.
Can't test that until this weekend, but at least that will be with FSD12.3.4 unless they push another version

 

[Joboo7777]

I seem to recall Elon saying there would be significant updates to V12 in both March and April. Does anyone know what was expected in April?

 

[JulienW]

I seem to recall Elon saying there would be significant updates to V12 in both March and April. Does anyone know what was expected in April?

Yes, in April we will get "two weeks" and in May "by the end of the year".

 

[mgs333]

With news that Tesla laid off 10%+ of their global workforce and that some groups were hit close to 20%, I wonder what was the damage to the AP/FSD team and whether they're still planning on updates every 2 weeks. My guess would be it was less than 10% to the AP/FSD team as that is the current savior per Tesla's narrative.

 

[OxBrew]

Although potentially the observed latency is actually longer than the time it takes 12.x to process photons to control end-to-end as it could be carrying over human reaction time from training.

That's a scary thought. 1-OK/cool if it's a normal move (like slowly changing lanes when no one is around), 2- very not cool if its reacting to an emergency.

How would they define "emergency" without any "coding"?

Lane wobble is really bad in several spots for me, to the point where it just gives up with no warning. Very repeatable, like 80% of the time in one spot. Just turns off FSD and keeps rolling slowly.

Other spots it often does wild s turns back and forth across 4 lanes in one direction, while a phantom double yellow line jumps around from lane to lane, never using any signals. It's just totally irrevocably lost in those moments.

 

[Daniel in SD]

Hopefully they fired whoever was in charge of Chuck’s UPL. Ridiculous that they can’t get a single nine of performance. I’m done betting on this until Ashok leaves.

Chuck is not accepting excuses.

https://twitter.com/x/status/1779593871451721790

 

[Todd Burch]

Hopefully they fired whoever was in charge of Chuck’s UPL. Ridiculous that they can’t get a single nine of performance. I’m done betting on this until Ashok leaves.

Chuck is not accepting excuses.

https://twitter.com/x/status/1779593871451721790

Not sure if joking.

You remember a few weeks ago when vehicles were gathering data at that intersection?

I assume you understand that data has not gone through a cycle of training and is not incorporated in the neural network in 12.3, correct?

 

[Todd Burch]

Meanwhile, FSD drove 45 minutes to work and another 50 minutes back today. I only intervened to keep the speed up and to initiate a lane change a few times because I was impatient.

And there was decently heavy traffic today--successful zipper merges, correct decisions to go, etc. Can't wait to see the next iteration of the network.

 

[AlanSubie4Life]

I’m done betting on this until Ashok leaves.

Glad you finally have come to your senses.

I assume you understand that data has not gone through a cycle of training and is not incorporated in the neural network in 12.3, correct?

It is not like they haven’t had a year to train the network to do these turns! I think that is the disappointment here. End to end has been around for a year!

Remember it struggles at other unprotected lefts when there is traffic present. This is part of driving. Not being able to do a key skill seems like a major shortcoming of the original training.

It makes me suspect that the additional training will not solve the problem. But hopefully I am wrong!

 

[JB47394]

2- very not cool if its reacting to an emergency.

What emergencies are you concerned about where the gap in FSD's reaction time vs an "ideal" reaction time would make the difference?

 

[Daniel in SD]

Not sure if joking.

You remember a few weeks ago when vehicles were gathering data at that intersection?

I assume you understand that data has not gone through a cycle of training and is not incorporated in the neural network in 12.3, correct?

I don’t think you understand how many beers I’ve lost to @AlanSubie4Life. They’ve been there many times over the years. How much data do they need? How much more data to get the second nine of performance?

https://twitter.com/x/status/1551605579419979784

https://twitter.com/x/status/1558051922820595714

https://twitter.com/x/status/1558213384881016837

https://twitter.com/x/status/1709987871082316163

https://twitter.com/x/status/1761060354891673943

https://x.com/chazman/status/1765455089127285203?s=46

 

[Daniel in SD]

It does sound like they have got three people working on it now so maybe we’ll see some progress. Still not betting on it!

https://twitter.com/x/status/1766115061959893249

Hopefully with the loss in headcount they’ll have more money to beef up the UPL data collection team.

 

[enemji]

I know I am totally wrong and Tesla has never published nor acknowledged it but yet I have a feeling that some localization is happening.

Why do I have that feeling?
Everytime there is an update, the capabilities appear to be regressed in how unique road situations are handled and then over the next 2-5 days, they all come back as they were before. This makes me believe that during an update everything is wiped out and then the localized data of major metropolitan areas(at least) is trickle fed back to those vehicles in that region.

This has happened most recently with v11.3.9. I was very satisfied in how it managed certain situations, and then they disappeared when it was updated to v12.3.3 and then v12.3.4. But now today it has been handling those unique situations like they were on v11.3.9.

Oh well…

 

[Tronguy]

I know I am totally wrong and Tesla has never published nor acknowledged it but yet I have a feeling that some localization is happening.

Why do I have that feeling?
Everytime there is an update, the capabilities appear to be regressed in how unique road situations are handled and then over the next 2-5 days, they all come back as they were before. This makes me believe that during an update everything is wiped out and then the localized data of major metropolitan areas(at least) is trickle fed back to those vehicles in that region.

This has happened most recently with v11.3.9. I was very satisfied in how it managed certain situations, and then they disappeared when it was updated to v12.3.3 and then v12.3.4. But now today it has been handling those unique situations like they were on v11.3.9.

Oh well…

Or.. and I sneak my personal weird thinking in here.. the car is doing some self-training, neural net style.

Yeah, yeah, I know: It Can't Possibly Be Doing That. Everybody says so.

 

[AlanSubie4Life]

It does sound like they have got three people working on it now so maybe we’ll see some progress. Still not betting on it!

https://twitter.com/x/status/1766115061959893249

Hopefully with the loss in headcount they’ll have more money to beef up the UPL data collection team.

Would be great. This UPL is basically driving FSD capability forward and has been for years. It requires very specific and important abilities to execute correctly. It cannot do it because FSD still lacks certain fundamental capabilities at the moment. These abilities are now obscured and obfuscated by the NN. But the capabilities are still required.

What emergencies are you concerned about where the gap in FSD's reaction time vs an "ideal" reaction time would make the difference?

It’s long been claimed that one of the advantages of FSD is superhuman ability. It would be good to realize that.

Also, if FSD lacks anticipation (seems extremely difficult), it will need to make up for it with faster-than-human reactions.

We are talking about 0.3 seconds here. That’s huge in terms of how much kinetic energy can be reduced prior to a collision.

 

[OxBrew]

What emergencies are you concerned about where the gap in FSD's reaction time vs an "ideal" reaction time would make the difference?

Everything sub one second to avoid a crash, that human reaction times (about 1 second on average) causes a crash?

Its a computer, it just feels like it should respond instantly, and delete any delay from training that's only due to human reaction times.

 

[cyborgLIS]

Yes, often it does, but sometimes it's way too close. I've seen tailgating only 2-3 times, but it's pretty dramatic.

One scenario where it consistently tailgates is when another car changes into the same lane as you, ending up just a short distance in front of you. FSD will not slow down to create distance. That short distance seemingly becomes its set following distance. I haven't let it continue long enough to see what happens if the car in front front hits the brakes.

 

[cyborgLIS]

Speaking of latency, has there been any determination of how much delay there is between the visualization and reality? If the delay was substantial, it would make the visualization less reliable for checking rearward traffic and blindspots, unless the relative positioning of cars was fairly static. I'm sure there are people who shun the use of the visualization for checking blindspots regardless.

 

[JB47394]

Everything sub one second to avoid a crash, that human reaction times (about 1 second on average) causes a crash?

People primarily rely on their reaction times because they are not 24/7/360 vigilant. Stuff sneaks up on them. But a computer system can be eternally vigilant. So it doesn't need fast reaction times; it will see the problem coming, just as people who pay attention do.

Certainly there are cases where even a vigilant driver cannot escape. In those cases either they cannot see the threat and their reaction time doesn't matter, or they can see the threat but can't do anything about it and again their reaction time doesn't matter.

So what's an accident that can be avoided by that eternally vigilant computer with 50 ms reaction times, but not 500 ms reaction times? (Assuming that FSD is that delayed)

 

[aronth5]

Would anyone else like to see an improvement when choosing a destination for FSD to have an option to automatically zoom into the destination allowing you to place the pin where you want the drive to end/park? Sometimes I do this but often forget which then becomes a hassle to try and change when driving. Assumes of course Tesla does a better job of ending at the Pin or close to it like my driveway.