FSD Beta v11.x

The next big milestone for FSD is 11. It is a significant upgrade and fundamental changes to several parts of the FSD stack including totally new way to train the perception NN.

From AI day and Lex Fridman interview we have a good sense of what might be included.

- Object permanence both temporal and spatial
- Moving from “bag of points” to objects in NN
- Creating a 3D vector representation of the environment all in NN
- Planner optimization using NN / Monte Carlo Tree Search (MCTS)
- Change from processed images to “photon count” / raw image
- Change from single image perception to surround video
- Merging of city, highway and parking lot stacks a.k.a. Single Stack

Lex Fridman Interview of Elon. Starting with FSD related topics.

Here is a detailed explanation of Beta 11 in "layman's language" by James Douma, interview done after Lex Podcast.

Here is the AI Day explanation by in 4 parts.

Deep Understanding Tesla FSD Part 1: HydraNet

From Theory to Reality, Analyze the Evolution of Tesla Full Self-Driving

saneryee-studio.medium.com

Here is a useful blog post asking a few questions to Tesla about AI day. The useful part comes in comparison of Tesla's methods with Waymo and others (detailed papers linked).

Twelve Questions about Tesla Autopilot Technologies Exposed at 2021 AI Day

On Aug. 19 2021, Telsa held an AI day , exposing more technologies about Autopilot , from perception, planning, data labeling, simulation…

yuhuang-63908.medium.com

 

[mgs333]

Do you actually believe any of the dates he posts? No one I know takes them to be worth anything, even the most ardent Tesla fanboy.

I've posted this before and made the moderators mad but I'll post it again - at this point anyone making comments like "Elon promised it last week and it's still not here!" is either too dense to participate in a conversation or a troll and not worth having a discussion with.

I still see plenty of people online who, at least on the surface, seem to believe his words including his date predictions.

 

[uscbucsfan]

Sure. I even said some of this earlier.

But the fact there's 400,000 cars driving tens of millions of miles a month suggests they don't do it very often otherwise there'd be lots of actual accidents--- versus the narrative from folks who insist it can't go a single mile without "doing something dangerous"




Can you link to them please? I asked a page or two ago for accident cites and nobody had anything other than the one guy who "thought" he heard some blamed on FSD and the one video of an FSD car having its driver ignore a "flooding ahead" type sign and then the car hydroplaned....which is entirely on the driver since all FSD was doing was driving in a straight line.

I'm not saying the # is 0--- Tesla themselves reports an accident rate on FSDb in their 2022 impact report of 0.31 per 1 million miles driven. This is roughly 5 times lower than the 1.53 accidents per 1 million miles in the general US vehicle fleet. I'm saying the number is very low for something the overly cautious among us keep insisting is super dangerous junk.

Thus supporting the case their issues in reporting a lot higher disengagement rates might be due to their overly cautious nature rather than the situations and behavior of the car actually being as dangerous as they perceive.

Both wrecks on are the FSD Beta Facebook group. One merged into another car and another turned into a car into a home depot parking lot.

Both should have been avoided by good testers, but that's the point. Is FSD Beta safer than the average driver or is FSD Beta + a ridiculously attentive driver safer? It's obviously the latter. Beta would wreck at a higher rate if we all just allowed it.

11.4.X does more dangerous things on a higher frequency than 11.3.6 or 10.69. It's safer than old versions like 10.3 because it couldn't make maneuvers. Most testers state the same, including YTers like Chuck. He's had 11.4.X run a red light in heavy traffic and try to enter a lane when a car was coming. Both times he stopped it.

 

[uscbucsfan]

Do you actually believe any of the dates he posts? No one I know takes them to be worth anything, even the most ardent Tesla fanboy.

I've posted this before and made the moderators mad but I'll post it again - at this point anyone making comments like "Elon promised it last week and it's still not here!" is either too dense to participate in a conversation or a troll and not worth having a discussion with.

Some in the investment thread do. They actually believe FSD Beta is beyond Waymo and Cruise and think Robotaxis should be here next year, but most watch Whole Mars and have no clue where the product really is.

It's gotten better, but not at an exponential rate. It's a slow process where every update has advancements mixed with major regressions.

It's why only 1 or 2 people in this forum think it'll be level 5 before 2030.

I'll say, I still love using it, but it's not super close to reliable level 2, much less level 5 at this stage.

 

[MP3Mike]

11.4.6 is still the best version by far for me. Today I only drove 13 miles and only had two disengagements, both because I wanted to take a different route than the nav wanted.

The best new feature is the "keep up with the pack". When entering a construction zone out didn't show to 25 MPH it stayed at 37 MPH matching the traffic. I always had to adjust the speed before so as to not cause a backup.

 

[willow_hiller]

Is FSD Beta safer than the average driver or is FSD Beta + a ridiculously attentive driver safer?

You believe that all 400,000 FSD Beta testers are ridiculously attentive? Or even a significant portion of them?

 

[uscbucsfan]

You believe that all 400,000 FSD Beta testers are ridiculously attentive? Or even a significant portion of them?

Yes, the ones who actually use it, absolutely.

Most people who aren't either strike out, ehich there are a long list who have numerous times, or don't use it at all. There's a large group here who say they rarely, if ever, activate it.

The car forces you to be and the behavior does as well.

 

[BBTX]

Context: 11.4.4 has been terrible on the main road of my neighborhood on every drive to date, for however long we've all been on this *sugar* version, whereas 11.3.6 was pretty decent at it.

Two way, 35mph, no stripe or divider. Tons of phantom braking for no observable reason (not panic stop, just ~5-7 mph slowdown), and hogging the center and rudely waiting too late to move over for oncoming cars.

Not one good drive in all this time. Since it's the main road in and out of my place, I engage here all the time in all kinds of traffic and lighting conditions, and disengage as it screws up and report it constantly and sometimes angrily, sometimes re-engaging to catch more mistakes several times in one run just out of spite.

Today, out of nowhere, it was suddenly much better. It still did both bad behaviors a little bit, but there were only two phantom slowdowns by ~3-4mph versus the usual several, and it hogged the center only once or twice (instead of constantly), and it moved out of oncoming traffic's way much earlier and more politely. I didn't even disengage the whole way to my driveway.

I'm thoroughly convinced it wasn't just conditions or randomness. Something changed. Maybe some stealth update of the dynamic map data in my neighborhood or something?

 

[kpanda17]

Two back to back reboots and stuck on 26.1/11.3.6
Trying to get to ?/11.4.6 or 7

 

[willow_hiller]

Yes, the ones who actually use it, absolutely.

Most people who aren't either strike out, ehich there are a long list who have numerous times, or don't use it at all. There's a large group here who say they rarely, if ever, activate it.

The car forces you to be and the behavior does as well.

Google entirely gave up on Level 3 after they determined it was impossible to ensure their paid test drivers would pay attention: Google/Waymo Stopped Testing Level 3 Self-Driving Tech After Testers Literally Fell Asleep While Using It, Switched To Full Autonomy - CleanTechnica

That's incredible that Tesla managed to do with customers what Google thought was impossible with employees.

 

[uscbucsfan]

Two back to back reboots and stuck on 26.1/11.3.6
Trying to get to ?/11.4.6 or 7

You've been told like 50 times it's not possible to go backwards. You just have to wait.

 

[uscbucsfan]

Google entirely gave up on Level 3 after they determined it was impossible to ensure their paid test drivers would pay attention: Google/Waymo Stopped Testing Level 3 Self-Driving Tech After Testers Literally Fell Asleep While Using It, Switched To Full Autonomy - CleanTechnica

That's incredible that Tesla managed to do with customers what Google thought was impossible with employees.

Tesla isnt going to level 3. It's level 2 with wheel nag and eye/head tracking.

 

[zoomer0056]

That's incredible that Tesla managed to do with customers what Google thought was impossible with employees.

Yeah, a lot of paying (some are vocally angry at being a tester) beta testers instead of paid testers. Once a tester pays in they can't get out. Hotel California: you can check out, but you can never leave.

 

[kabin]

Sure. I even said some of this earlier.

But the fact there's 400,000 cars driving tens of millions of miles a month suggests they don't do it very often otherwise there'd be lots of actual accidents--- versus the narrative from folks who insist it can't go a single mile without "doing something dangerous"




Can you link to them please? I asked a page or two ago for accident cites and nobody had anything other than the one guy who "thought" he heard some blamed on FSD and the one video of an FSD car having its driver ignore a "flooding ahead" type sign and then the car hydroplaned....which is entirely on the driver since all FSD was doing was driving in a straight line.

I'm not saying the # is 0--- Tesla themselves reports an accident rate on FSDb in their 2022 impact report of 0.31 per 1 million miles driven. This is roughly 5 times lower than the 1.53 accidents per 1 million miles in the general US vehicle fleet. I'm saying the number is very low for something the overly cautious among us keep insisting is super dangerous junk.

Thus supporting the case their issues in reporting a lot higher disengagement rates might be due to their overly cautious nature rather than the situations and behavior of the car actually being as dangerous as they perceive.

This puts a little damper on transparency. TSLA claims accident info is confidential information. Apparently NHTSA would need to take legal action to deny TSLA's claims. Seems underhanded as well as an expensive path for the taxpayer.

 

[Dewg]

You've been told like 50 times it's not possible to go backwards. You just have to wait.

You might want to reread that slower.

 

[willow_hiller]

Tesla isnt going to level 3. It's level 2 with wheel nag and eye/head tracking.

The level isn't the point. The point is that Google couldn't find any way to make their drivers pay attention. And now you're saying that Tesla has a magic solution to this problem that makes every customer super attentive.

 

[tmoz]

I'll stop watching this thread until a new release comes out. A few dozen pages will scroll by until then. See you in a few weeks.

 

[uscbucsfan]

The level isn't the point. The point is that Google couldn't find any way to make their drivers pay attention. And now you're saying that Tesla has a magic solution to this problem that makes every customer super attentive.

You don't seem to understand the difference. Google was asking drivers to watch the car drive and it didn't have any forced measures. It was level 3, they weren't driving at all.

Tesla turns things off if you stop paying attention for a short moment.

 

[uscbucsfan]

You might want to reread that slower.

He's on 2023.20.1 and wants to get on 11.4.X, which is 2023.7.X reboot won't do anything.

 

[aronth5]

11.4.6 is still the best version by far for me. Today I only drove 13 miles and only had two disengagements, both because I wanted to take a different route than the nav wanted.

The best new feature is the "keep up with the pack". When entering a construction zone out didn't show to 25 MPH it stayed at 37 MPH matching the traffic. I always had to adjust the speed before so as to not cause a backup.

Have you tried 11.4.6 merging onto a highway where the highway traffic is going slowly? (5mph-30mph). While highway merging has improved this scenario is a disaster for FSD in my experience. I've had many cases where FSD is going 2-3 times the speed of the highway traffic and then runs out of the merge lane room. FSD just doesn't understand it needs to slow down to merge in rush hour traffic.

 

[willow_hiller]

You don't seem to understand the difference. Google was asking drivers to watch the car drive and it didn't have any forced measures. It was level 3, they weren't driving at all.

Tesla turns things off if you stop paying attention for a short moment.

It's possible to be looking forward and holding the steering wheel and still be inattentive. And people will always find workarounds.

If it was as easy as monitoring the driver, I'm sure Google could have done that back in 2017. But they concluded it was not feasible to keep their drivers attentive.