FSD v12.x (end to end AI)

[Cherry Wine]

24MYLR got 2024.3.6 pushed last night from 2024.2.7. Downloading now. 20MXLR still on 2024.2.7.

Looking forward to seeing if 12.3.x lives up to the hype - I found v11 pretty disappointing and never used it.

 

[JB47394]

Driving on FSD Beta, on the roads that FSD Beta is typically driven on, with the disengagements that drivers have been making, is safer than driving a Tesla without active safety features, on the roads those Teslas are typically driven, with the driver attributes of those without them.

I hesitate to jump in here, but I don't believe the statistics demonstrate that.

The bars show

1. The accident rate per mile for Teslas driven on highways that use Autopilot.
2. The accident rate per mile for Teslas driven on secondary roads that don't use Autopilot.

I claim that because V11 is going to be used for highway miles far more than secondary. It's why V12 came into being. So Tesla owners will use Autopilot on highways and drive manually on secondary roads.

This is just a bias statement because there are clearly going to be people with Teslas who manually drive highway miles. The point is that the bias is there, and I would assume that it's quite strong for the long bars in the safety report. They are almost all highway miles.

What's needed is a breakdown of highway and secondary road miles for both owners and Autopilot because not all miles are created equal. Perhaps the breakdown should be by speed limit ranges. Secondary roads can have high speed limits, suggesting that safety officials deem eating up miles on those roads is safer.

 

[YW555]

First time using FSD with the one-month trial that's being given out on 2024.3.6. I have some observations after a 30 mile morning commute:

• No taillight visualization? I watched some of the FSD videos where they showed FSD visualization showing brake lights and turn signals on other vehicles. I guess it's disabled with the latest FSD?
• It's hanging out behind a semi-truck on a 5-lane highway, I don't need another cracked windshield. Manual turn signal doesn't influence lane change, kind of disappointing.
• Almost missed my exit because it's too conservative on the merging, had to take over. I might try 'Aggressive' behavior setting and see what happens.
• No new Autopark? I guess that's still slowly rolling out.

I have a 2023 Model Y with HW3.

 

[zoomer0056]

FSD (Supervised)
Free one month trials are going out. Expect an email and a note in the car to switch it on if your car is chosen.

I may give it a try since it's free. I've already received my training of many months from this thread.

Post in thread 'When will trial start rolling out?' When will trial start rolling out?

 

[MP3Mike]

Manual turn signal doesn't influence lane change, kind of disappointing.

Sure it does. But not at the three blink level, you have to push it to the full activate level.

 

[GWord]

• No taillight visualization? I watched some of the FSD videos where they showed FSD visualization showing brake lights and turn signals on other vehicles. I guess it's disabled with the latest FSD?

This was with the previous Version 11 that had more “visibility” into what the vision system was seeing. I think the general consensus here is that the current end-to-end approach doesn’t easily allow for queries into specifics of what the model is seeing. It seems like there is a V11 object recognition “lite” running in parallel to produce some sort of visualization, but it’s basic compared to previous versions (probably to minimize compute load).

 

[lzolman]

I think Elon jumped the gun a bit with the free FSD trial. He should have waited until the speed issues were mostly fixed first, which I don’t think is too far off. The conversion rate would probably be significantly higher.

Having said that, there is a portion of the population who disliked FSD due to phantom braking, unnecessary braking, jerky steering, and indecisiveness which is vastly improved in v12 so there is that…

This may be a stretch (or just track my wish for demonstrating continuous improvement), but to me the timing could work out well if the speed issue--and hopefully others--are effectively addressed in an upcoming update. It'll be the start of that march of nines, something I absolutely need to see to be able to feel truly positive about FSD.

 

[sleepydoc]

If I had a nickel for every time someone suggested he got his car checked....

Kinda says something, don’t you think?

Edit: I’ve noted this before but Alan really seems to obsess over small nuances of driving style and with that seems to have difficulty appreciating that there is a range of normal driving and that a giving style, while different from his own preference, may still be completely acceptable.

Now I will readily admit that there seems to be a fair amount of variability in FSD behavior between cars, locations, etc, so it is possible that his car actually is behaving abnormally but from the totality of his posts I tend to think it’s more perceptual on his part.

 

[sleepydoc]

Looks like a new wave started this evening

Wasn’t 3.2.1 the version installed on new cars following Our Dear Leader’s dictum?

 

[sleepydoc]

Poor @TK211X because of their unbounded optimism on get the lates updates has been moved behind even me on getting them. So that may be awhile.

Well, someone’s got to keep @j0shm1lls company on the V11 thread!

 

[PianoAl]

Also, look at the specific defect: consistently going lower than the speed limit. This is *not* how humans drive. So if the whole premise is "based on millions of video clips of humans driving" ... how do you figure? Also, how is something like this "fixed" with the current video/AI model if you're not writing explicit code to correct behavior? What, they took video clips of people driving faster? Doesn't add up.

I literally glanced at the avatar to make sure this wasn't one of my own posts. I've concluded that I just don't understand the "based on millions of video clips of humans driving" thing.

I'm going say something that I don't believe, and no one else in this thread has said (unless I missed it). It's something that needs to be brought up, so don't flame me.

What if this whole "based on millions of video clips of humans driving" thing doesn't work and Tesla has to to back to explicit coding?

 

[primedive]

0 for 1 on unprotected lefts on 12.3.1, @Daniel in SD. The hangs by a mere thread.

We'll see if he goes 9/9 on the next 9! I'm sure it was just the rain; those cars were super hard to see with their headlights and all.

No normal human could perform this turn in these conditions. This is a highly skilled human trained to land on aircraft carriers in a nuclear-capable (depth charges - seems like overkill?) plane resembling a flying Hoover vacuum (S-3 Viking).

Also failed here (doesn't count):

Florida is nuts. That road could be one lane on each side with roundabouts and still handle the traffic just fine, and also not be a death trap. Last time I was in Florida I saw a crash almost every five minutes.
If FSD can reliably handle these intersections then I will be truly impressed.

 

[Todd Burch]

First time using FSD with the one-month trial that's being given out on 2024.3.6. I have some observations after a 30 mile morning commute:

• No taillight visualization? I watched some of the FSD videos where they showed FSD visualization showing brake lights and turn signals on other vehicles. I guess it's disabled with the latest FSD?

Interesting, works for me.

• It's hanging out behind a semi-truck on a 5-lane highway, I don't need another cracked windshield. Manual turn signal doesn't influence lane change, kind of disappointing.

Interesting, manual lane change works for me while FSD is engaged. But V12 is definitely not as aggressive as v11 since the speed issues make it difficult for the system to know how fast you want to be going. Assume they’re working on that.

• Almost missed my exit because it's too conservative on the merging, had to take over. I might try 'Aggressive' behavior setting and see what happens.

Highway driving is still using v11 code. But both v11 and v12 often don’t get in the proper lane for a turn or exit until the last half mile which can increase the pucker factor. Admittedly v12 on city streets has not missed a turn for me yet, though I wish it would get in the proper lane sooner.

 

[GWord]

What if this whole "based on millions of video clips of humans driving" thing doesn't work and Tesla has to to back to explicit coding?

I wonder if it could it also be just as likely that they find some new, as of yet unknown, solution that is neither? Some hybrid approach maybe?

 

[Cherry Wine]

24MYLR got 2024.3.6 pushed last night from 2024.2.7. Downloading now. 20MXLR still on 2024.2.7.

Looking forward to seeing if 12.3.x lives up to the hype - I found v11 pretty disappointing and never used it.

Initial reaction on 12.3.2.1: it lives up to the hype.

I couldn’t go half a mile before disengaging on previous versions over 2.5 years. I just had 30+ minutes of nearly flawless drives through places that always stumped FSD in the past.

I had two disengagements - both when the car wanted to treat a wide bike lane as a vehicle lane - and one accelerator intervention when the car treated a three-way stop as if cross traffic didn’t have to stop. There were a couple of spots where I could have intervened if I were being really picky in how the car was driving.

Auto speed setting was fine.

If I were grading the car I’d give it a B, maybe a B-.

I’m impressed, and I’ve been very skeptical of FSD in the past.

 

[Todd Burch]

What if this whole "based on millions of video clips of humans driving" thing doesn't work and Tesla has to to back to explicit coding?

It’s a valid question. After all, nobody has solved this problem yet. It’s an order or two magnitude harder than landing on the moon, and that was hard.

And yes we all know this initial v12 release has some issues still-nobody except Whole Mars really thinks it’s perfect (only teasing, neither does he). But after experiencing v12 and comparing to v11.4.9, do you honestly feel like the coded approach was better?

 

[Todd Burch]

I wonder if it could it also be just as likely that they find some new, as of yet unknown, solution that is neither? Some hybrid approach maybe?

I think the hybrid approach is in use now. Neural net for the control, but with a C++ guardrail layer on top which has overriding control.

 

[kpanda17]

It’s a valid question. After all, nobody has solved this problem yet. It’s an order or two magnitude harder than landing on the moon, and that was hard.

And yes we all know this initial v12 release has some issues still-nobody except Whole Mars really thinks it’s perfect. But after experiencing v12 and comparing to v11.4.9, do you honestly feel like the coded approach was better?

how can it fail?
edge cases are the only risk and they just need to capture edge case same videos and it will learn how to correct those

 

[rlsd]

Initial reaction on 12.3.2.1: it lives up to the hype.

I couldn’t go half a mile before disengaging on previous versions over 2.5 years. I just had 30+ minutes of nearly flawless drives through places that always stumped FSD in the past.

I had two disengagements - both when the car wanted to treat a wide bike lane as a vehicle lane - and one accelerator intervention when the car treated a three-way stop as if cross traffic didn’t have to stop. There were a couple of spots where I could have intervened if I were being really picky in how the car was driving.

Auto speed setting was fine.

If I were grading the car I’d give it a B, maybe a B-.

I’m impressed, and I’ve been very skeptical of FSD in the past.

Thanks God you didn't teach me and my kids.

 

[AlanSubie4Life]

and say they are fudging the numbers

They are not.

This is comparing Tesla to Tesla, so the demographics of Tesla drivers vs the general population is inconsequential.

Nope.

But it absolutely refuses the idea that FSD Beta is unsafe

Not claiming that.

are proven wrong with the data.

Nope.