FSD v12.x (end to end AI)

[mgs333]

big wave according to teslafi

I'm seeing no pending versions except for a few for 8.7.

View attachment 1033611

Is this right, did they prune all other FW pending.

Yeah I'm seeing the same and wondering if it's some error or if they pulled everything. That 94 for 8.7 is now down to 6

 

[diplomat33]

I would say the main safety critical issue I see with V12 in my daily drives, is the decision to make those unprotected turns with oncoming traffic. I've been hyper vigilant so there have been no accidents thank goodness. Sometimes, V12 has been really smooth and smart but other times it is very hesitant, kind of jerks the wheel like it is thinking of making the turn but then stops. I am paying attention to oncoming cars and thinking "either go now or wait but make up your mind". I guess I just don't have super confidence that V12 is accurately assessing when to turn in front of oncoming cars to clear the turn safely. So I feel like that scenario is a bit iffy.

 

[syncword]

big wave according to teslafi

Seems so. Been on 2024.2.7 since Feb 21 and a manual check yesterday had no update. This morning they pushed .3.6. So far my car has received updates only after I see a thousand or two before me getting theirs. I wonder if .8.7 also has FSD 12.3.2.1.

Edit: like others, got the one month trial too. While the email says supervised, in the car it's FSD beta as usual. Perhaps unusually, also being a user and small-time developer of Comma's openpilot on my Kia Niro (satisfies my urge to more directly participate in ADAS progress)... been looking forward to trying FSD 12. Comparing Tesla's and Comma's E2E driving models, Tesla's is much larger. The Comma C3X I have is built on basically a cell phone SoC that includes a small GPU module.

 

[AndrewZ]

Not sure what you're asking. Without debating robotaxi chances, it's logical that there would be a minimum capable configuration, yet higher performing configurations to follow.

Does a robotaxi need to avoid every pothole? If not, do you think it would be better if HW4 end-to-end could avoid more of those potholes? There's potential for the newer hardware (cameras and compute) to see more with higher resolution and respond faster to be better such as increased comfort.

Accepting liability at all is already an extremely high bar, Elon has thrown numbers out there in the March of 9 spiels like needing to be 1,000% or even 10,000% safer than the average human. When you get there, I don’t know how much further there is to go?

Guess I don’t picture Tesla taking liability until it’s borderline perfect, and then I don’t know how you improve upon perfection. Either way it wouldn’t matter for the end user because the liability is binary, it’s on Tesla to replace any damaged wheels/tires etc.

Heck the car would just drive itself to get the wheel repaired while you’re at work or something and then drive itself back

 

[rjack1282]

One of our MYs just received the FSD trial this morning. 2024.3.6, FSD Beta 12.3.2.1. Our second MY (which I previously used with FSD V10) has not received the trial.

 

[mgs333]

Pending versions now climbing for 8.7, other versions still 0 but still seeing new installs of 3.6.

Teslascope was still unchanged so might be a Teslafi issue.

Edit: now pending is going back down from 196 to ~76

 

[Supcom]

Much more important would be for Ashok and Elon tell the FSD free trial testers what doesn't work like school busses, school zones, emergency vehicles, pot holes etc. Those are real safety issues new FSD users need to know about.

You mean, the way that they've communicated such things to us over the entire course of FSDb development?

 

[TK211X]

Pending versions now climbing for 8.7, other versions still 0 but still seeing new installs of 3.6.

Teslascope was still unchanged so might be a Teslafi issue.

Edit: now pending is going back down from 196 to ~76

Might be real. Maybe the installs are being reported from previously downloaded FW.

 

[RowdyMY]

I'm seeing no pending versions except for a few for 8.7.

Yeah I'm seeing the same and wondering if it's some error or if they pulled everything. That 94 for 8.7 is now down to 6

TeslaFi went down ~0730 EDT for about 30 minutes this morning. Probably still recovering.

 

[Supcom]

I have no idea if there will be many more accidents with millions of FSDb cars out there, being driven by millions of people who have not become accustomed to the dangers. But one would think that there will be at least an increase in fines, suspension damage, wheel damage and possibly parking lot fender benders.

There are already over 400,000 users of FSDb who were not accustomed to the dangers when they began using it. And, FSDb was much worse when they did start using it then it is for new users today.

Every time FSDb access has been expanded, there have been predictions of mass mayhem. But, somehow, the newbies seem to figure out pretty quickly to be cautious and there is no carnage. Maybe it's that self-preservation instinct?

 

[SidetrackedSue]

Pending versions now climbing for 8.7, other versions still 0 but still seeing new installs of 3.6.

Teslascope was still unchanged so might be a Teslafi issue.

For my current version there were 34 of us this morning on Teslafi and one pending download. That pending is gone now, but the 34 is unchanged.

Teslafi still shows my car waiting to download 2023.44.30.14. So nothing has changed as far as my car is concerned (if tesla had pulled the update it would show as 'installed' and the new offering would be listed as "new" and "waiting for wifi to download").

 

[JB47394]

Here's an interesting clip from AI DRIVR's live stream yesterday. The navigation path showed going past the destination, taking a left, making a U-turn, taking a right, then arriving at the destination. FSD ignored all that and took a left to reach the destination.

It looks like FSD doesn't go by the V11 visualization, but what's going on with the navigation path? Is that not being used as well? I seem to recall other cases where FSD would go off the planned route (in small ways).

 

[lzolman]

Which brings up an interesting conundrum. In the future when people become more aware and understanding of self driving cars they will be more likely to cut them off, walk in front of them, block them in, etc, since self driving cars have no passion and will willing yield. We are stating to see these type of behavior towards Waymo.

It makes sense to me that once cars are capable of L3+, such as the (slow) Mercedes, they'll need that visual indicator the car's driving. Then we will absolutely have some interesting new dynamics. I'm kinda looking forwad to the "fun" of it...

 

[aronth5]

Did not say it wasn’t!!!

Of course you did. My point was very specific- part of the delay is the creep required to move further forward to make a safe decision on low visibility intersections.

You said "Too slow. Slowness not related to bad visibility."

 

[SidetrackedSue]

There are already over 400,000 users of FSDb who were not accustomed to the dangers when they began using it. And, FSDb was much worse when they did start using it then it is for new users today.

The difference between the existing FSDb users is that, in general, they are interested in the product. A good number of those bought it and have been riding the development train, others have paid for a subscription.

You may be correct, I may be correct. Either way it won't show up in the safety stats that tesla totes if this version of FSDb is mostly the fender benders and damage due to hitting potholes (or ladders) on the road that I predict. And of course, fines don't show up at all, According to the small print on the safety stats page, " In practice, this [crash] correlates to nearly any crash at about 12 mph (20 kph) or above, depending on the crash forces generated." I'll assume that most drivers will instinctively react to circumstances leading to that sort of crash and deactivate and avoid collisions.

 

[lzolman]

It would be really helpful if Ashok or Elon posted a list of general issues they’re aware of and addressing.

I would feel much better if Tesla acknowledged that v12 drives too slow, creeps too slow, hesitates at some intersections, etc. so that I had more confidence that they were working on a fix.

How could they possibly not know? I think they tend to play things close to the vest for competetive reasons and also not to provide too much fodder for the FUDsters. We all want to know this stuff, but "normal" people probably couldn't care less anyway.

 

[TK211X]

I want..... summon.....

 

[kabin]

I would say the main safety critical issue I see with V12 in my daily drives, is the decision to make those unprotected turns with oncoming traffic. I've been hyper vigilant so there have been no accidents thank goodness. Sometimes, V12 has been really smooth and smart but other times it is very hesitant, kind of jerks the wheel like it is thinking of making the turn but then stops. I am paying attention to oncoming cars and thinking "either go now or wait but make up your mind". I guess I just don't have super confidence that V12 is accurately assessing when to turn in front of oncoming cars to clear the turn safely. So I feel like that scenario is a bit iffy.

Agreed. Misjudging oncoming traffic continues to be FSD's Achilles' heel. We all probably noticed Chuck's tweet of another ADAS test vehicle in his neighborhood this week.

Hopefully these one month FSD(S) free trials happen without a hitch as the majority of drivers on the road didn't sign up to be crash test dummies.

 

[Pdubs]

Morning drive with 12.3.2.1 went very well. It's amazing how much better it is than 12.3 in the DC area. Actually worth the monthly subscription.

I do think it accelerates a bit too quick for my liking on both chill and normal. Only real issue I have (I also don't want to eat through my tires).

Auto speed was perfect, although I do wonder if it's smart enough to know to slow down a speed camera.

Really looking forward to seeing how this technology gets better this year. Very promising right now.

I don't know it it's just me, but I haven't noticed any discernible difference between any of the FSD modes.

 

[AlanSubie4Life]

Of course you did

You said "Too slow. Slowness not related to bad visibility."

Correlation is not causation.

It’s slow on many intersections, irrespective of visibility, and sometimes creeps for a long time (the 2-4 seconds you mention, and more), when visibility is perfectly fine.

It can take over 10 seconds to turn after coming to a stop (it takes about 5-10 seconds to come to a stop), for it to commit, on completely unobstructed intersections.

Original sequence of discussion:

I'd say that it was generally a little too timid when creeping out at intersections with poor visibility.

Get used to this on intersections with poor visibility.

just expect a few second delay on low visibility intersections. People behind you will continue to be a little irritated and there is nothing you can do.

Again, it’s like this on many intersections, not just low visibility. And there is something you can do when people are behind!!! Press the accelerator and make it GO, quickly. Even on low visibility intersections you can do this - some of it before the decision point, and some of it to push through the decision point. Another point I disagreed with.

When you can see cross traffic well enough to make a decision you'll note that FSD creeps another 2 feet. That slow creep is a 2-4 seconds of added time.

Why wouldn’t it just go to that point initially? (Not counting the NHTSA stop which already happened 10 seconds prior.) What information does it gather by stopping before it can see?