FSD v12.x (end to end AI)

[sleepydoc]

Speed limits are also made up numbers...
I think 10 seconds is the UNECE regulation.

I think this refers to things like the car being on fire. Which should be detectable by someone in the car but not by the car itself.

By that logic all numbers are made up, except for e, π, c, 6.02*10^23...
Speed limits are numbers that have been decided upon and codified. "10 seconds" appears to have been made up by people here.

The SAE document specifically mentions suspension issues as an example of an issue that should be detectable by fallback person but again, there's a contradiction in that an remote fallback person would not be able to detect any of these.

FYI, the UNECE document states:

"The initiation of the transition demand shall be such that sufficient time is provided for a safe transition to manual driving"

and

"In case the driver is not responding to a transition demand by deactivating the system (either as described in paragraph 6.2.4. or 6.2.5.), a minimum risk manoeuvre shall be started, earliest 10 s after the start of the transition demand"

This was the only reference to 10 seconds that I could find. I haven't read the entire document but the UNECE guidelines appear to be much more prescriptive than the ASE guidance. That could lead to more uniformity but could also be problematic if it's overly prescriptive in ways that aren't appropriate for local driving customs and conditions. It will be interesting to see how it all proceeds; I can easily see the EU adoption the UNECE rules and the U.S. having its own (or even worse, every state having its own.)

Oh, not me. It took me 15 minutes to get out of bed this morning.

Joking aside, I'm highly skeptical we ever see wide rollouts of L3 systems with these short 10-15 second takeover times. Humans are bad drivers already under normal conditions, let alone acquiring situational awareness within a few seconds. Slightly more likely, IMO, is a system with a strong enough safety record that the manufacturer would be comfortable with no driverless operation on highway and the only L3 handbacks are triggered at off-ramps where there is no time pressure.

Like I mentioned above, 'level 3' is broad enough that you could have a 'level 3.1' system and a 'level 3.5' system that is still level 3 but more reliable and capable. Regardless, especially in the beginning I would expect there to be signifiant numbers of circumstances in which L3 systems would need to time out and have the driver take over. The key is how to do this safely. It's very possible to have the system alert and then either pull over or simply slow down and stop with the hazards on if there is no response. The UNECE document specifically references this. If you can't handle these situations then a L3 system will never happen.

 

[sleepydoc]

One addition here about the no safety regressions, they are just double confirming that and then they are good to go. OMG, this is it, the first 9.

And much smoother! Remarkable. I thought there was a distinct group of people here who said it had no smoothness issues.

And no I don’t think it will be smooth and I don’t think it will be particularly good, but I am betting on the first 9 this time.

Though that is extraordinarily dumb (it’s me after all), because past experience has shown that none of the point releases result in any changes to driving behavior (it’s not retrained AFAICT). So really it should be 12.5 which does it. Though maybe they can prompt-engineer it to make it “Faster” “Make it really fast”, etc. since that is all that is needed. 3-second crossings should do it.

No, Alan, it will never be smooth enough for you. You've made it abundantly clear that you like to set impossible standards that are five 9's outside of the norm for everyone else and then complain that the car doesn't achieve them.

It's ok - the rest of us will be happy. And we can enjoy your unhappiness while we enjoy our beers! (I'm going to enjoy a nice, smooth porter)

 

[AlanSubie4Life]

No, Alan, it will never be smooth enough for you

Apparently wasn’t smooth enough for the Autopilot team either. Vindication. We’ll see whether it is smooth enough for them now. I suspect not - they have insane standards.

 

[sleepydoc]

Apparently wasn’t smooth enough for the Autopilot team either. Vindication. We’ll see whether it is smooth enough for them now. I suspect not - they have insane standards.

not really - the question is whether the released version is smooth enough for you!
odd you say they have insane standards yet you haven't been satisfied with the past releases. What does that say about your standards?

Like everyone else I'm still tooling around on 12.3.6 and my wife has zero problems riding in the car. (This is the first version she has not complained about.)

 

[Nostromo]

The county just repaved the suburban four lane divided highway by my house. There are sections with no markings on the asphalt. FSD didn't even blink. It still visualized the lanes, including turn lanes, and kept its lane like always.

I doubt it visualized the lanes at all, most likely it relied on map data. I say this because my M3 consistently fails to make a turn in my neighborhood where the road is narrow (maybe 2 1/2 lane widths) and the curb is formed asphalt, not painted.

 

[JB47394]

I doubt it visualized the lanes at all, most likely it relied on map data.

"Visualized" in the sense that they appeared in the visualization, not "perceived".

 

[kabin]

One addition here about the no safety regressions, they are just double confirming that and then they are good to go. OMG, this is it, the first 9.

And much smoother! Remarkable. I thought there was a distinct group of people here who said it had no smoothness issues.

View attachment 1059848

And no I don’t think it will be smooth and I don’t think it will be particularly good, but I am betting on the first 9 this time.

Though that is extraordinarily dumb (it’s me after all), because past experience has shown that none of the point releases result in any changes to driving behavior (it’s not retrained AFAICT). So really it should be 12.5 which does it. Though maybe they can prompt-engineer it to make it “Faster” “Make it really fast”, etc. since that is all that is needed. 3-second crossings should do it.

The usual Nancy Pelosi response - we have no idea how safe it will be until after the build.

And I agree, 12.4.2 will likely be no smoother or safer. The improvements are like an unrefined Parkinsonian robot on simulated carbidopa/levodopa data.

The connection between less safe and smooth is bizarre but makes sense from the FSD short term fix/crutch perspective. An attentive safe human driver is smooth and predictable. I bet the team spends 50% of their time dicking with simulation data for training.

 

[kabin]

not really - the question is whether the released version is smooth enough for you!
odd you say they have insane standards yet you haven't been satisfied with the past releases. What does that say about your standards?

Like everyone else I'm still tooling around on 12.3.6 and my wife has zero problems riding in the car. (This is the first version she has not complained about.)

My family doesn't want to use FSD when driving in the car. They say it's rough, indecisive, and in general embarrassing.

 

[Dewg]

I thought there was a distinct group of people here who said it had no smoothness issues.

And there is a distinct group of people who thought it needed additional smoothing. You're talking about a subjective issue. Is the current smoothness satisfactory for many people? Yes. Can it be smoother? Yes. Smoothness is about an acceleration and deceleration curve. What is the correct curve for ultimate smoothness? There is no such thing, because you will never satisfy everyone. Think of it like room temperature. You will never find the "correct" room temperature for everyone. Just try to please as many as you can and hope the rest find it "satisfactory".

 

[Daniel in SD]

This was the only reference to 10 seconds that I could find.

Yes, UNECE decided that 10 seconds is the minimum number. It seems like enough time to regain situational awareness. If it's not then it will change, just like speed limits.

 

[AlanSubie4Life]

And there is a distinct group of people who thought it needed additional smoothing. You're talking about a subjective issue. Is the current smoothness satisfactory for many people? Yes. Can it be smoother? Yes. Smoothness is about an acceleration and deceleration curve. What is the correct curve for ultimate smoothness? There is no such thing, because you will never satisfy everyone. Think of it like room temperature. You will never find the "correct" room temperature for everyone. Just try to please as many as you can and hope the rest find it "satisfactory".

That is not how things work.

Sometimes people just need to take the L.

 

[aronth5]

Apparently wasn’t smooth enough for the Autopilot team either. Vindication. We’ll see whether it is smooth enough for them now. I suspect not - they have insane standards.

No not vindication. You have no idea what criteria they used and how they tested. Doesn't mean you could not be correct but stating as fact does not make it so.

 

[gsmith123]

You're talking about a subjective issue. Is the current smoothness satisfactory for many people? Yes. Can it be smoother? Yes. Smoothness is about an acceleration and deceleration curve. What is the correct curve for ultimate smoothness? There is no such thing, because you will never satisfy everyone.

In some cases people have preferences (acceleration speed for a stop sign), but there’s a lot of cases today where the car is simply jerk in ways that are undebateably bad.

Offhand examples:
- jerky steering through my local roundabouts
- yo-yoing follow distance
- uneven deceleration slowing to red lights (watch the regen bar)

 

[Dewg]

In some cases people have preferences (acceleration speed for a stop sign), but there’s a lot of cases today where the car is simply jerk in ways that are undebateably bad.

Offhand examples:
- jerky steering through my local roundabouts
- yo-yoing follow distance
- uneven deceleration slowing to red lights (watch the regen bar)

I'm not arguing that it can't improve. And I don't experience jerky steering. And I don't find the deceleration curve annoying. But I also don't mind the office at 72 degrees, while some of the staff are freezing and have space heaters.

I like large gaps between me and the lead car, while others lose their sh$t over it.

You find the perfect deceleration curve for you, and I guarantee someone will say it started braking too soon. And someone else will say it started braking too late.

 

[AlanSubie4Life]

In some cases people have preferences (acceleration speed for a stop sign), but there’s a lot of cases today where the car is simply jerk in ways that are undebateably bad.

Offhand examples:
- jerky steering through my local roundabouts
- yo-yoing follow distance
- uneven deceleration slowing to red lights (watch the regen bar)

Exactly. I never said there weren’t people who found it smooth. (This is subjective.)

I said it wasn’t smooth. (This is objective, quantifiable, decades of support, well established limits for human comfort.)

That’s a key distinction.

Anyway they are making it smoother because it wasn’t smooth enough, that’s the key. They have enough priorities that they wouldn’t waste time making something smoother if it were smooth enough.

 

[hey8you]

I think there may be some sort of automated or rapid map adjustment mechanisms. On V12 I've had a couple of incorrect lane choices that have occurred exactly once. One mistake, one disengagement and voice report, and then it's never made the mistake again.

I wish V12 would do that for me. There is a spot where the right lane is about 1 block long(not a turn lane). For some reason V12 loves trying to hit that lane even though it ends after a block.

 

[Dewg]

Exactly. I never said there weren’t people who found it smooth. (This is subjective.)

I said it wasn’t smooth. (This is objective, quantifiable, decades of support, well established limits for human comfort.)

That’s a key distinction.

Anyway they are making it smoother because it wasn’t smooth enough, that’s the key. They have enough priorities that they wouldn’t waste time making something smoother if it were smooth enough.

Ugh, I'm done arguing with you over this. You literally used the word comfort, which is a subjective term, in your description of objective.

I hope Tesla gets the smoothness you're looking for some day. I really do.

 

[AlanSubie4Life]

I hope Tesla gets the smoothness you're looking for some day. I really do.

They will. No question about it.

 

[FSDtester#1]

This info comes from the Tesla Institute of Technology

I didn't know that you were interested in the Tesla Institute of Technology (TIT).
Interesting to hear that you pay close attention to it and have a grasp and feel of what is going on with them.
Your eyeballs must stay glued to any updates.

 

[Yelobird]

Ugh, I'm done arguing with you over this. You literally used the word comfort, which is a subjective term, in your description of objective.

I hope Tesla gets the smoothness you're looking for some day. I really do.

Applaud the effort but you didn’t Really think you would sway or win a debate with the Subie? The Alan Subie System (ASS) is designed to wear down any alternate view not his own. You tried, as many have before you so there’s that!