Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register
This site may earn commission on affiliate links.
New users just don't have the experience of knowing when to intervene and trusting an all new v12 system
Even "expert" owners don't know what an all new v12 system should be good or bad at because Tesla doesn't tell us. We've had to figure out from subtle clues that 12.x so far hasn't been trained for freeways, but fortunately there's a pretty good fallback of 11.x behavior. Similarly, only from watching or experiencing poor parking stall performance of 12.x is it clear this wasn't intentionally trained on this yet either, but this doesn't have a good fallback until actually smart summon.

But Tesla probably also doesn't fully know the extent of 12.x current capabilities and limitations especially for new behaviors such as going around puddles. How often does it correctly or incorrectly drive through water at what depths and speed; and what thresholds become unsafe or "just" uncomfortable? Apply this to other new behaviors like U-turns or even just pull aside parallel parking. Some behaviors might already be good enough with general end-to-end training, but others will need dedicated training to be safer for release to everyone.
 
  • Like
Reactions: OxBrew and JB47394
Wonder if the "stuttering" speed on straight roads, with no lead car is related to the training videos it learns from. If shown several videos of cars going straight by themselves they will be driving at different speeds and they will all be correct. Maybe the NN see this and thinks that to be driving correct it must drive at all the speeds it has learned.
 
Wonder if the "stuttering" speed on straight roads, with no lead car is related to the training videos it learns from. If shown several videos of cars going straight by themselves they will be driving at different speeds and they will all be correct. Maybe the NN see this and thinks that to be driving correct it must drive at all the speeds it has learned.
I think it's more likely the case that the guard rail system doesn't feel confident enough when there is no lead car and the guard rails are activated leading to the conservative and bumpy behaviour.

The car seems to do fine with a lead car so it understands to keep a smooth "non-low" speed is preferred in other situations.
 
It's too bad Tesla didn't release V12 to all the beta testers, they lost out on all of the data which would have been collected to feed back into the system. Beta testers are capable of managing V11an its idiosyncrasies, why not V12.
I personally don't want to see it until it's better than V11 across the board. If it's not, they've got plenty of issues to work on from just the testers they've already got...
 
But anyway I would argue that nearly every time a competent human disengages due to something outside the comfort zone, that was a mistake by the car - it was deviating from prudent behavior. I don't think these type of systems should routinely engage in behaviors that would result in reduced margins, since they make it too difficult for the human to recover.


The words 'nearly' and 'competent' are doing a lot of a heavy lifting in that sentence in my opinion. There I times I drop out of FSDbollocks because I don't want to have to put the attention into anticipating every mistake it is going to make. It is easier to just drive the car myself. Not every time is my comfort zone reasonable according to the drivers here on TMC. For instance, in heavy rush hour traffic, I leave 2+ car lengths in front of me and never take offense if someone fills the lane in front of me, I just drop back to recreate that cushion. There is a school of thought here that makes me incompetent.

Comfort zones are a tricky thing. So I disengage a lot but most of the time it is because FSDbollocks has not demonstrated that I should trust it. So the 'nearly' does apply. Most importantly, I'm never testing it in tricky circumstances or on city streets. I don't have the energy to deal with its inability to drive the way I drive, slowing down in anticipation of situations (like red lights) that the car has not seen yet (or at least are not on the screen as being in its line of sight.)

Where I do use it is the route to my son's. That is 5+ minutes of city streets to the city freeway (no FSDb used at all because I don't use it on that sort of driving.) Then freeway to 4 lane divided highway interchange which in good weather FSDb can handle that but taking it at a higher speed (25% higher than the suggested speed) than I'm comfortable with so I scroll it down and back up. If the driving conditions are suspect (potential for black ice, wet, snowy), I disengage because my version of FSDb doesn't respond to driving conditions (except sunlight on the cameras.)

There's a slowdown at a particular point (have my foot on the go-pedal to nudge since this is not a real or imaginary speed change), then slowing down as the highway enters the outskirts of a town and becomes a city street, with two through lanes and assorted turning lanes, including an alternating shared left-turn lane for a section. At the opposite end of the town (~8 blocks and one traffic light) I need to pull into a dedicated left turn lane and turn onto a 2 lane highway (with passing lanes every 10km or so to get around slow moving vehicles, slower through a couple of villages, and a couple of traffic lights) and proceed 28km to my son's town where I disengage because of city streets.

I disengage before at the end of the 4 lane highway because it won't slow down smoothly from 100kph to 50kph as the highway ends and it keeps me in the right lane (I'm set to Chill) which is actually the worse lane for speed (several dozen business driveways) and difficulty to get into the required left turn lane due to traffic volume of cars in the better lane avoiding all the right-hand turning folks into the businesses. If I ask it to move to the left lane while on FSDb, it will but it moves back to the right lane as fast as possible, including within a half block of the beginning of the left turn lane I need. I've never made it through the down without disengaging. I've reengaged while in the left turn lane waiting at the red light and it has pulled forward when it got its protected turn light but since its path was into the (occupied) lane of the stopped cross traffic, the only time I tried it (an early version of V11), it failed, and I'm not in the mood to give it another chance until it can make it into town in the first place without me needing to disengage.

It has got better about moving into the passing lanes on the 2 lane highway but still dives into turning lanes (not consistently in the same spots, though.) Because I don't speed, I need to allow others to pass so I appreciate it being reliable on moving out of the way of others but it is unreliable for moving back when the lane ends; we still get occasional (but not predictable) drives along the shoulder when the lane ends and the car is 'surprised' (that we experiment with only when there are no other cars in sight.) Sometimes it will move into the through lane if we signal, sometimes it won't. We never signal asking for the lane change if there are other vehicles in the lane so we have no idea why it works sometimes and not others.

So there's no trust whatsoever, even on relatively simple parts of a mostly highway drive. Each time we update, we let it try (and for a brief time in the fall the predictable slowdown on the highway went away but it is back again with our current version.)

It is frustrating for me to see posts about all the improvements others see but while highway is much smoother, we are still getting occasional dives into turning lanes off 2 lane highways and no improvement on the stretch through the edge of town.

So 'nearly' every time I switch out it is because I'm anticipating or experiencing bad behaviour. Will the car crash if I don't disengage? Likely not but I might get pulled over for DUI-like behaviour (part of the drive is past a provincial police station so this behaviour being seen by an officer is not as unlikely as one might think).

As I only use FSDb for mostly highway drives, I'm not in any rush for V12 since the majority of my driving would still be under V11. But it has such a long way to go to get over 'comfort-level' disengagements, I doubt I'll see a big change in 2024 at all. I need my car to be consistent and safe all the time, not just mastering a 4 way stop or a particular UPL in the US while not managing to deal with the different road markings and speed change signage we have in my province.
 
Anyone notice that the car has ultrasonic sensors!
Impossible. Noone with USS has ever had an issue parking. TMC has established that they are literally the finest technology man kind has invented, several times over.

Makes this guy seem like even more of an idiot though, so not only did he let v12 handle a situation nobody at Tesla has claimed it can handle, then sat there and supervised it messing it up, he also did it in the face of the parking sensors telling him it was about to happen.
 
It just ran into a parked car!
In term of longer distances Andrej is only showing a single video clip, the distance is only 30m and there are no statistics about how well it actually works. It's also a very simple scenario where the cars are moving in the same direction. Watch Chuck's ULT videos which have much longer distances. He still hasn't managed to get nine successful turns in ten attempts (one "9" in the "march of nines"). I win a beer if V12 does it so I'm optimistic.

I don't think you can tell distance estimation precision from collisions or lack thereof. After all, Waymo has lidar sensors that have distance accuracy down to millimeters and it still hit the same truck twice.

For example, if you took an average human and asked them to estimate the distance from their bumper to the vehicle ahead, I wouldn't be sure they would necessarily be able to give a number that is more accurate than the functions used by FSD Beta (at least for V11, not sure what the equivalent would be for E2E). Yet that doesn't impede them from driving safely. What humans do a better job at is intuitively figure out when to give a larger margin as appropriate. For example, most humans would not make that turn as confidently.

It's unsure what the car was attempting. Looking at other videos of V12, even for cars with USS, it does not transition to parking automatically, but rather would just stop in the travel lane.
 
  • Like
Reactions: rlsd
Anyone notice that the car has ultrasonic sensors!
I don't disagree at all that the car needs better camera coverage of the front bumper area.

Having said that, I'd like to ask why the general assumption is that v12 was actually engaged when this incident occurred. Is it because the person reporting it is (apparently) a friend of a friend of the community?

This is a genuine question, though I completely understand that it might come across as some fanboy attempt to exonerate Tesla. I can only say that's not my purpose in asking; I really want to understand.

The whole thing seems odd to me. No one else has reported that v12 has hit another car (nor any large object), moving or stationary. The comment about expecting it (or hoping for it) to make a three-point turn seems very odd, and certainly would seem to go along with a bit of extra caution on the part of an experienced FSD-er. We certainly cannot blame the driver for not running with a YouTuber cabin camera setup, but this simply leaves us with no validation one way or the other.

If the driver is wrong or about what happened, that doesn't have to be chalked up to dishonesty - it's very easy to be confused in the moment. I know that I tend to drive FSD with my foot hovering over the accelerator more than the brake, because it's much more likely to require a little push then an emergency stop. Then in this "three-point turn" attempt, a sudden brake/disengage intention could result in a nudge on the wrong pedal - we know this happens all the time in Level Zero driving and parking lot accidents.

I suppose Tesla knows, because they have access to the multi-camera video plus ancillary status telemetry.

I'll admit that I could be missing something, and I don't know much about the situation beyond a few recent posts here, so I'd be grateful for any clarification.

EDIT: this topic is being discuss simultaneously in the v11 and the v12 threads. The thrust of my comment may have more to do with the other conversation but I think it is a v12 topic, so I'll leave it here.
 
Last edited:
  • Like
Reactions: JB47394
Having said that, I'd like to ask why the general assumption is that v12 was actually engaged when this incident occurred. Is it because the person reporting it is (apparently) a friend of a friend of the community?
Just seems easiest to assume it was. Not really any downside and seems very reasonable.

Obviously we could talk back and forth about it but we’ll never know. Just have to make the best assumption, whatever that is.
 
There have been other accidents on FSD Beta if you use Tesla's numbers, I was more surprised when Tesla put out those numbers and we had never seen any of the incidents on social media or anything. But of course the majority of testers are testing and not videotaping it.

Kinda amusing that WholeMarsBlog and others are going on the offensive because the guy late last year was receiving warnings when using FSD Beta and I think was suspended at one point
 
The words 'nearly' and 'competent' are doing a lot of a heavy lifting in that sentence in my opinion. There I times I drop out of FSDbollocks because I don't want to have to put the attention into anticipating every mistake it is going to make. It is easier to just drive the car myself. Not every time is my comfort zone reasonable according to the drivers here on TMC. For instance, in heavy rush hour traffic, I leave 2+ car lengths in front of me and never take offense if someone fills the lane in front of me, I just drop back to recreate that cushion. There is a school of thought here that makes me incompetent.

Comfort zones are a tricky thing. So I disengage a lot but most of the time it is because FSDbollocks has not demonstrated that I should trust it. So the 'nearly' does apply. Most importantly, I'm never testing it in tricky circumstances or on city streets. I don't have the energy to deal with its inability to drive the way I drive, slowing down in anticipation of situations (like red lights) that the car has not seen yet (or at least are not on the screen as being in its line of sight.)

Where I do use it is the route to my son's. That is 5+ minutes of city streets to the city freeway (no FSDb used at all because I don't use it on that sort of driving.) Then freeway to 4 lane divided highway interchange which in good weather FSDb can handle that but taking it at a higher speed (25% higher than the suggested speed) than I'm comfortable with so I scroll it down and back up. If the driving conditions are suspect (potential for black ice, wet, snowy), I disengage because my version of FSDb doesn't respond to driving conditions (except sunlight on the cameras.)

There's a slowdown at a particular point (have my foot on the go-pedal to nudge since this is not a real or imaginary speed change), then slowing down as the highway enters the outskirts of a town and becomes a city street, with two through lanes and assorted turning lanes, including an alternating shared left-turn lane for a section. At the opposite end of the town (~8 blocks and one traffic light) I need to pull into a dedicated left turn lane and turn onto a 2 lane highway (with passing lanes every 10km or so to get around slow moving vehicles, slower through a couple of villages, and a couple of traffic lights) and proceed 28km to my son's town where I disengage because of city streets.

I disengage before at the end of the 4 lane highway because it won't slow down smoothly from 100kph to 50kph as the highway ends and it keeps me in the right lane (I'm set to Chill) which is actually the worse lane for speed (several dozen business driveways) and difficulty to get into the required left turn lane due to traffic volume of cars in the better lane avoiding all the right-hand turning folks into the businesses. If I ask it to move to the left lane while on FSDb, it will but it moves back to the right lane as fast as possible, including within a half block of the beginning of the left turn lane I need. I've never made it through the down without disengaging. I've reengaged while in the left turn lane waiting at the red light and it has pulled forward when it got its protected turn light but since its path was into the (occupied) lane of the stopped cross traffic, the only time I tried it (an early version of V11), it failed, and I'm not in the mood to give it another chance until it can make it into town in the first place without me needing to disengage.

It has got better about moving into the passing lanes on the 2 lane highway but still dives into turning lanes (not consistently in the same spots, though.) Because I don't speed, I need to allow others to pass so I appreciate it being reliable on moving out of the way of others but it is unreliable for moving back when the lane ends; we still get occasional (but not predictable) drives along the shoulder when the lane ends and the car is 'surprised' (that we experiment with only when there are no other cars in sight.) Sometimes it will move into the through lane if we signal, sometimes it won't. We never signal asking for the lane change if there are other vehicles in the lane so we have no idea why it works sometimes and not others.

So there's no trust whatsoever, even on relatively simple parts of a mostly highway drive. Each time we update, we let it try (and for a brief time in the fall the predictable slowdown on the highway went away but it is back again with our current version.)

It is frustrating for me to see posts about all the improvements others see but while highway is much smoother, we are still getting occasional dives into turning lanes off 2 lane highways and no improvement on the stretch through the edge of town.

So 'nearly' every time I switch out it is because I'm anticipating or experiencing bad behaviour. Will the car crash if I don't disengage? Likely not but I might get pulled over for DUI-like behaviour (part of the drive is past a provincial police station so this behaviour being seen by an officer is not as unlikely as one might think).

As I only use FSDb for mostly highway drives, I'm not in any rush for V12 since the majority of my driving would still be under V11. But it has such a long way to go to get over 'comfort-level' disengagements, I doubt I'll see a big change in 2024 at all. I need my car to be consistent and safe all the time, not just mastering a 4 way stop or a particular UPL in the US while not managing to deal with the different road markings and speed change signage we have in my province.
Um. Didn't you say, in some thread or other, that you had managed, with Tesla's help, to uninstall FSD since it was a life-threatening danger of the first water?

If I remember correctly, we all wished you a happy life and goodbye.

So, you're using it again?