FSD v12.x (end to end AI)

[gsmith123]

It seems like this stuff would naturally automatically fall out of training, since virtually every training clip would have this defensive-driving behavior strongly demonstrated. So I'm curious if it is showing up! If it doesn't, I have no explanation for why it would not.

I have a somewhat poor theory about why it wouldn't show up:

They might have trained V12 heavily on synthetic data, for the "easy" driving situations. Synthetic data meaning artificially generated video where the driving decisions are similar to V11 (perhaps even reusing the V11 code!), rather than genuine human driving clips.

(Why might they do this? V11 is pretty safe for the situations you described, or at least not prone to being at fault, so they may have wanted to train V12 to a similar baseline before starting to address the details you're talking about. They may be focusing their data collection from human drivers on harder situations.)

I share your curiosity btw, and annoyance. This is on my list of top things to test in V12!

 

[Mardak]

Anyway it seems pretty clear that the situation is hopeless, based on what we are seeing with v12

There will be diminishing returns on a given architecture, but where do you think we're on that curve with 12.x so far? Potentially end-to-end and continuous disengagement feedback loop is still towards the left side of each of these AlphaZero Elo charts:

And one major difference from AlphaZero's random self-play is that Tesla is using human knowledge to proactively seek out problematic scenarios to hopefully keep up the rate of improvement.

 

[AlanSubie4Life]

There will be diminishing returns on a given architecture, but where do you think we're on that curve with 12.x so far? Potentially end-to-end and continuous disengagement feedback loop is still towards the left side of each of these AlphaZero Elo charts:

View attachment 1023090

And one major difference from AlphaZero's random self-play is that Tesla is using human knowledge to proactively seek out problematic scenarios to hopefully keep up the rate of improvement.

No idea. Not familiar with all these AI projects to know if applicable and how they worked exactly. And I definitely don't know about Elo scores, how linear they are, how/whether they would reflect performance for this task, etc.

Given that:
They've been working on this for a long time so suspect they're way up there, maybe 90% of the way there? Probably much harder to make improvements than it was originally. But no idea.

 

[Tronguy]

I would disagree vehemently about your characterization of "real time" as a "marketing" term. Briefly: with a real-time system you can know the worst-case latency when, for example, sampling A-to-D inputs at a high rate and knowing you'll have enough uninterrupted CPU cycles to process each sample without losing any data. That isn't a requirement of most general-purpose platforms. If you don't really need it, you can get better throughput on a non real-time system. But when you do need it, it's a whole different software architecture.

It's not just marketing.

All right. I didn’t exactly want to go into this, but what the heck. I’ll try and make it short.

Say that one has a satellite that zipping along between Earth and Neptune and it’s going to take years to do so. It’s got its star and planet cameras and can take angles and such as it goes, so it sort of knows where it is. Once in a while it does a course correction. What’s realtime for this thing? Sure doesn’t have to do all that processing and computation every microsecond; once a day is probably more than fine. Even in a gravity well, once every hour is probably fine. So a slow, low power CPU is called for.

Opposite extreme: One has a RADAR/ballastic gun attempting to shoot a supersonic jet out of the sky, and said jet is maneuvering. How fast does this CPU need to be? Well, it’s 6.18 us per RADAR range mile, the targeting RADAR is bouncing pulses off the target every 150 us and, well, things get nasty. New results 10us after receiving the latest data? So, now you know why things like Patriot missile systems have lots of trailers filled with compute systems and generators to keep all that power-hungry stuff happy.

There are likely situations where that Patriot system isn’t fast enough. I’ve programmed DSP systems that do audio processing that have to chug along with results faster than the Nyquist limit. But then there’s phased array RADARS that have to go faster than that. And what about 50 GSa/s o’scopes that have to dunk the data on the screen?

In all of these cases, it’s not about whether a piece of hardware is “realtime” or not. It’s strictly if the collection of hardware/software is as fast as it needs to be. Period.

First one figures the needed response time; then one looks at how to implement that, usually as cheaply as possible. Under these circumstances, some marketeer throwing around the word, “realtime” is just trying to attract eyeballs. Realtime for what, pray tell?

Cycling back to Teslas, one can make, I suppose, an argument that that’s a realtime computer in there. Fine: is that computer fast enough to track a hypersonic missile, or to control a spacecraft doing aerobraking when entering the atmosphere? No? Then it’s not realtime enough, is it?

Marketing term, that’s all.

 

[rlsd]

I agree that potential exists, in terms of broad-based functionality - e.g. the parking lots have looked fairly good (hard to tell from a video if it is remotely usable, but for sure way better than anything before!!!). I think the issue will be more likely the fails which will still crop up on occasion without much warning, in nearly every area.


I'd guess compute power, followed by resolution, followed by camera positions. Camera positions only impact a small set of domains. Resolution is an issue. But compute seems like the biggest issue. Not saying it can be "solved" by any of this of course.

=====================

For people with v12, who have used it on larger multi-lane surface streets (not freeways), some questions. Haven't really seen the focus on this in the videos, but haven't really watched any:

1) Does the car stay out of people's blind spots now, when traffic is light enough to allow it? And does it no longer just drive next to other people (not in blind spot, just next to them, annoyingly).
2) Does the car shift lane position slightly whenever anyone is a little off-center in their lane? (Not in situations where other car is making any lane changes, or drifting over the line, or cutting in - I just mean too off center to ignore, and would normally automatically mean a human would shift slightly to maintain buffer space?
3) Does the car subtly change speed, imperceptibly, to accomplish this maintenance of buffer zones, prioritizing them over follow distance?
4) In general, is the car no longer largely "locked" to the middle of the given lane? Does it wander around in the lane when appropriate, like a human?

(I saw the video avoiding puddles but that's obviously WAY beyond what I'm talking about here - and was honestly a bit odd and hard to explain.)

It seems like this stuff would naturally automatically fall out of training, since virtually every training clip would have this defensive-driving behavior strongly demonstrated. So I'm curious if it is showing up! If it doesn't, I have no explanation for why it would not. If the movement to avoid puddles showed up magically out of training, it seems like this subtle stuff should definitely show up, and we should be seeing it now.

There should be no longer a locking to the middle of the lane, nor should there be a fixed distance to traffic in front (instead prioritizing buffer zones, just as basically all good humans do).

Links to videos at proper time stamps, consistently showing this behavior, appreciated.

Not doing these behaviors is possibly the most serious issue increasing accident risk with routine v11 use, so it would be good to see that issue addressed consistently. I'm kind of surprised the early release people haven't been focused on it (maybe they have mentioned it and I missed it). Would make it nearly dreamy on large busy streets.

It’s a complicated problem, and solutions are highly dependent on traffic density and many other factors. Would be hard to code for every case, but should be solved easily by NNs.

I can't answer all your questions because I don't have an objective way to measure the behaviors of the car. I would need a second person sitting next to me to record detailed observations while I am concentrating on the road and trying to avoid the nags.

My general impression is the car moves smoothly most of the time on the streets. I just sit in the car listening to the music and enjoy watching people struggling with driving in stop and go traffic or doing bad things while driving (reading the phone, texting, shaving with a big razor blade, looking at the mirror to put on makeup,...).

One of the big improvements is FSD correctly estimates the speed and acceleration of the leading car to avoid blocking the intersection. It doesn't completely stay away of the intersection, most of the time half of the car is still in the intersection but it gives enough room for other cars to make turns. I think human drivers do the same thing to beat heavy and slow traffic. I was very embarrassed when my car was in the middle of the intersections with v11. A Tesla car blocking the intersection raises more eyebrows than other cars.

If the following things are fixed then V12 is very useful for my daily driving:

1. Creeping behavior for unprotected right turn. Unprotected left turn does not give me stress like unprotected right turn because my view and the car view are not blocked.

2. Not recognizing freeway entrance throttling lights. FSD run through red light all the time.

3. Transition between city stack and freeway stack. When entering freeway ramp FSD changes lane back to the ramp after it successfully merges to freeway.
When preparing to exit freeway, FSD takes the exit before the right exit (even when the 2 exits are more than 1 mile apart). This happens more at night and more frequently than V11.

4. Too slow to make turn or go through intersection after a stop sign.

 

[JulienW]

AI DRIVR you are FIRED. 2 weeks of v12 and only one video made the day you got it.

Totally sucks Tesla is excluding most of the normal influencers. Now all we get is new wannabes with a cell phone stuck to the sunroof and a horrible view of the sun visors. Come on Tesla and at least follow the "normal" procedure so we can have some quality videos to "lust" over while we wait.

Also Elon you need to My Space about how great and 12.3 is so we can build up more unrealistic hope of getting it soooooooon.

 

[Madmolecule]

My happiness is not dependent on whether I was scammed by Elon or not.
As an electrical engineer, batteries, AC inverters, and electric motors are pretty straightforward and that is the main reason I bought an electric vehicle. I also wanted Dog mode

I will continue to not vent but express by dissatisfaction in the Kickstarter scam called FSD, and all the shills that continue to promote it like it will be released as a product that will drive my car without me having to pay 100% attention and be 100% responsible for its actions. Tesla has made no attempt at getting FSD approved without driver supervision . They have shown no plan on how this is going to happen or how you can insure the vehicle.

I will continue to drive my electric car, it is paid for it only has 80,000 miles, I just spent four grand upgrading the stereo, from the crap premium system that comes with it. Also adding a lot of noise insulation, which sucks on the 2018s

I will continue to express my opinion as an owner on an owners forum, not to try to change the fanboys, but maybe to prohibit somebody from getting scammed by buying into the FSD BS.

The richest man in the world has made a lot of promises. I am waiting to see if he’ll back up any of them. I am a Tesla Long, that invested in the product, and not just playing the stock market. The fact is the product was sold with a lot of vaporware, and hype. And yes, I will join the lawsuit when it gets enough momentum. There is no sympathy in the courts for computers killing people. Once Dojo wipes out a few more pedestrians. I still remember how much Elon wines about the pedestrian warning system they had to put in the car. The 2018 somehow cannot have it, even though I got the new upgrade computer. If you driven an electric car long enough, you have definitely been behind somebody in a parking lot that did not know you were there. Boombox never happened. You would think they would use the PWS system to notify the people crossing the street to go ahead, I thought the car was smart. At least play a fart noise to let them know to proceeded.

Currently, I am being held 100% responsible for controlling a cars computer that has documented hallucinations, and only uses vision. The other senses do not seem to matter to Tesla, only vision. This might be Tunnel vision, ultrasonic and radar companies will be going out of business very soon. LiDAR is for suckers. Single stack end to end BS is the only way, watch Tesla and learn.

That is another reason why I will keep the car is to remind potential buyers and the forum of the false promises, bravado which can only be interpreted as a scam.

I keep waiting to be proved wrong. Do you think 10 years is the proper time for me to wait?
Do you still believe that they will be releasing a mass produced long haul semi for coast to coast travel. Do you still believe a roadster will be coming out, that can fly with cold air thrusters?

I want to believe, but come on. Book my ticket to Mars, I got an extra hundred grand lying around.

 

[gottagofast]

My happiness is not dependent on whether I was scammed by Elon or not.
As an electrical engineer, batteries, AC inverters, and electric motors are pretty straightforward and that is the main reason I bought an electric vehicle. I also wanted Dog mode

I will continue to not vent but express by dissatisfaction in the Kickstarter scam called FSD, and all the shills that continue to promote it like it will be released as a product that will drive my car without me having to pay 100% attention and be 100% responsible for its actions. Tesla has made no attempt at getting FSD approved without driver supervision . They have shown no plan on how this is going to happen or how you can insure the vehicle.

I will continue to drive my electric car, it is paid for it only has 80,000 miles, I just spent four grand upgrading the stereo, from the crap premium system that comes with it. Also adding a lot of noise insulation, which sucks on the 2018s

I will continue to express my opinion as an owner on an owners forum, not to try to change the fanboys, but maybe to prohibit somebody from getting scammed by buying into the FSD BS.

The richest man in the world has made a lot of promises. I am waiting to see if he’ll back up any of them. I am a Tesla Long, that invested in the product, and not just playing the stock market. The fact is the product was sold with a lot of vaporware, and hype. And yes, I will join the lawsuit when it gets enough momentum. That is another reason why I will continue to the car and continue to remind people and the forum of the false promises, bravado which can only be interpreted as a scam.

I keep waiting to be proved wrong. Do you think 10 years is the proper time for me to wait?
Do you still believe that they will be releasing a mass produced long haul semi for coast to coast travel. Do you still believe a roadster will be coming out, that can fly with cold air thrusters?

I want to believe, but come on. Book my ticket to Mars, I got an extra hundred grand lying around.

I feel like FSD does mostly what it’s called, I let it drive through my local town and I didn’t touch the accelerator or brake and only touched the steering wheel if it asked me too, isn’t that basically a car driving itself?

Just the car stopping at stop lights and going at green lights by itself without me doing anything is impressive me to, even something as simple as that wasn’t even possible less than 10 years ago

 

[Madmolecule]

I feel like FSD does mostly what it’s called, I let it drive through my local town and I didn’t touch the accelerator or brake and only touched the steering wheel if it asked me too, isn’t that basically a car driving itself?

Just the car stopping at stop lights and going at green lights by itself without me doing anything is impressive me to, even something as simple as that wasn’t even possible less than 10 years ago

It doesn’t fully drive your car at all. You are responsible for 100% supervision and to take control when it screws up.
There is 0 miles logged in self driving mode. The billions of miles have been in driver assisted mode.
The promises of the car driving itself or being a Robo taxi has no hope in happening in the future as I see it.
Does your insurance or does Tesla insurance allow for the car to be driving without your 100% supervision, no
Who is liable when the car has an accident when you are not paying attention, the answer is you? How can that be full self driving?
The promises that were made by Elon were not that it would stop at a parking light and do parlor tricks
The promises that it would drive to coast without supervision. He promised that you could fly to New York and have your car meet you there
This is what I was sold. I also was sold that it would be the new valet. That would drop you off at the grocery store and pick you up in the rain. My attempts using summon has been a disaster, dangerous, and embarrassing. Advanced summon is no better.
After five years, I can’t even use it to park in my garage, or back out of my garage safely

I’m sorry I’m not impressed with visualizing traffic cones, or stopping at a stoplight. Most new cars have advanced driver assistance that can keep you in your lane.

No, this was not possible 10 years ago. But seven years ago, Elon stated that they have it working. That was clearly a lie and hype to pump stock. I actually heard him say that it was going remember your tire positions to be able to navigate through a parking garage without GPS.
Please tell me this was not BS, but you cannot

As a senior citizen, I hope that this car would take over the driving for me and my golden years, But this is clearly not going to happen.

Over there updates = they never have to complete the software

They’ve given up on the old cars and are working on the new models. How big of a development team do they have on hardware 3.5

 

[gottagofast]

It doesn’t fully drive your car at all. You are responsible for 100% supervision and to take control when it screws up.
There is 0 miles logged in self driving mode. The billions of miles have been in driver assisted mode.
The promises of the car driving itself or being a Robo taxi has no hope in happening in the future as I see it.
Does your insurance or does Tesla insurance allow for the car to be driving without your 100% supervision, no
Who is liable when the car has an accident when you are not paying attention, the answer is you? How can that be full self driving?
The promises that were made by Elon were not that it would stop at a parking light and do parlor tricks
The promises that it would drive to coast without supervision. He promised that you could fly to New York and have your car meet you there
This is what I was sold. I also was sold that it would be the new valet. That would drop you off at the grocery store and pick you up in the rain. My attempts using summon has been a disaster, dangerous, and embarrassing. Advanced summon is no better.
After five years, I can’t even use it to park in my garage, or back out of my garage safely

I’m sorry I’m not impressed with visualizing traffic cones, or stopping at a stoplight. Most new cars have advanced driver assistance that can keep you in your lane.

No, this was not possible 10 years ago. But seven years ago, Elon stated that they have it working. That was clearly a lie and hype to pump stock. I actually heard him say that it was going remember your tire positions to be able to navigate through a parking garage without GPS.
Please tell me this was not BS, but you cannot

As a senior citizen, I hope that this car would take over the driving for me and my golden years, But this is clearly not going to happen.

Over there updates = they never have to complete the software

They’ve given up on the old cars and are working on the new models. How big of a development team do they have on hardware 3.5

Yes I’m responsible if the car wrecks, but technically it’s still driving it self, I’m not making it go or stop or turning it, it’s doing that itself

 

[lzolman]

All right. I didn’t exactly want to go into this, but what the heck. I’ll try and make it short.

Say that one has a satellite that zipping along between Earth and Neptune and it’s going to take years to do so. It’s got its star and planet cameras and can take angles and such as it goes, so it sort of knows where it is. Once in a while it does a course correction. What’s realtime for this thing? Sure doesn’t have to do all that processing and computation every microsecond; once a day is probably more than fine. Even in a gravity well, once every hour is probably fine. So a slow, low power CPU is called for.

Opposite extreme: One has a RADAR/ballastic gun attempting to shoot a supersonic jet out of the sky, and said jet is maneuvering. How fast does this CPU need to be? Well, it’s 6.18 us per RADAR range mile, the targeting RADAR is bouncing pulses off the target every 150 us and, well, things get nasty. New results 10us after receiving the latest data? So, now you know why things like Patriot missile systems have lots of trailers filled with compute systems and generators to keep all that power-hungry stuff happy.

There are likely situations where that Patriot system isn’t fast enough. I’ve programmed DSP systems that do audio processing that have to chug along with results faster than the Nyquist limit. But then there’s phased array RADARS that have to go faster than that. And what about 50 GSa/s o’scopes that have to dunk the data on the screen?

In all of these cases, it’s not about whether a piece of hardware is “realtime” or not. It’s strictly if the collection of hardware/software is as fast as it needs to be. Period.

First one figures the needed response time; then one looks at how to implement that, usually as cheaply as possible. Under these circumstances, some marketeer throwing around the word, “realtime” is just trying to attract eyeballs. Realtime for what, pray tell?

Cycling back to Teslas, one can make, I suppose, an argument that that’s a realtime computer in there. Fine: is that computer fast enough to track a hypersonic missile, or to control a spacecraft doing aerobraking when entering the atmosphere? No? Then it’s not realtime enough, is it?

Marketing term, that’s all.

I don't think you've invalidated what I said. A "real-time" system makes it possible to know worst-case latency for a computation, other systems simply don't. That's why there's OSs such as QNX, because the real-time-ness has to be at the OS level. Its why C++ is the language used for landers and rovers (and autonomy! I've taught advanced C++ to about 5 different autonomous driving teams at various companies) and such, and not Java or some other managed language.

It's not just a marketing term.

 

[JB47394]

And one major difference from AlphaZero's random self-play is that Tesla is using human knowledge to proactively seek out problematic scenarios to hopefully keep up the rate of improvement.

Another is that they could make the network as large as was needed for the task - which was comparatively simple. Given enough good data, the system may train up as shown in those charts, but once you're trained up, what do you have? A system that can handle complex urban navigation but which cannot maintain a constant speed on clear secondary roads? Or a system that can squeeze through tight sections in traffic, but also crowds pedestrians?

My point is that, apart from training up, there will also be a period of packing, where the Tesla engineers will have to learn how to cram as much competency into the available network capacity as possible. I suspect that it won't be as simple as "urban driving is 20 units", "pedestrian management is 10 units" and so on. I suspect that it will involve lots of thinking along the lines of Venn diagrams, where some tasks overlap others, and you pay full price if you include only one, but less than full price if you take both of the related tasks. It'll be a bit like herding cats.

As far as I know, that 'packing' task isn't something that's been done before, making it another area of research. As customers, I can imagine that it might be a pretty aggravating experience. "It was doing so well spotting dogs. Now it's rubbish." "Yeah, but look how well it manages sharp curves on the highway. That was always a problem". And so on.

I’ll try and make it short.

Keep at it. You'll get there eventually.

 

[kabin]

All right. I didn’t exactly want to go into this, but what the heck. I’ll try and make it short.

Say that one has a satellite that zipping along between Earth and Neptune and it’s going to take years to do so. It’s got its star and planet cameras and can take angles and such as it goes, so it sort of knows where it is. Once in a while it does a course correction. What’s realtime for this thing? Sure doesn’t have to do all that processing and computation every microsecond; once a day is probably more than fine. Even in a gravity well, once every hour is probably fine. So a slow, low power CPU is called for.

Opposite extreme: One has a RADAR/ballastic gun attempting to shoot a supersonic jet out of the sky, and said jet is maneuvering. How fast does this CPU need to be? Well, it’s 6.18 us per RADAR range mile, the targeting RADAR is bouncing pulses off the target every 150 us and, well, things get nasty. New results 10us after receiving the latest data? So, now you know why things like Patriot missile systems have lots of trailers filled with compute systems and generators to keep all that power-hungry stuff happy.

There are likely situations where that Patriot system isn’t fast enough. I’ve programmed DSP systems that do audio processing that have to chug along with results faster than the Nyquist limit. But then there’s phased array RADARS that have to go faster than that. And what about 50 GSa/s o’scopes that have to dunk the data on the screen?

In all of these cases, it’s not about whether a piece of hardware is “realtime” or not. It’s strictly if the collection of hardware/software is as fast as it needs to be. Period.

First one figures the needed response time; then one looks at how to implement that, usually as cheaply as possible. Under these circumstances, some marketeer throwing around the word, “realtime” is just trying to attract eyeballs. Realtime for what, pray tell?

Cycling back to Teslas, one can make, I suppose, an argument that that’s a realtime computer in there. Fine: is that computer fast enough to track a hypersonic missile, or to control a spacecraft doing aerobraking when entering the atmosphere? No? Then it’s not realtime enough, is it?

Marketing term, that’s all.

Semantic distractions are a dead end. There will always be technological/cost/weight/size limitations. We have to work within real world constraints.

More importantly, is FSD using enough sensors and collecting sufficient data to safely respond to the environment? I've yet to see v11 or v12 equipped vehicle respond quicker than 1.5 seconds even though processing keeps up with realtime input data.

 

[Tronguy]

Semantic distractions are a dead end. There will always be technological/cost/weight/size limitations. We have to work within real world constraints.

More importantly, is FSD using enough sensors and collecting sufficient data to safely respond to the environment? I've yet to see v11 or v12 equipped vehicle respond quicker than 1.5 seconds even though processing keeps up with realtime input data.

Agreed about the semantic distractions. But, reading over your response, I'm curious.

On the one hand, say the car is going straight down a road, highway or otherwise. Pretty clearly there's some kind of control loop running that's keeping the car in the middle of a lane on, say, a highway. I think I'd argue that said loop has a reaction time with respect to bumps on the road, wind gusts blowing sideways, and other disturbances that's a heck of a lot shorter than 1.5 seconds.

And I've been in a couple-three situations where, ahead of me, some car went into a panic stop; in those cases, the car in FSD/EAP seemed to react to those events very appropriately, quickly, and faster than 1.5 seconds.

Now, I think you may have a point about, I dunno, driving around a dump truck in front of one. Or deciding to take a turn with cars all over. Or what? Comment, please?

 

[AlanSubie4Life]

I've yet to see v11 or v12 equipped vehicle respond quicker than 1.5 seconds even though processing keeps up with realtime input data.

Yes. This is mysterious, but 1-second plus reaction seems to be the norm. Not sure why.

And I've been in a couple-three situations where, ahead of me, some car went into a panic stop; in those cases, the car in FSD/EAP seemed to react to those events very appropriately, quickly, and faster than 1.5 seconds.

I have not seen any documented (with video) event showing superhuman reaction times - except POSSIBLY to lead vehicles.

I do agree that to lead vehicles, reaction time may be faster. Still not superhuman, probably. And even there the reaction is often exaggerated and non-optimal and sometimes delayed depending on the stopping profile of the vehicle in front.

But that would be a very limited (though important) special case. For FSD there are many other cases, all of them appearing to be horribly delayed at this point.

I could understand Tesla deliberately slowing reaction time to human reaction time (could be problematic to react too quickly in some cases), but this is way slower than that.

 

[swedge]

Since ver 12 is Ai based isn’t it supposed to be constantly improving on its own with not much input from Tesla?

My understanding is that the "P" in Chat GPT stands for pre-trained. Training is done at great expense in computation to create a model. Once that is done, using the model to answer questions or drive the car is what we do, and this typically does not further train the model. Feedback from users is used to improve the training for the next version. So the system does learn, but not in the individual cars.

 

[Madmolecule]

Yes I’m responsible if the car wrecks, but technically it’s still driving it self, I’m not making it go or stop or turning it, it’s doing that itself

Yes, you paid a lot of money, to drive the car yourself

 

[gottagofast]

Yes, you paid a lot of money, to drive the car yourself

I actually got for free, it was already on the car I bought lol, I’m not driving when I use it, I’m just paying attention to what’s around me, I’m more like a passenger that can intervene if I need to

I wish there was an option where you could make the car favor being closer to edge of the road though when your on a 2 lane road with oncoming traffic

 

[FSDtester#1]

Yes, you paid a lot of money, to drive the car yourself

You and anyone else who bought or pay for FSD via a subscription did so without a gun to your head.

It was barely useful when you purchased it. Buying something with the promise it will get better in the future, is a buyer beware situation.
It seems you are very Dissatisfied with Elon and his promises.

I would suggest you sell any Tesla's you own and go purchase another brand that has autonomous driving solved already.

Oh, sorry, they don't exist.

This is not the Fsd therapy thread. I am certain there are others on Tmc.

Take the complaints somewhere else. They don't belong here.

This thread are for people who want to share their experiences of the current software and it's abilities, good or bad, and their speculation about what direction they think it's going. If anyone wants to continually trash the state of current fsd, as long as they own a Tesla with fsd, and are speaking from first hand experiences, not youtube videos, Fire away and let the debate begin. This is not what you are arguing.

I am sorry you feel duped, I really am, but this is the wrong thread for these style complaints.

 

[gottagofast]

Will the nag eventually make a noise if you don’t notice it? There’s been times when I even had my hand on the wheel but wasn’t giving it enough force I guess, If it does make a noise does that count as a strike? If it makes a noise how much time do you have before it counts as a strike