FSD discussion

[Singuy]

Your argument is that CS is right to priority driver monitoring. Due to Bluecruise being the winner for having the best driver monitoring and perhaps the most cautious system, my argument is that this lost the plot on why these systems even exist.

This is my take

If an adas system can keep the car driving without crashing for 5 additional minutes when someone fell asleep, that 5 extra minutes worth of time that person could potentially wake up and not die. The alternative is the person fell asleep going 80mph and crashes within 10 seconds on manual driving. I take the 5 extra minutes over the 10 second any day.

You start losing the plot when you try to make your system so difficult to enable and kept enabled with a bunch of geo maps and monitoring, that it fails to serve the purpose, which is to give that person a surviving chance as that person falls asleep. So performance and ease of use should be the ONLY metrics that matters here. I understand why they decide to add so much monitoring and geo fencing, which is to prevent people from falling asleep PURPOSELY while reducing the chance of crash WHEN the system is enabled to a min . This is a backwards IMO, especially when it shuts itself off while the driver is doing something dangerous like closing their eyes. The system is suppose to shine the most when trying to keep a person who kept their eyes closed alive as long as possible, not to shut itself off to cause a crash.

 

[AndrewZ]

I’ve never used BlueCruise, but I doubt it just shuts off and sends you careening ahead when it detects your eyes are closed when even the basic lane keeping was more robust.

AP is also supposed to throw warnings and take action when it detects disengagement.

I haven’t argued anything towards CS’s ranking, I have argued that some of the geofencing criticism from Munro’s video was silly. And it was silly, because you’re also not supposed to be using AP in situations similar to those that were locked out in BlueCruise. And I’ve argued that engagement monitoring is and will continue being very important, because none of these systems are perfect and won’t be perfect for who knows how long.



Have you used BlueCruise? Or from where are you passing judgment on its use case

 

[Singuy]

I’ve never used BlueCruise, but I doubt it just shuts off and sends you careening ahead when it detects your eyes are closed when even the basic lane keeping was more robust.

AP is also supposed to throw warnings and take action when it detects disengagement.

I haven’t argued anything towards CS’s ranking, I have argued that some of the geofencing criticism from Munro’s video was silly. And it was silly, because you’re also not supposed to be using AP in situations similar to those that were locked out in BlueCruise. And I’ve argued that engagement monitoring is and will continue being very important, because none of these systems are perfect and won’t be perfect for who knows how long.



Which ADAS systems have you used exactly?

Based on the Munro video, blue cruise shut itself off due to the curve and and the driver didn't even know as the audible chime was very soft at hwy speed.

I have only use Tesla's. My friend use volvos and it can't lane keep for longer than 30 seconds before running him off the road.

 

[AndrewZ]

Based on the Munro video, blue cruise shut itself off due to the curve and and the driver didn't even know as the audible chime was very soft at hwy speed.

I have only use Tesla's. My friend use volvos and it can't lane keep for longer than 30 seconds before running him off the road.

Where do you see that? I just watched it again, Sandy noticed BlueCruise alerting him of the upcoming sharp curve, took control, and the guy with Ford pointed out how good it was that the display graphics etc caught his attention as they’re intended to. That video is closing in on two years old now, and things can be tweaked.

We’re taking Lane Keeping from Volvo or what system exactly? I imagine you can cite the latest version of FSDBeta in the fleet right now, or 10.69, or v11.

 

[Singuy]

Where do you see that? I just watched it again, Sandy noticed BlueCruise alerting him of the upcoming sharp curve, took control, and the guy with Ford pointed out how good it was that the display graphics etc caught his attention as they’re intended to. That video is closing in on two years old now, and things can be tweaked.

We’re taking Lane Keeping from Volvo or what system exactly? I imagine you can cite the latest version of FSDBeta in the fleet right now, or 10.69, or v11.

Volvo's motion drive or whatever they call the thing. The top of the line stuff. He sends me video and is just frustrated that it can't even do straight line right. He is a Tesla owner, Volvo Ev forced on him through work.

 

[daniel]

Road slippery = don't engage AP/FSD.

Other than that FSD has come a long way. It is quite good even at sharp turns. I've a few winding roads with nearly 90 degree turns and FSDb used to struggle early last year. Now it can handle those smoothly.

I've never used FSD or FSDb. My issue with them is that they're intended for city use, and in the city, things happen much more quickly than on the highway. On the highway, I can see well ahead when I'm about to enter an area where I would not trust EAP. In the city there are more obstacles, pedestrians, cross traffic, etc. There are lane changes due to traffic flow or navigation, unprotected left turns, etc.

From what I've seen on here, I don't think that FSD is ready. But as noted, I've not used it and I don't know what the latest improvements have changed. I think it is fraudulent advertising to call a Level 2 system "Full Self-Driving."

... I drove a rental 2021 F150 Lariat for work two years ago that had simple lane keeping, and I was doing long highway hauls. Even with a basic implementation like that, it would detect you weaving between lane lines. First detection it would shake the wheel. Second detection it would shake the wheel and start throwing warnings that you should pull over. In a more advanced implementation that is likely headed our way, such a system would pull the vehicle over. ...

A system capable of pulling over and stopping safely is years, maybe decades ahead of a lane-keeping system. When you say "in a future implementation..." I think you're talking ten or twenty years from now. Maybe more.

I agree that driver-assist systems make cars much safer, when used as intended. And true self-driving cars will be much safer. But let's not forget that simple enforcement of traffic laws, and treating infractions with the severity their consequences merit, would go a long way to making our roads safer.

 

[AndrewZ]

I've never used FSD or FSDb. My issue with them is that they're intended for city use, and in the city, things happen much more quickly than on the highway. On the highway, I can see well ahead when I'm about to enter an area where I would not trust EAP. In the city there are more obstacles, pedestrians, cross traffic, etc. There are lane changes due to traffic flow or navigation, unprotected left turns, etc.

From what I've seen on here, I don't think that FSD is ready. But as noted, I've not used it and I don't know what the latest improvements have changed. I think it is fraudulent advertising to call a Level 2 system "Full Self-Driving."



A system capable of pulling over and stopping safely is years, maybe decades ahead of a lane-keeping system. When you say "in a future implementation..." I think you're talking ten or twenty years from now. Maybe more.

I agree that driver-assist systems make cars much safer, when used as intended. And true self-driving cars will be much safer. But let's not forget that simple enforcement of traffic laws, and treating infractions with the severity their consequences merit, would go a long way to making our roads safer.

Yeah this is why I think driver engagement should get an emphasis, because I also believe vehicles doing this stuff reliably is off in the distant future. In the interim it's all about the human-machine interface and leveraging strengths while working to eliminate weaknesses. Autopilot is one of the most capable systems, and it can be made better -- all these systems can and will be made better. Autopilot hasn't had a meaningful update in a long time and other systems have closed the gap, hopefully single stack will change that.

I personally haven't used BlueCruise, or SuperCruise, or Volvo's system, or almost every system on CR's list. I have used Highway Driving Assist II on my Genesis, which is the newer version of the system CR tested, and I've gone hours on the highway with barely touching the wheel. The Genesis also watches you constantly in all drive modes and throws warnings when it notices you checking your phone etc.


And I would not pay one red cent extra for the lane keeping I experienced in the 2021 F150, because I know what my experience was: it all but encouraged me to disengage from the driving task and take risks. Maybe it'll work out, maybe it won't, but I don't need that temptation. Shaking the wheel and throwing alerts when weaving in the lane lines though, that was solid.

 

[STS-134]

I’ve joked on here before that Tesla could make FSDBeta the default drive mode and the human would be there to monitor. Accident rates would plummet, psychiatric bills and anxiety med usage would skyrocket.

Accident rates would plummet during conditions that FSDb works, and skyrocket whenever humans have to take over. If you don't maintain your skills, but somehow are expected to take over only in the most difficult conditions, you're going to fail in those conditions.

 

[AndrewZ]

Accident rates would plummet during conditions that FSDb works, and skyrocket whenever humans have to take over. If you don't maintain your skills, but somehow are expected to take over only in the most difficult conditions, you're going to fail in those conditions.

I had never thought about that but feels very apt and a good example of trade-offs implementing this type of stuff in the real world that is full of complexity

Maybe the car should force you to play a driving simulator while it operates to keep your skills sharp until it needs you to take over

 

[STS-134]

I had never thought about that but feels very apt and a good example of trade-offs implementing this type of stuff in the real world that is full of complexity

The purpose of automation should be to reduce workload on the driver. It should never, EVER be used in a manner that allows the driver to lose situational awareness by not paying attention. If the system suddenly requires you to take control, and you've been reading a news article or watching videos on your phone, you don't really have the 10-20 seconds to figure out what's going on around you; you'd better already be aware of what's happening.

 

[EVNow]

I've never used FSD or FSDb. My issue with them is that they're intended for city use, and in the city, things happen much more quickly than on the highway. On the highway, I can see well ahead when I'm about to enter an area where I would not trust EAP. In the city there are more obstacles, pedestrians, cross traffic, etc. There are lane changes due to traffic flow or navigation, unprotected left turns, etc.

Ofcourse FSD is not ready ... it is not "fully autonomous". You need careful monitoring.

BTW, FSDb is very useful to me in my suburban area. I use it 95%+ of the time. I'm very careful and put both hands on the wheel where it might screw up (basically any intersection or near pedestrians). Infact I'd say, I use it similar to how I use NOA/EAP. I let it take the exits sometimes, but many times I disengage and make the exits or drive to another freeway. After 10.69.x came out I almost never have to intervene at turns, FSD handles them ok. It is just bad at Roundabouts and picking the right lane which cause most of my disengagements.

BUT, it IS better than EAP. Remember EAP is now some 3 years old with hardly any improvements. If you give FSD the same conditions as EAP, it will do better. But there might be some regressions ... thats why V11 is not out yet.

I think it is fraudulent advertising to call a Level 2 system "Full Self-Driving."

FSD is just a trade term now - don't expand, doesn't mean what it might say. LG doesn't mean Life is always Good

 

[AndrewZ]

I've personally never thought LG's name suggested that Life is always Good, I think many people would argue that Full Self-Driving did sell a certain perception along with what was sold by Elon.

In reality, anyone who has purchased a HW3 vehicle was only sold a Level 2 ADAS that will exceed human safety by Tesla's metrics and those vehicles will never reach what most would consider Full Self-driving or unlock robotaxis, massively increased utilization, revenue generated while you sleep or at work, etc.

Put this all together and I can't help but be in awe at the skillful marketing

 

[daniel]

Ofcourse FSD is not ready ... it is not "fully autonomous". You need careful monitoring.

BTW, FSDb is very useful to me in my suburban area. I use it 95%+ of the time. I'm very careful and put both hands on the wheel where it might screw up (basically any intersection or near pedestrians). Infact I'd say, I use it similar to how I use NOA/EAP. I let it take the exits sometimes, but many times I disengage and make the exits or drive to another freeway. After 10.69.x came out I almost never have to intervene at turns, FSD handles them ok. It is just bad at Roundabouts and picking the right lane which cause most of my disengagements.

View attachment 906136

BUT, it IS better than EAP. Remember EAP is now some 3 years old with hardly any improvements. If you give FSD the same conditions as EAP, it will do better. But there might be some regressions ... thats why V11 is not out yet.

I don't doubt that "FSD" is "better" than EAP. My concern is that it's intended for use in conditions (city driving) where it will be more difficult for the driver to react in time to conditions the car cannot handle. On the highway, where AP/EAP is intended to be used, reaction times are not as critical. Tesla apparently agrees that FSDb is problematic, since you have to "qualify" to get it. First you pay, and then you have to convince them that you're responsible enough and capable enough to use it safely. (And I approve of them being very particular about who gets it.)

FSD is just a trade term now - don't expand, doesn't mean what it might say. LG doesn't mean Life is always Good

LG's motto "Life is good" is universally understood to be a platitude. Nobody buys an LG TV believing that their life will change. Tesla aficionados understand that "FSD" is just a marketing term and does not mean "full self-driving." But many people in the general public believe that Tesla cars drive themselves, and cause accidents. When there's an accident, the news reports don't say "Idiot driver ignored the instructions to stay aware." They say "Tesla FSD crashed the car."

This is exacerbated by Musk's promises over the years that a Tesla will be able to drive across the country without human intervention "real soon now." (Even if they do solve actual self-driving, AFAIK there are no plans to automate connecting to a supercharger. So Musk is misrepresenting both Tesla's capabilities AND its intentions.)

 

[Bet TSLA]

I don't doubt that "FSD" is "better" than EAP. My concern is that it's intended for use in conditions (city driving) where it will be more difficult for the driver to react in time to conditions the car cannot handle. On the highway, where AP/EAP is intended to be used, reaction times are not as critical.

Be "concerned" all you like. Many who don't drive EVs are "concerned" about charging problems that don't exist. Just don't pretend that your opinion is worth anything when it comes from common sense applied to ignorance.

Posting about something you don't use on a thread populated mostly by people with extensive experience will just get you into trouble.

Tesla apparently agrees that FSDb is problematic, since you have to "qualify" to get it. First you pay, and then you have to convince them that you're responsible enough and capable enough to use it safely. (And I approve of them being very particular about who gets it.)

No, you don't. The safety score qualification was eliminated.

... But many people in the general public believe that Tesla cars drive themselves, and cause accidents. When there's an accident, the news reports don't say "Idiot driver ignored the instructions to stay aware." They say "Tesla FSD crashed the car."

Many people believe all sorts of idiotic things. It's sad, but that doesn't mean Tesla has any obligation to make non-customers less ignorant.

This is exacerbated by Musk's promises over the years that a Tesla will be able to drive across the country without human intervention "real soon now." (Even if they do solve actual self-driving, AFAIK there are no plans to automate connecting to a supercharger. So Musk is misrepresenting both Tesla's capabilities AND its intentions.)

Seriously? You imagine that Tesla would make you aware of all their plans?

Agreed that Musk's endless optimism in the face of being wrong over and over is not a good thing. But by the time full autonomy shows up, I imagine Tesla will have done something to address the autonomous charging problem. It's a much, much easier problem than autonomous driving.

 

[EVNow]

I've personally never thought LG's name suggested that Life is always Good, I think many people would argue that Full Self-Driving did sell a certain perception along with what was sold by Elon.

In reality, anyone who has purchased a HW3 vehicle was only sold a Level 2 ADAS that will exceed human safety by Tesla's metrics and those vehicles will never reach what most would consider Full Self-driving or unlock robotaxis, massively increased utilization, revenue generated while you sleep or at work, etc.

Put this all together and I can't help but be in awe at the skillful marketing

I think this has been argued ad nauseum.

Pre-March '19 FSD language on the site was closer to full autonomy, but after that the language is more practical. It is debatable whether Elon was just plain wrong about how easy it was to achieve FSD or was lying. I think there are enough indications to think he was plain wrong (like his statements about how things with NN will improve "exponentially" or comparisons to NN learning to play games slow in the beginning and quickly achieving super-human abilities).

One of the issues I think is "Elon is so brilliant, he could not have been wrong" thinking. The fact is Elon has been very wrong about FSD .... atleast in terms of timelines. Whether he will be just late or the current hardware / vision only approach can't achieve autonomy still needs to be seen.

 

[AndrewZ]

I think this has been argued ad nauseum.

Pre-March '19 FSD language on the site was closer to full autonomy, but after that the language is more practical. It is debatable whether Elon was just plain wrong about how easy it was to achieve FSD or was lying. I think there are enough indications to think he was plain wrong (like his statements about how things with NN will improve "exponentially" or comparisons to NN learning to play games slow in the beginning and quickly achieving super-human abilities).

One of the issues I think is "Elon is so brilliant, he could not have been wrong" thinking. The fact is Elon has been very wrong about FSD .... atleast in terms of timelines. Whether he will be just late or the current hardware / vision only approach can't achieve autonomy still needs to be seen.

I can buy that it was mostly being plain wrong at the front end but somewhere along the line it became lying by omission, very carefully wording things to maintain the pre-established perception despite fully knowing reality. Things on the website were quietly tweaked over time to limit raising alarms, and those tweaks were likely made in response to things happening behind closed doors. Elon being asked about Level 4-5 in interviews and on earnings calls since then? Carefully deflected, the narrative subtly adjusted.

Around the same time as the language change we had leaked emails and letters from 2019-2020 between Tesla and the Cali DMV that laid out exactly the path we're on now. People didn't want to believe it and came up with all sorts of reasons why Tesla might be claiming City Streets is and would remain SAE Level 2, surely it was just so development wouldn't be hampered while they march towards robotaxis.

I could also buy those theories when believing upgrades would be offered throughout but now we know, HW3 is all but done. It's on to the next iterations exactly as described to the DMV, we're even talking about HW5 now. A HW3 upgrade isn't required because FSD as sold has been achieved, nothing beyond this was promised. I'd love to be proven wrong but really think this is the way and it had been the private plan for many years now.

 

[Buckminster]

Nice update:

https://twitter.com/x/status/1624693067143069702

 

[EVNow]

I can buy that it was mostly being plain wrong at the front end but somewhere along the line it became lying by omission, very carefully wording things to maintain the pre-established perception despite fully knowing reality.

IMO, the AP team is more realistic than Elon - who continued (continues?) to believe robotaxi is around the corner. He even talked about it as the next vehicle on ER that got very cold reception (and SP dropped).

If Elon continues to say end of the year FSD will be ready - one of these years, it might come true !

My personal belief is that FSD will get very good by 2025 (1 disengagement in 100 miles like NOA) and possibly autonomous (1 disengagement in 10k miles) by the end of the decade.

 

[AndrewZ]

IMO, the AP team is more realistic than Elon - who continued (continues?) to believe robotaxi is around the corner. He even talked about it as the next vehicle on ER that got very cold reception (and SP dropped).

If Elon continues to say end of the year FSD will be ready - one of these years, it might come true !

My personal belief is that FSD will get very good by 2025 (1 disengagement in 100 miles like NOA) and possibly autonomous (1 disengagement in 10k miles) by the end of the decade.

FSD can theoretically but it doesn't necessarily mean Gen 3 hardware will get there. Just watch Gen 4 will be HD radar, I see a new GPS antenna on the recent Euro approval, it's probably more redundancies across the board, who knows what

 

[EVNow]

FSD can theoretically but it doesn't necessarily mean Gen 3 hardware will get there. Just watch Gen 4 will be HD radar, I see a new GPS antenna on the recent Euro approval, it's probably more redundancies across the board, who knows what

Not talking about hardware at all - my prediction is just on software front. I think the gap in hardware is much smaller than the gap in software. Just see all the other companies (like Waymo, Cruise) - they all struggle with $100k worth of hardware and every sensor they want to expand beyond their Geo boundaries. If you see how Tesla FSD performs on routes that people on YT have compared to Waymo & Cruise, there isn't a big difference. So, I'm not convinced much better HW gets anyone much closer to autonomy.