FSD discussion

[Dikkie Dik]

Still can't edit ...

In case you’re still looking for that edit button and you’re on a phone, go landscape.

Annoyingly it’s missing in portrait but landscape is fine.

 

[diplomat33]

The problem with Cruise and Waymo are the inability for these solutions to SCALE UP in any reasonable time-frame.

Specifically:
1) they are geofenced to TINY TINY areas
2) they require cars with thousands, sometimes tens of thousands of dollars of extra equipment
3) until recently Cruise could only operate in low-traffic environments (and both still completely avoid highway driving) - tell-tale sign of "cherry picking" the variables under which your system has the best chance for functionality (can't blame them, it's reasonable - but it is telling)

I disagree.

1) We are already seeing the geofence areas get significantly bigger, going from just a 10 sq mi area to now doing entire metro areas. They are not tiny, tiny anymore. Waymo is operating fully driverless, 24/7 in the entire downtown Phoenix, Scottdale, Tempe and Chandler areas. That is not a tiny tiny area. Waymo is also doing 24/7 driverless in the entire SF area. That is not tiny tiny. And Waymo will be launching driverless 24/7 in a significant geofence in the LA area soon.

2) The cost is coming down and will continue to come down. Lidar used to cost $75k per unit. Now it costs less than $1k per unit.

3) Leave Cruise aside as they are not the AV leader. Waymo is doing fully driverless, 24/7, including in rush hour traffic. Waymo is also operating fully driverless in rain and fog. So Waymo is proving they can do reliable autonomous driving in wide and difficult conditions. They are not limited to just easy, cherry picked conditions. They are doing driverless in pretty much the same conditions as humans do.

The biggest challenge is achieving safe, reliable driverless that works in wide conditions. Once you do that, bringing down costs and scaling to more geolocations should be relatively straight forward. Waymo has safe, reliable autonomous driving that is fully driverless in suburban, dense urban, in rain and fog. So bringing down costs and scaling to more cities should be doable in a reasonable time-frame IMO.

I did an intervention-free, door-to-door drive in FSDb 11.3.6 this weekend from Orange County, CA to north San Diego County (~70 miles). It involved city streets with heavy traffic, entering and exiting the freeway, and traversing multiple construction zones. That's a first for me with FSDb, but I'm seeing marked improvements in a system that can generally handle all driving conditions (not specific ones that are easier). I also find that most disengagements currently are not safety-related, but preference related (i.e. I take over because I'm an aggressive driver and would prefer to be in the left lane going faster passing people, etc.).

The problem for FSD beta is getting to "driverless" level of disengagement rates for a non-geofenced area. Going 70 mi with zero interventions is actually super easy. To remove driver supervision you need like 10k miles with zero interventions at least. And FSD beta would need to achieve that intervention rate for the ENTIRE US, that is nearly 4M miles of roads in all conditions. Basically, FSD Beta needs to improve its intervention rate by at least a factor of 10,000 and do so for the entire US. There are huge differences in different places in the US (different roads, different rules, different traffic). Some FSD beta users do see good intervention rates in some areas, while others see worse intervention rates in other places. I still have safety critical inteventions with FSD Beta in my area. So actually getting to at least 10k miles per intervention in the entire US is a monumentally huge challenge. It is one reason why companies like Waymo start with geofences because it far easier to achieve 10k miles per intervention in a geofence area first and then scale the geofence area than it is to achieve 10k miles per intervention for the entire US first. In fact, I would argue that Tesla's challenge of achieving driverless everywhere is more difficult than Waymo's challenge of scaling because FSD Beta has more problems to solve.

 

[bkp_duke]

I disagree.

1) We are already seeing the geofence areas get significantly bigger, going from just a 10 sq mi area to now doing entire metro areas. They are not tiny, tiny anymore. Waymo is operating fully driverless, 24/7 in the entire downtown Phoenix, Scottdale, Tempe and Chandler areas. That is not a tiny tiny area. Waymo is also doing 24/7 driverless in the entire SF area. That is not tiny tiny. And Waymo will be launching driverless 24/7 in a significant geofence in the LA area soon.

2) The cost is coming down and will continue to come down. Lidar used to cost $75k per unit. Now it costs less than $1k per unit.

3) Leave Cruise aside as they are not the AV leader. Waymo is doing fully driverless, 24/7, including in rush hour traffic. Waymo is also operating fully driverless in rain and fog. So Waymo is proving they can do reliable autonomous driving in wide and difficult conditions. They are not limited to just easy, cherry picked conditions. They are doing driverless in pretty much the same conditions as humans do.

The biggest challenge is achieving safe, reliable driverless that works in wide conditions. Once you do that, bringing down costs and scaling to more geolocations should be relatively straight forward. Waymo has safe, reliable autonomous driving that is fully driverless in suburban, dense urban, in rain and fog. So bringing down costs and scaling to more cities should be doable in a reasonable time-frame IMO.



The problem for FSD beta is getting to "driverless" level of disengagement rates for a non-geofenced area. Going 70 mi with zero interventions is actually super easy. To remove driver supervision you need like 10k miles with zero interventions at least. And FSD beta would need to achieve that intervention rate for the ENTIRE US, that is nearly 4M miles of roads in all conditions. Basically, FSD Beta needs to improve its intervention rate by at least a factor of 10,000 and do so for the entire US. There are huge differences in different places in the US (different roads, different rules, different traffic). Some FSD beta users do see good intervention rates in some areas, while others see worse intervention rates in other places. I still have safety critical inteventions with FSD Beta in my area. So actually getting to at least 10k miles per intervention in the entire US is a monumentally huge challenge. It is one reason why companies like Waymo start with geofences because it far easier to achieve 10k miles per intervention in a geofence area first and then scale the geofence area than it is to achieve 10k miles per intervention for the entire US first. In fact, I would argue that Tesla's challenge of achieving driverless everywhere is more difficult than Waymo's challenge of scaling because FSD Beta has more problems to solve.

Bigger? Metro areas? TWO metro areas, SF and Phoenix. Not even the "hard" metros like LA and NYC. Those metro areas represent < 0.001% of the driveable roads in the USA. At this rate, the entire USA might have full coverage by . . . (does back of napkin math) . . . 2150. Again, NOT a scalable solution.

Meanwhile, Tesla is collecting as much data from their fleet in a few HOURS as Waymo and Cruise collect in . . . 1-2 years.


Lidar costs - still will always be FAR FAR more expensive than Tesla's cameras. MobileEye announced that their expected pricing to OEMs in 2025 for their hardware will be $6000. Sorry, but FSD hardware now costs a fraction of that, and has since 2021. I know a C-suite exec at a Lidar company, have known him since college, and even he admits that cost is a hurdle they are not sure they can overcome to get to Tesla levels.

What you and people looking at these solutions fail to see with FSDb is that we are already in the "march of the 9s" progression. The limiting factor for FSDb is AI training time, and that's accelerating. The intervention rate is arguably 10X less than it was just 6 months ago. That's an exponential progression, whereas Waymo and Cruise are on linear progression curves.

 

[EVNow]

In case you’re still looking for that edit button and you’re on a phone, go landscape.

Annoyingly it’s missing in portrait but landscape is fine.

TMC app on Android is fixed to portrait!

 

[diplomat33]

What you and people looking at these solutions fail to see with FSDb is that we are already in the "march of the 9s" progression. The limiting factor for FSDb is AI training time, and that's accelerating. The intervention rate is arguably 10X less than it was just 6 months ago. That's an exponential progression, whereas Waymo and Cruise are on linear progression curves.

And what I am trying to tell you is that FSD Beta is still at the very early beginning of the "march of 9's". FSD Beta still has at least 5 9's to go. So, FSD beta needs to improve at least 10,000 times. So FSD Beta is still very very far from driverless.

 

[bkp_duke]

And what I am trying to tell you is that FSD Beta is still at the very early beginning of the "march of 9's". FSD Beta still has at least 5 9's to go. So, FSD beta needs to improve at least 10,000 times. So FSD Beta is still very very far from driverless.

Fortunately, it's not marketed or sold as "driverless".

Still my intervention-free rides in SoCal say you are missing the bigger point.

 

[diplomat33]

Fortunately, it's not marketed or sold as "driverless".

Still my intervention-free rides in SoCal say you are missing the bigger point.

I don't care about "no intervention" rides. I care about "no supervision" rides. Until, FSD Beta is "no supervision" rides, it has not achieved what I want.

And I am happy for your intervention free drives. But I don't have intervention free drives where I live. So it is still not intervention free for me yet. And that's bigger point: FSD Beta needs to be intervention free for everybody, not just a lucky few who live in SoCal.

 

[bkp_duke]

I don't care about "no intervention" rides. I care about "no supervision" rides. Until, FSD Beta is "no supervision" rides, it has not achieved what I want.

And I am happy for your intervention free drives. But I don't even have intervention free drives where I live. So it is still not intervention free for me yet.

What you want isn't what is advertised. And supervision-free rides in Waymo and Cruise are of such short distances, at such low speeds, in such restricted areas, that their practical "value" is minimal. I'll take FSDb over that any day, as it provides me substantially more value on a day to day basis, even in the current form.

 

[Baumisch]

I'm confused as why there's no consensus among us - but hey - the consensus seams to be it's either Waymo or Tesla or some chinese company as a wildcard?
So I got GOOG and TSLA covered for the 10x uptick when FSD hits the valuation - correct?

 

[InTheMiddle]

Preface: Pretty sure this will result in 2 disagrees rather than a meeting-in-the-middle, but here goes...

Is it possible you are both right in some respects because what's important to each of you is different?

Conjecture: Tesla FSDb is the best ADAS on the market by leaps and bounds over the competition, and headed towards a high number of "no intervention" drive-assists, while Waymo (followed by Cruise) is the best Robotaxi on the market today (and for the next X years) <Each of you can put in whatever value of X you'd like, as I'm quite certain that each of us would put a different value in there.>

Potentially for at least those <X> years, Tesla FSDb continues to be the only realistic and ever-improving option for those who want to *own* cars (myself included). Each year, my Teslas with FSDb (and eventually dropping the 'b') will take over the march-of-9s of the driving task from me, but with me continuing to retain responsibility and having to supervise it, and that is a trade-off I'm happy to make in order to *own* my car. Even when the rides are vast-majority-intervention-free, Tesla may not (shouldn't for some time, IMHO) take responsibility for those times when intervention is needed, and it may be quite some time before Tesla wants responsibility for a human to be taken from point A to point B without a licensed driver in the car. But people will still buy Teslas in ever-increasing numbers expressly because of FSDb (on top of all the other things that makes Teslas amazing).

Potentially for at least those <X> years, Waymo continues to be the ever-expanding Robotaxi option for those who want a supervision-free ride in a car they do not own themselves. Even if Waymo is amazing, bottom line is that I cannot *buy* a Waymo vehicle of my own, that is dedicated to me, keeps my stuff in it, etc. Fortunately for Waymo, not everyone wants to own a car. Waymo can certainly displace human-driven taxis for high-taxi-ride-dense areas as they expand coverage (expanding existing coverage w/in a market and expanding to additional markets). They may be slowed in their scaling by the high cost of each vehicle and the challenges of entering each new market, but if they can make it profitable (a *huge* 'if', to be sure, but there are ways, especially with advertising) then I could see it displacing taxi fleets areas of dense usage, relegating human-driven rideshares to only rides outside the Waymo Robotaxi-served areas. Perhaps they also displace human-driven city busses running fixed routes even beyond their Robotaxi-served areas.

It seems to me that the real question is something like, "After those <X> years, will the world be dominated by A) Teslas running FSD w/o supervision or B) Waymo/Cruise Robotaxis eliminating car ownership?" Frankly, IMHO, <X> is a large enough number of years that I don't think any of our crystal balls show that picture. It isn't just the technology, it's all of humanity, it's the pace and curve of culture. I cannot imagine not wanting to own a car, but for many (especially those younger than me), they do not understand my desire to own a car.

TLDR: I can understand how Waymo is of far more interest / seems far more promising to someone who wants to get from their point As to their point Bs (within geos Waymo covers or is likely to cover soon) without them having to supervise it. I can understand how Tesla FSDb is of far more interest / seems far more promising to someone who wants to get from their (different) point As to their (different) point Bs and doesn't mind supervising as needed in exchange for A) owning the car and B) having points A and B which are at least sometimes outside the areas that Waymo covers. I don't see Waymo covering everywhere and offering cars for sale (not pay-per-ride) anytime soon, nor do I see Tesla assuming legal responsibility and hiring a huge call center of people monitoring my car to be able to remote-control if something goes wrong anytime soon. As of right now, they are different products serving different markets. Very good arguments have been made here for why each is the right solution for what the poster considers important, but hey, different things are important to different people. And that's OK.

 

[EVNow]

So far even geofenced system is not all that impressive compared to FSDb as we see them taking easy routes, slams on the brakes, getting stuck, and giving passengers heart attacks as it swerve into pedestrians as it tries to avoid other pedestrians.

This is the funny part - this kind of FUD is not very different from saying all EVs are terrible, get you stranded on the streets and catch fire all the time (isn't that what you read in the media ?!).

Here at TMC - we don't just go by headlines or accounts of one or two people. We investigate and learn more. Need to do that for all new tech ... there will always be a lot of FUD until the tech is mature.

 

[EVNow]

So I got GOOG and TSLA covered for the 10x uptick when FSD hits the valuation - correct?

If and when FSD is achieved - it won't happen overnight. It is a slow and not even steady process.

More importantly it is not at all clear robotaxis are all that profitable.

More likely multiple players will get to the "finishing line" of the marathon not very distant from each other - so no one will have the monopoly. That will reduce profits further.

 

[Singuy]

This is the funny part - this kind of FUD is not very different from saying all EVs are terrible, get you stranded on the streets and catch fire all the time (isn't that what you read in the media ?!).

Here at TMC - we don't just go by headlines or accounts of one or two people. We investigate and learn more. Need to do that for all new tech ... there will always be a lot of FUD until the tech is mature.

I was comparing L4 robotaxies to FSD and its performance despite having solved localization and a decade focusing on very small areas, they still make similar mistakes which doesn't make them all that impressive, or self driving is just that difficult. I didnt say it doesn't work or it's dangerous although I don't think a Tesla will ever hit a bus in a straight line. Maybe hit a bus doing a roundabout or a turn, but not a bus right in front of it.

 

[EVNow]

I was comparing L4 robotaxies to FSD and its performance despite having solved localization and a decade focusing on very small areas, they still make similar mistakes which doesn't make them all that impressive, or self driving is just that difficult.

I don't think they make similar mistakes. Their disengagement rates are in the range of 1 in 10k miles and Tesla's is 1 in 10 miles.

I don't even try to compare Robotaxis with FSD. They have a completely different business model (which doesn't seem to work at present). Tesla's FSD aim is to offer good ADAS and potentially L3 level at some point on consumer cars - so I compare that to MobilEye and other competitors.

 

[bkp_duke]

I don't think they make similar mistakes. Their disengagement rates are in the range of 1 in 10k miles and Tesla's is 1 in 10 miles.

I don't even try to compare Robotaxis with FSD. They have a completely different business model (which doesn't seem to work at present). Tesla's FSD aim is to offer good ADAS and potentially L3 level at some point on consumer cars - so I compare that to MobilEye and other competitors.

Where is the data supporting this 1 in 10k disengagement rate? Seems to be pulled out of thin air. The YT videos I've watched certainly don't support that assertion, not even remotely close.

 

[Jeff N]

Certainly, there have been some stalls with Cruise and Waymo that got a lot of media attention but they are rare compared to the total driverless miles. To argue that the stalls are happening all the time and that the rides are terrible with constant phantom braking is complete hyperbole IMO.

One things I have noticed on Cruise is that I’m averaging almost one apparent remote intervention per ride in SF. As in, the car will get stuck or confused and the passenger info screen will say “we’re helping to move your car…”.

Often this will be only momentary and without the notice on screen a passenger would probably be unaware. Other times the car turns on the emergency blinkers and it can take 15-30 seconds before things are corrected.

Recent examples include confusion at an intersection with non-functioning signal lights, getting stuck too close behind a double-parked truck and needing to back up, and being wedged at an intersection with narrow streets and parallel parked cars when we were trying to turn left and another car suddenly appeared and tried to turn right. Another time, the car appeared over-eager to bypass a stopped vehicle ahead of it only to find itself in the middle “suicide lane” adjacent to a hospital emergency entrance and blocked by a cement road divider ahead of it. It’s possible that it was spooked by the ambulances nearby.

Deployment of the emergency lights typically results in honking and yelling from other drivers which is awkward and embarrassing when you are sitting alone in the back seat…. So far these remote interventions have been automatic without my having to press the “help” button to call the support operator. In fact, a couple of times a support operator has spontaneously called into the car while the remote intervention is happening.

I wonder if all of these apparent remote interventions are from humans or if some simpler cases are handled automatically by a bigger and smarter computer back at HQ. I assume that over time the car software will be improved so that fewer remote (human) interventions are required so that this is not a long-term scaling issue.

I have never been part of a truly stuck situation requiring manual human intervention on scene.

 

[bkp_duke]

Preface: Pretty sure this will result in 2 disagrees rather than a meeting-in-the-middle, but here goes...

Is it possible you are both right in some respects because what's important to each of you is different?

Conjecture: Tesla FSDb is the best ADAS on the market by leaps and bounds over the competition, and headed towards a high number of "no intervention" drive-assists, while Waymo (followed by Cruise) is the best Robotaxi on the market today (and for the next X years) <Each of you can put in whatever value of X you'd like, as I'm quite certain that each of us would put a different value in there.>

Potentially for at least those <X> years, Tesla FSDb continues to be the only realistic and ever-improving option for those who want to *own* cars (myself included). Each year, my Teslas with FSDb (and eventually dropping the 'b') will take over the march-of-9s of the driving task from me, but with me continuing to retain responsibility and having to supervise it, and that is a trade-off I'm happy to make in order to *own* my car. Even when the rides are vast-majority-intervention-free, Tesla may not (shouldn't for some time, IMHO) take responsibility for those times when intervention is needed, and it may be quite some time before Tesla wants responsibility for a human to be taken from point A to point B without a licensed driver in the car. But people will still buy Teslas in ever-increasing numbers expressly because of FSDb (on top of all the other things that makes Teslas amazing).

Potentially for at least those <X> years, Waymo continues to be the ever-expanding Robotaxi option for those who want a supervision-free ride in a car they do not own themselves. Even if Waymo is amazing, bottom line is that I cannot *buy* a Waymo vehicle of my own, that is dedicated to me, keeps my stuff in it, etc. Fortunately for Waymo, not everyone wants to own a car. Waymo can certainly displace human-driven taxis for high-taxi-ride-dense areas as they expand coverage (expanding existing coverage w/in a market and expanding to additional markets). They may be slowed in their scaling by the high cost of each vehicle and the challenges of entering each new market, but if they can make it profitable (a *huge* 'if', to be sure, but there are ways, especially with advertising) then I could see it displacing taxi fleets areas of dense usage, relegating human-driven rideshares to only rides outside the Waymo Robotaxi-served areas. Perhaps they also displace human-driven city busses running fixed routes even beyond their Robotaxi-served areas.

It seems to me that the real question is something like, "After those <X> years, will the world be dominated by A) Teslas running FSD w/o supervision or B) Waymo/Cruise Robotaxis eliminating car ownership?" Frankly, IMHO, <X> is a large enough number of years that I don't think any of our crystal balls show that picture. It isn't just the technology, it's all of humanity, it's the pace and curve of culture. I cannot imagine not wanting to own a car, but for many (especially those younger than me), they do not understand my desire to own a car.

TLDR: I can understand how Waymo is of far more interest / seems far more promising to someone who wants to get from their point As to their point Bs (within geos Waymo covers or is likely to cover soon) without them having to supervise it. I can understand how Tesla FSDb is of far more interest / seems far more promising to someone who wants to get from their (different) point As to their (different) point Bs and doesn't mind supervising as needed in exchange for A) owning the car and B) having points A and B which are at least sometimes outside the areas that Waymo covers. I don't see Waymo covering everywhere and offering cars for sale (not pay-per-ride) anytime soon, nor do I see Tesla assuming legal responsibility and hiring a huge call center of people monitoring my car to be able to remote-control if something goes wrong anytime soon. As of right now, they are different products serving different markets. Very good arguments have been made here for why each is the right solution for what the poster considers important, but hey, different things are important to different people. And that's OK.

Not a disagree, I just don't see Waymo's solution as either cost-effective or scalable long-term. It's more a proof-of-concept demonstration (which is at least better than what MobileEye has done).

Tesla's solution, while funded by FSDb early adopters, has the potential to properly scale up and through L2, L3, L4, and L5, while remaining cost-effective from a hardware standpoint.

 

[EVNow]

Where is the data supporting this 1 in 10k disengagement rate? Seems to be pulled out of thin air. The YT videos I've watched certainly don't support that assertion, not even remotely close.

CA reports.

 

[diplomat33]

I wonder if all of these apparent remote interventions are from humans or if some simpler cases are handled automatically by a bigger and smarter computer back at HQ. I assume that over time the car software will be improved so that fewer remote (human) interventions are required so that this is not a long-term scaling issue.

I am not aware of Cruise using a big computer to do remote assistance but it is possible of course. I do know that they have human remote assistance that will give the car a hint but do not remote control the cars.

I can think of a few possibilities:
1) you might have a case that momentarily confuses the car but the car itself is able to solve it a few seconds later on its own.
2) Other cases might be more complicated so the car gets stuck, a human back in HQ gives the car a "hint" and with the hint, the car is able to solve it a few moments later.
3) And then you have the more serious cases where the car gets completely stuck or has some hardware malfunction like a computer crash and the human back in HQ is not able to solve it remotely, so the car requires someone to physically come and manually drive the car away.

And yes, these stalls will be reduced over time as the software gets better. Cruise is continually training their stack to be better at prediction and planning so that it can solve these edge cases on its own.

 

[bkp_duke]

CA reports.

Source please, not just the graph. I want to dig into how that is reported, b/c something smells off there. That has not been my personal experience (in the last 3 months) with Cruise in SF.