FSD discussion

[InTheMiddle]

How anyone can deny that the rate of progress since single stack hasn't accelerated materially just dumbfounds me. And we're already on the doorstep of another big release coming out. Past 6 months, but especially the past 3 months have significantly made me more bullish for FSD adoption rate and revenue in the near term (next 12 months). Before I was thinking 2025 was the more realistic timeline for mass FSD adoption rate.

In my experience, the switch to single-stack was a significant positive change, and FSD has improved since then. However, I would disagree with the characterization "that the rate of progress since then has accelerated materially". In my >100 miles of driving today, there were far too many errors that remain uncorrected vs the fewer areas I can say have definitely been improved since single-stack to believe the rate of progress has accelerated. Unquestionably YES, FSD *is* improving every release, just not necessarily at a materially increasing rate. It is still incredible what FSD can do - it's so many orders of magnitude better than any other commercial option available to me, with an ever-growing lead. But as an investor, my own expectation for any 'mass FSD adoption rate' which might impact Tesla's financials materially isn't until much later this decade (depending on how one defines 'mass FSD adoption rate', of course).

Spoiler: Sample of today's FSD failures

Many of today's failures were the 'changing out of the correct lane into the incorrect lane' type - for example, surface street with 2 lanes in each direction and a shared center turn lane, ego is in the left southbound lane with 0.4 miles before the upcoming left turn at a light (where the shared center turn lane becomes a dedicated left turn lane). The right blinker kicks on and the screen shows "Changing lanes to follow route" (not changing lanes to try to speed around traffic or anything)...it thinks it needs to change lanes *right* rather than staying in the left lane and being able to get into the dedicated left turn lane that is about to appear. It did this multiple times today, but the time with 0.4 miles to go before the upcoming left-at-a-light was the most egregious. Also, note that the setting for "Minimal lane changes this trip" does not reduce these errors, as this isn't a lane change to a faster lane, etc...it thinks it *needs* to change lanes to follow the route, even when in these instances if I allow it to change into the wrong lane, it frequently cannot get back left in time to make the turn, resulting in either a recalculation for a new route or a risky dive towards the left turn lane in a manner which leads me to take control again.

Another particularly poorly-handled incident today was heading eastbound on E Lincoln Dr approaching N 24th St, with the navigation planning a right hand turn onto N 24th St southbound. The dedicated right turn lane breaks off before the intersection, and the in-car map shows this; however, ego remains eastbound on E Lincoln Dr until after the gore-zone separator appears, then tries to cross the gore-zone separator until I yanked control back over.

Google Maps

Find local businesses, view maps and get driving directions in Google Maps.

www.google.com

 

[Carl Raymond]

Suddenly Gen 3 platform makes a whole lot of sense. If I’m repeating something below that Lodger has said six times, sorry. It’s getting real now.

The day before robotaxi legally approved, you don’t want to make too many, they‘re dead inventory, a maintenance overhead. You can’t know for certain at that point that approval is imminent.
The day after, you want to be making as many as possible, on fully ramped lines with a fully ramped supply chain.
How is this possible? By having a common design Model 2 come Robotaxi. Subtraction-only to switch from one to the next.
Giga Mexico becomes a gold mine.

 

[Knightshade]

....what?

The "wait for approval" thing remains a red herring talking point.

You can legally deploy Robotaxis today in at least a half dozen US states, covering a fair bit of the US population, without needing any magic "convince the regulators" type approval at all. In some cases you need do nothing at all besides have insurance. In a few more you basically just file a form that says "Our robotaxis are safe. We promise. And we have insurance."

Whenever Tesla decides their system is actually safe enough for RT use they need not wait on anybody but themselves to actually enable that.

 

[joh01652]

If, like me, you dismissed FSD for failing on roundabouts, I recommend watching this. My faith has been restored.
In the vid it expertly handles other drivers driving badly. Two who barge in. One who changes exit plan mid-way around. Entry is assertive when clear, yielding when not. Crosswalks also handled perfectly, no hesitation when clear.
Robotaxi. Is. Coming. OMG.

Yeah 11.3.6 has really been a step-change in improvement for me. I didn't have to touch the brake or steering once in a ~30mi trip from Flagstaff, AZ to Sedona, AZ. That has never happened until now, not even close. The route takes me from congested residential streets, to a busy four-lane 40mph road, to entering an interstate for two miles at about 70 mph, to exiting interstate to a state highway with winding roads and roundabouts. I have never done this drive without multiple disengagements--the only interventions were on the accelerator to prevent impatient tailgaters from rear-ending me because of non-human-like overly-cautious behavior from FSD.

And now it can handle all sharp turns on every marked road--prior to 11.3.6 it would often cross center lines and get "confused" and fail to continue--it doesn't do that anymore either.

Also with 11.3.6, I have yet to witness the spastic herki-jerkiness of the steering wheel, and lurching stops-starts as it struggles to make unprotected left turns, which for me personally was one of the most disconcerting behaviors because I had no idea if the car was going to bolt into oncoming traffic. The sudden, unpredictable jerkiness is something I won't tolerate, it feels so unsafe and does it at the most inopportune times in heavy traffic. It hasn't done that since (not that I've had a lot of time to play with it)--I'm hoping that characteristic is gone for good. It is so much smoother now.

Even my partner is surprised by the progress. Prior to 11.3.6 she didn't like it when I used FSD, but now doesn't mind, even displays mild curiosity and enthusiasm as she's not afraid anymore.

It's not perfect by any means, but yeah my faith has definitely been bolstered as well. Definitely getting better!

 

[Singuy]

In my experience, the switch to single-stack was a significant positive change, and FSD has improved since then. However, I would disagree with the characterization "that the rate of progress since then has accelerated materially". In my >100 miles of driving today, there were far too many errors that remain uncorrected vs the fewer areas I can say have definitely been improved since single-stack to believe the rate of progress has accelerated. Unquestionably YES, FSD *is* improving every release, just not necessarily at a materially increasing rate. It is still incredible what FSD can do - it's so many orders of magnitude better than any other commercial option available to me, with an ever-growing lead. But as an investor, my own expectation for any 'mass FSD adoption rate' which might impact Tesla's financials materially isn't until much later this decade (depending on how one defines 'mass FSD adoption rate', of course).

Spoiler: Sample of today's FSD failures

Many of today's failures were the 'changing out of the correct lane into the incorrect lane' type - for example, surface street with 2 lanes in each direction and a shared center turn lane, ego is in the left southbound lane with 0.4 miles before the upcoming left turn at a light (where the shared center turn lane becomes a dedicated left turn lane). The right blinker kicks on and the screen shows "Changing lanes to follow route" (not changing lanes to try to speed around traffic or anything)...it thinks it needs to change lanes *right* rather than staying in the left lane and being able to get into the dedicated left turn lane that is about to appear. It did this multiple times today, but the time with 0.4 miles to go before the upcoming left-at-a-light was the most egregious. Also, note that the setting for "Minimal lane changes this trip" does not reduce these errors, as this isn't a lane change to a faster lane, etc...it thinks it *needs* to change lanes to follow the route, even when in these instances if I allow it to change into the wrong lane, it frequently cannot get back left in time to make the turn, resulting in either a recalculation for a new route or a risky dive towards the left turn lane in a manner which leads me to take control again.

Another particularly poorly-handled incident today was heading eastbound on E Lincoln Dr approaching N 24th St, with the navigation planning a right hand turn onto N 24th St southbound. The dedicated right turn lane breaks off before the intersection, and the in-car map shows this; however, ego remains eastbound on E Lincoln Dr until after the gore-zone separator appears, then tries to cross the gore-zone separator until I yanked control back over.

Google Maps

Find local businesses, view maps and get driving directions in Google Maps.

www.google.com

"Rate of progress since then has accelerated materially" is referring to the rate of improvement since V11's wide release. You have not gotten any additional updates since the V11 wide release version of 11.3.6 so the strawman argument you have put forth is invalid. This is merely referring to V11.4's update notes and V11.4.1 update notes being pretty substantial even though development time seems to be relatively quick.

 

[InTheMiddle]

"Rate of progress since then has accelerated materially" is referring to the rate of improvement since V11's wide release. You have not gotten any additional updates since the V11 wide release version of 11.3.6 so the strawman argument you have put forth is invalid. This is merely referring to V11.4's update notes and V11.4.1 update notes being pretty substantial even though development time seems to be relatively quick.

Yes, it is true that I haven't received any additional updates since 11.3.6 (v2022.45.15) - all 3 of my cars are on that still to this day! Nice 'strawman' of picking the very latest version as the baseline! Of course, each of those 3 were previously on prior V11 releases, and I guess I can ignore the improvement on Cave Creek that the 11.3.6 point-release brought over the first V11 release downloaded to my X. Certainly better for me to stop basing progress rates on real-world first-hand experience and instead make investment decisions based on assumptions of how the 11.4 release notes and an online video will translate into future real driving experiences. /s

In all seriousness, if the original poster was referring merely to a hypothetical / potential rate of progress increase based on release notes and watching a video online, then my prior post and this can be deleted. My assumption was that their claim was made based on an actual observed rate of progress increase from the first publicly-released V11 builds, and I would refute that based on my experiences. I would also caution that point-releases have been known to be hyped more than the actual build justifies each time, and it is unfortunate IMHO when a release with very good updates is overshadowed by the even-greater-than-possible expectations some set for the release prior to each of us being able to experience it and put it thru a varied range of paces.

 

[Buckminster]

Those release notes are impressive. The % improvement numbers are significant.

 

[heltok]

11.4.1 notes release. Apparently even .1 releases come with all new notes.

https://twitter.com/x/status/1656396186532839424

Imo it's under-appreciated how Tesla has been able to rapidly grow, while investing a rather large amount into improving the speed of future development. This is an investment that has a longer term payoff that we now see the results of. FSD development has been suffering because Tesla have been busy doing rewrites that once they were done did make developing faster than the previous system. Going from image space to BEV was one such step, Autolabeler another, multipass autolabeler another. Dojo and 10x the compute/year is another. All these have been very hard to do and has taken time and investments to do.

Now with a lot of hard work done, they are in a place where they can make these great improvements rapidly, when their engineers can pick and choose any part of the system and improve it without breaking the rest of the system, where an engineer can just request a few thousand new clips fully labelled, add it to the training data and see performance improve. Where after a few weeks they can add multiple improvements, some that required 80k clips of 30s video data, 24fps, by 8 cameras, with 150 labels per frame, where they can improve a part of the autolabeler and rerun it on millions of clips and see performance go up. These are insane numbers that we now take them for granted, but it took a lot of work and vision to get here...

 

[Buckminster]

Summon update coming in the summer:

https://twitter.com/x/status/1656602293457010688

 

[Mengy]

Suddenly Gen 3 platform makes a whole lot of sense. If I’m repeating something below that Lodger has said six times, sorry. It’s getting real now.

The day before robotaxi legally approved, you don’t want to make too many, they‘re dead inventory, a maintenance overhead. You can’t know for certain at that point that approval is imminent.
The day after, you want to be making as many as possible, on fully ramped lines with a fully ramped supply chain.
How is this possible? By having a common design Model 2 come Robotaxi. Subtraction-only to switch from one to the next.
Giga Mexico becomes a gold mine.

This is the way.

 

[jerry33]

....what?

The "wait for approval" thing remains a red herring talking point.

You can legally deploy Robotaxis today in at least a half dozen US states, covering a fair bit of the US population, without needing any magic "convince the regulators" type approval at all. In some cases you need do nothing at all besides have insurance. In a few more you basically just file a form that says "Our robotaxis are safe. We promise. And we have insurance."

Whenever Tesla decides their system is actually safe enough for RT use they need not wait on anybody but themselves to actually enable that.

Regulatory approval has more to do with the EU than the US.

 

[Singuy]

....what?

The "wait for approval" thing remains a red herring talking point.

You can legally deploy Robotaxis today in at least a half dozen US states, covering a fair bit of the US population, without needing any magic "convince the regulators" type approval at all. In some cases you need do nothing at all besides have insurance. In a few more you basically just file a form that says "Our robotaxis are safe. We promise. And we have insurance."

Whenever Tesla decides their system is actually safe enough for RT use they need not wait on anybody but themselves to actually enable that.

Tesla's system doesn't have geo fencing and they have no intention of geo fencing as that only cost Tesla in operating cost as it's expensive to constantly monitor, update, and enforce. Since teslas are in the hands of customers and not Tesla, there's no stopping someone using it across state borders. So even though half a dozen states auto approved robotaxies, Tesla need all states to approve it before going live.

 

[rikkert]

Tesla's system doesn't have geo fencing and they have no intention of geo fencing as that only cost Tesla in operating cost as it's expensive to constantly monitor, update, and enforce. Since teslas are in the hands of customers and not Tesla, there's no stopping someone using it across state borders. So even though half a dozen states auto approved robotaxies, Tesla need all states to approve it before going live.

Seems to me that geo fencing would be fairly easy to implement, just monitor GPS location and disable FSD outside fence borders.

 

[Singuy]

Seems to me that geo fencing would be fairly easy to implement, just monitor GPS location and disable FSD outside fence borders.

Given the amount of bugs that come from the map data, I'm sure people would have a hoot trying to use a robotaxi in cities bordering another state.

 

[Zero CO2]

Here is the question for all of our FSD beta testers who think the recent improvements are not meaningful or significant enough.

Are you willing to give FSDb up on your vehicle?
I know i am not ... it is already a valuable feature and has improved highway driving significantly... and will be buying/subscribing FSD on future Tesla vehicles...

Tesla will be able to capture significant monthly subscribers when they offer the free FSDb trial in the coming months.... no need for robotaxi to start generating some meaningful revenue

 

[Knightshade]

Tesla's system doesn't have geo fencing and they have no intention of geo fencing as that only cost Tesla in operating cost as it's expensive to constantly monitor, update, and enforce. Since teslas are in the hands of customers and not Tesla, there's no stopping someone using it across state borders. So even though half a dozen states auto approved robotaxies, Tesla need all states to approve it before going live.

Tesla absolutely geofences between countries for other features (china most noticably, but not exclusively-- another more relevant example might be that FSD/EAP in Europe is different from the FSD/EAP in the US due specifically to certain things being unregulated in the US but restricted in the EU).

It'd be trivial to do the same for robotaxi across state lines to only enable it where it's already legal- which again it is, today, in a bunch of em... (even easier really, since RT would require a destination input at the start it'd know before the trip even began if it was crossing into a state it wasn't allowed to operate or not).

So again, anytime you hear "RT can't be a thing until Tesla gets regulators to agree" you're hearing something that's flat out false unless they're discussing some specific location it's not ALREADY ok without that.

Teslas lack of an RT service today isn't anything to do with regulators, it's to do with Tesla not believing the system is safe enough to be used that way yet. The day that second bit changes they could turn it on in multiple states THAT day.

 

[Tony73]

Tesla absolutely geofences between countries for other features (china most noticably, but not exclusively-- another more relevant example might be that FSD/EAP in Europe is different from the FSD/EAP in the US due specifically to certain things being unregulated in the US but restricted in the EU).

It'd be trivial to do the same for robotaxi across state lines to only enable it where it's already legal- which again it is, today, in a bunch of em... (even easier really, since RT would require a destination input at the start it'd know before the trip even began if it was crossing into a state it wasn't allowed to operate or not).

So again, anytime you hear "RT can't be a thing until Tesla gets regulators to agree" you're hearing something that's flat out false unless they're discussing some specific location it's not ALREADY ok without that.

Teslas lack of an RT service today isn't anything to do with regulators, it's to do with Tesla not believing the system is safe enough to be used that way yet. The day that second bit changes they could turn it on in multiple states THAT day.

I doubt anyone including Tesla is assuming the liability of turning on a robotaxi service statewide in any state in the next decade.
Robotaxis will be geofenced for the foreseeable future.
Tesla needs to get FSD into wider adoption. That is pure profit and can happen in a much shorter timeframe than robotaxis

 

[InTheMiddle]

Here is the question for all of our FSD beta testers who think the recent improvements are not meaningful or significant enough.

Are you willing to give FSDb up on your vehicle?
I know i am not ... it is already a valuable feature and has improved highway driving significantly... and will be buying/subscribing FSD on future Tesla vehicles...

Tesla will be able to capture significant monthly subscribers when they offer the free FSDb trial in the coming months.... no need for robotaxi to start generating some meaningful revenue

I would absolutely *not* give up FSDb on my vehicles...and in fact, FSDb is one key reason (of several) for me to continue to buy Teslas instead of other EVs, because even in its current state FSDb is worth more to me than the $15,000 price tag.

Spoiler: Definitely meaningful and significant enough, just no Robotaxi around the corner

(To be clear, I am not one of those "who think the recent improvements are not meaningful or significant enough"...the improvements are definitely both meaningful and significant enough to keep me confident that progress will continue. But my bar for "enough" still slopes towards many, many years before RoboTaxi, and until then I don't think FSDb take-rate will materially increase until then, because I've come to accept that most other people won't see the current-capabilities-vs-current-cost as worth it (like I do) in the meantime. It might be far better for Tesla to market FSDb more as "the current FSDb builds are the best, most capable ADAS you can buy" and less as "pay us now for amazing robotaxi in the future" and actual take-rate may increase then, if more people can grow to appreciate it for what it actually is today.

 

[Singuy]

Tesla absolutely geofences between countries for other features (china most noticably, but not exclusively-- another more relevant example might be that FSD/EAP in Europe is different from the FSD/EAP in the US due specifically to certain things being unregulated in the US but restricted in the EU).

It'd be trivial to do the same for robotaxi across state lines to only enable it where it's already legal- which again it is, today, in a bunch of em... (even easier really, since RT would require a destination input at the start it'd know before the trip even began if it was crossing into a state it wasn't allowed to operate or not).

So again, anytime you hear "RT can't be a thing until Tesla gets regulators to agree" you're hearing something that's flat out false unless they're discussing some specific location it's not ALREADY ok without that.

Teslas lack of an RT service today isn't anything to do with regulators, it's to do with Tesla not believing the system is safe enough to be used that way yet. The day that second bit changes they could turn it on in multiple states THAT day.

There is a major difference between geo fencing haphazardly for a specific region vs geo fencing in higher definition for a robotaxi as that will start to get pretty detailed when any error can lead to traffic jams and stuck customers. This is a major liability and difficult issue especially without HD map. Right now Tesla can do whatever and the driver just shrug if you can't enable fsd at any particular area or enable fsd at an area that you are not supposed to and the system shuts down with "take over immediately" warning.

Looking at waymo and cruises earnings will tell you how daunting the operating cost can be from geofencing in high detail as a legitimate robotaxi service.

 

[Tony73]

There is a major difference between geo fencing haphazardly for a specific region vs geo fencing in higher definition for a robotaxi as that will start to get pretty detailed when any error can lead to traffic jams and stuck customers. This is a major liability and difficult issue especially without HD map. Right now Tesla can do whatever and the driver just shrug if you can't enable fsd at any particular area or enable fsd at an area that you are not supposed to and the system shuts down with "take over immediately" warning.

Looking at waymo and cruises earnings will tell you how daunting the operating cost can be from geofencing in high detail as a legitimate robotaxi service.

Geofencing is not a big deal. Cars have very accurate location information because they stick to roads for the most part, have high gain gps antennas and very accurate inertial sensors.
A geofenced robotaxi will only accept trips for which it can clearly ascertain that both end points and all transit points are within the geofenced area.
There are plenty of challenges with robotaxis but geofencing is not one of them.

The high operating cost of waymo and cruise has nothing to do with geofencing. If anything, geofencing is keeping operating cost down