FSD discussion

[Daniel in SD]

Disagree because this all depends on the contracts involved.

There is no contract between the person hit by an autonomous vehicle and the manufacturer.
Insurance is not even required for autonomous vehicles in California. You just have to show you have $5 million. Insurance does not change who is responsible for a collision, it only changes who pays (and only up to the liability limit).
I’m pretty sure Mercedes is liable for at fault collisions in the EU and California by regulation. I have no idea if they purchase insurance to cover that liability.

 

[Bet TSLA]

Of course, we are decades away from having fully-self-driving cars that are that good.

You say this as though it's a fact, but you are without facts. It's completely obvious that the advent of AGI would mean autonomous driving -- AI as smart as human plus a better sensorium plus better reflexes is obviously capable of driving as well as a human can. And nobody knows whether AGI is a tomorrow thing, a next week, next year, or next decade thing. Hardly anybody who knows much about the subject expect it to be much more than a decade or two.

So no, even in the extreme case of requiring AGI it could be here very soon. You certainly have no idea.

 

[daniel]

Insurance has nothing to do with whether or not a lawsuit is successful.

It does because the outcome of a case has far more to do with the skill and financial resources of the lawyers than with any facts of the case. Insurance companies have the resources to hire the best lawyers, pay for legal aids to do research, and pay for expert witnesses. And to appeal excessive awards.

It is my belief that at first other insurance companies will not be a good choice to insure Tesla vehicles with autonomous capability:
- either the contract will state that liability is not covered when no driver is behind the wheel, or you are always viewed "at fault" with regard to your insurer, meaning your premiums will be increased;
- or the insurance premiums would be insanely high since these companies won't "trust" Tesla FSD and they'll want to overcharge to minimize risk.

[...]

Side note: that's when in my mind Tesla cars will be fully autonomous, surpassing human level driving: the moment Tesla insurance covers you even without supervising the software/without anyone in the driver seat.

Insurance companies are experts at assessing risks. Once fully autonomous cars are actually safer than human drivers, they will lower their rates for them. And for the best autonomous systems, they'll lower their rates more.

This is NOT how the law is currently crafted, not nation-wide and not state by state. The liability still lies with the driver, 100%, and Tesla has VERY carefully crafted their EULA to make sure that users are the responsible party, not Tesla. Anyone using Autopilot or FSDb is literally saying "In the event of an accident, even if there might be concern that the system had a degree of fault, I accept FULL responsibility by using this system."

Apples and oranges. We're talking about fully driverless cars, once they are being sold to consumers. That's a very long way off. We all agree that today's Level 2 cars are the responsibility of the driver. And Level 3 is a grey area, since the driver still must be available to take over, so there's shared responsibility. That might make insuring Level 3 cars problematic.

You say this as though it's a fact, but you are without facts. It's completely obvious that the advent of AGI would mean autonomous driving -- AI as smart as human plus a better sensorium plus better reflexes is obviously capable of driving as well as a human can. And nobody knows whether AGI is a tomorrow thing, a next week, next year, or next decade thing. Hardly anybody who knows much about the subject expect it to be much more than a decade or two.

So no, even in the extreme case of requiring AGI it could be here very soon. You certainly have no idea.

We'll have to disagree on that. Nobody working in AI thinks we're anywhere near developing AGI. The notion that AGI might be a year away is an extreme fringe idea, not held by anybody working in the field. I would happily bet $100,000 that no Level 5 automobile will be widely available for purchase before the end of 2024. True AGI would, of course, make the development of a Level 5 car trivial.

 

[Buckminster]

https://twitter.com/x/status/1625745896808873985

11.3

 

[Buckminster]

https://twitter.com/x/status/1625745896808873985

11.3

https://twitter.com/x/status/1625858508691505156

 

[Bet TSLA]

We'll have to disagree on that. Nobody working in AI thinks we're anywhere near developing AGI. The notion that AGI might be a year away is an extreme fringe idea, not held by anybody working in the field. I would happily bet $100,000 that no Level 5 automobile will be widely available for purchase before the end of 2024. True AGI would, of course, make the development of a Level 5 car trivial.

Sadly you seem to not know the difference between stuff you believe for no apparent reason, unsupported by anything, and actual expert opinion. What you write seems to be based on ignorance and confusion.

Here's a small sample of what experts actually believe.

How far are we from artificial general intelligence (AGI)? The experts weigh in...

Are we close? Or, will it be beyond our lifetime?

aifuture.substack.com

There's nothing wrong with being ignorant. But why post here and display your ignorance for everybody? And, by the way, nobody said AGI is likely to come within the next couple of years. All I said is that we have no idea, and pretty much everybody who knows anything thinks that it will be here within a couple of decades. Note the prevalence of the 2029-30 timeframe in the referenced expectations for AGI, and the fact that various stepping stone predictions were achieved very early already.

 

[daniel]

... And nobody knows whether AGI is a tomorrow thing, a next week, next year, or next decade thing. ...

... And, by the way, nobody said AGI is likely to come within the next couple of years. ...

There are people in the field who think we're on the verge of AGI, and there are people in the field who think it's a very, very long way off. I find the latter's arguments more convincing.

 

[willow_hiller]

Interesting tidbit in the full recall report for FSD Beta: https://static.nhtsa.gov/odi/rcl/2023/RCLRPT-23V085-3451.PDF

"As of February 14, 2023, Tesla has identified 18 warranty claims, received between May 8, 2019, and September 12, 2022, that may be related to the conditions described above. Tesla is not aware of any injuries or deaths that may be related to such conditions."

So maybe 18 accidents in the first 3 years of FSD Beta, with no injuries or deaths.

 

[JRP3]

Perhaps the legal liability is put on the driver as the person who chooses the settings?

I was thinking more of Level 4 and 5 when no driver input is required. Will the car adhere to the set limit, and will people be annoyed if it does?

 

[Buckminster]

I haven't seen much discussion about greentheonly's reveal of whats involved in HW4.
I think its going to be very hard for Tesla to talk up HW4 without is becoming implicit that HW3 is infgerior, and very likely not capable of full FSD, or even FSD in circumstances outside of wholemars's daily commute...

If Tesla are to admit that HW3 can't do FSD, then they will have a lot of angry HW3 owners, and a lot of very angry FSD purchasers, given then have already said there is no upgrade path.

I think there is a big blindspot among tesla bull investors regarding this issue. The sensible thing would be for Tesla to either knuckle-down and work out how to provide a HW3-HW4 FREE upgrade for people who bought FSD, or to set aside funds to compensate people who very very clearly were sold 'FSD capable' vehicles which may not be. Its also TERRIBLE for PR either way, but worse if they just stick two fingers up to HW3 owners.

I'm a long term Tesla bull who bought an autopilot 1 car, then recently a HW3 model Y with FSD. Currently, my fancy £80,000 car has no parking sensors or equivlanet functionality, cannot autopark, and has trouble knowing when to turn its wipers on and off. Tesla REALLY need to get their act together regarding living up to their promises regarding TeslaVision, its pretty much the worse thing about the company right now...

Said as a long term investor and full supporter of the company for at least 8 years now.

 

[Buckminster]

This was already done at the last earnings call. And I think they squared that circle pretty sufficiently:

"Hardware 3 will not be as good as Hardware 4, but I'm confident that Hardware 3 will so far exceed the average -- the safety of the average human. So how do we get ultimately to -- let's say, for argument's sake, if Hardware 3 can be, say, 200% or 300% safer than human, Hardware 4 might be 500% or 600%. It will be Hardware 5 beyond that. But what really matters is are we improving the average safety on the road."

Yes, HW3 is inferior to HW4, but both will be safer than human drivers. If Tesla can deliver FSD with HW3, there are no broken promises and no need for an expensive retrofit.

 

[Buckminster]

FSD Rant On

There is a huge difference between 1) highly competent L2 FSD beta and 2) an actual robotaxi.

The difference is that an actually robotaxi needs to have an insanely high performance & accuracy, multiple of orders of magnitude better than a great L2 FSD beta.

HW3.0 will likely be able to deliver a highly competent L2 FSD, but it will not be able to deliver a robotaxi.

Why?

It can't see cross traffic well.

The cameras were never in the right position for detecting cross traffic when visibility is constrained.

The cameras don't have high enough resolution to detect high speed cross traffic with a high enough reliability from far enough away to basically never make a mistake.

These reasons alone are enough reason to forgot about thinking HW3.0 will be for robotaxis.

But that's okay.

Tesla will be making HW4.0 vehicles at such a rate that if indeed they get to robotaxi level, they'll be able to deploy hundreds of thousands per month.

Shed a tear for the HW3.0 owner who though they would have their car in the robotaxi fleet, but otherwise who cares?

HW4 complainers - HW4 exists because of inexorable semiconductor process improvements. Those happen every year. If you keep using the process tech from the initial year you started using your chip, you're leaving efficiency on the table and your competitors will leapfrog you. HW3 was revealed at the Autopilot Day in ~March 2019. It was designed in 2018 (Jim Keller and all that). That's 5 years ago! It is inevitable that Tesla would move to a new process tech, and I think they said so at the time.

One of the goals of a new process tech is to do the same job as the last chip, but for less power, generate less waste heat, and take up less room in the circuit board. HW4 does that, but they have added some new capability. Nobody should be saying HW3 can't do the job.

In 4-5yrs time there will be HW5. After that... HW6. etc. etc. Each time, the chips will use less power, occupy smaller space, generate less heat under the dash, and/or have more capability if Tesla decides to add that. Again... do not complain that HW3 has been usurped or is now proven to be unable to do the original job. That is nonsense. Tesla has revised the CPU in the MCU several times now. Can owners of the original 2012 MCU still run Navigation? Of course!

Having HW4 now makes it more believable that Cybertruck is going into production this year. It makes me happy as a future CT owner that they are locking another piece of the CT puzzle into solid production hardware. It's true to say that the HW4 unit will have had a little in-the-field testing on 4 or 5 other platforms before CT starts using it. Nice!

WTF does your last paragraph even mean? We arent asking them to stop anything, but they took anywhere from 4-15K from people for a feature that isn't fully baked. They already have a history of replacing hardware (HW2.5 to HW3) to implement these features. It may very well make everyone happy eventually but as of right now Tesla is implying the money people paid for a software they can't even recognize full revenue yet is going to be capped based on new hardware they are releasing (3X better than 10X from Elon)...I still fully expect Tesla to deploy a robotaxi in less than 24 months...which will evidently be based off HW4. As a shareholder...I couldn't be more excited, but as someone who paid them 10K for a product that now has questions to its end state...call me worried.

Nobody here is even suggesting Tesla stop innovating or slow down, actually quite far from it.

Have you read the details of the recall? Three out of the four items aren't bugs or faults, they're just differences in opinion between Tesla and the NHTSA on how FSD Beta should operate.

Here's the full list:


On item 1, FSD Beta is actually very proficient at estimating whether it has enough time to travel through an intersection on a yellow. But the NHTSA would prefer that Tesla comes to a stop on yellow instead of proceeding.

On item 2, other descriptions make it clear that the NHTSA thinks Tesla isn't stopping for a sufficient amount of time. They want the cars to wait longer.

On item 3, Tesla has optimized for passenger comfort over quickly matching set speeds. The NHTSA wants them to snap to the set speed faster.

All of those can be fixed almost immediately without any additional features to FSD Beta. The only one that might take some work is item 4, if they choose to fix it by solving the lane selection problem. As I said above, they might be able to fix it to the NHTSA's liking by simply never allowing FSD to attempt to proceed forward once it recognizes it's in a turn lane.

Super interesting point of view from Whole Mars Blog this evening which puts this NHTSA thing in perspective.



I hadn’t really thought about it this way, but while we’ve been looking at it from the glass half empty perspective, this really puts a different spin on it. If this is all NHTSA takes away from this, then this is a win.

Spoiler: Source

https://twitter.com/x/status/1626394348622577666

 

[heltok]

I asked chat GPT to come up with more improvements to v11.3:

1. Enhanced handling of construction zones and lane closures by utilizing real-time traffic information and construction site data to adjust FSD's planning and control algorithms accordingly.
2. Improved object recognition and response to pedestrian and cyclist traffic, especially in urban and suburban environments.
3. Integration of weather data and road surface conditions to improve FSD's response to changing weather and road conditions, such as rain, snow, ice, and potholes.
4. Improved routing and navigation based on real-time traffic conditions and historical data, taking into account factors such as accidents, road closures, and congestion.
5. Integration with smart city infrastructure and traffic management systems to enable FSD to optimize driving routes and reduce traffic congestion.
6. Improved communication and interaction with other vehicles on the road, including the ability to exchange information about road conditions, traffic flow, and potential hazards.
7. Enhanced safety features, such as improved collision avoidance systems and the ability to detect and respond to emergency vehicles, such as police cars, fire trucks, and ambulances.
8. Integration with advanced driver assistance systems (ADAS) to provide a seamless transition between manual and autonomous driving modes.
9. Enhanced user interface and driver feedback systems, including more intuitive controls and displays, and improved voice recognition and natural language processing.
10. Continued research and development of machine learning algorithms and artificial intelligence systems to improve FSD's performance and capabilities over time.

 

[replicant]

https://twitter.com/Winnersechelon posted lots of videos of FSD 11 today (starting here)

 

[MartinAustin]

This is a HUGELY underappreciated issue!

Its pretty clear to anybody with a Tesla in the UK, especially in the last few months, that the current suite of sensors is absolutely inadequate to provide FSD, even given an infinite amount of compute. It will probably be absolutely fine for California, I have no doubt, but so far this year, EVERY single day I've driven more than 10 miles in my new model Y, the car ends up with a horribly grimey reverse camera view, and its absolutely constant that I get warnings that autopilot cannot function because a side pillar camera or 'multiple cameras' have their view blocked or blinded.

There is no way this can be fixed by software. They have to have some kind of self-cleaning camera, and TBH its pretty ridiculous that tesla don't realize this. They should relocate the whole autopilot team to a state that gets constant bad weather, because right now, they are massively over-fitting for California sunshine.

The same problem is probably why they think USS can be replaced by vision only, or that USS isn't required to park. Not all weather is Californian. Not all roads are Californian width. I'd love to see Elon try and park a model X with no USS in a UK car parking space.

There are very few advantages other auto firms have over Tesla, but being aware of real world driving conditions is sadly one of them. I genuinely think this is a risk that investors underappreciate.

semi-off topic for this thread, but, I grew up in the UK and learned to drive there. One time in my first car, I was pulled over by the cops while not speeding, not doing anything untoward... they came up to the driver's side window and said "the reason we're pulling you over is because your rear license plate is so dirty no-one can read it... and that's an offence. clean it up!!!" It is a mucky life in the UK. Cars get covered in it... headlights get dimmed, windows are mucky etc. etc. - you simply have to clean your car often to use it to its proper spec. Cleaning computer cameras is no different. Heck, I even have to clean the rear camera on my pre-AP Model S. (oh, the humanity) There are even dirtier places than the UK. Are you saying that Tesla has to design a car for the India market (where they have smog pollution PLUS rain/mud) w/ external cameras that never need to be cleaned by the owner of the car?

 

[AndrewZ]

semi-off topic for this thread, but, I grew up in the UK and learned to drive there. One time in my first car, I was pulled over by the cops while not speeding, not doing anything untoward... they came up to the driver's side window and said "the reason we're pulling you over is because your rear license plate is so dirty no-one can read it... and that's an offence. clean it up!!!" It is a mucky life in the UK. Cars get covered in it... headlights get dimmed, windows are mucky etc. etc. - you simply have to clean your car often to use it to its proper spec. Cleaning computer cameras is no different. Heck, I even have to clean the rear camera on my pre-AP Model S. (oh, the humanity) There are even dirtier places than the UK. Are you saying that Tesla has to design a car for the India market (where they have smog pollution PLUS rain/mud) w/ external cameras that never need to be cleaned by the owner of the car?

How do you design and build a generalized autonomous vehicle without a steering wheel or pedals -- which Elon has mentioned several times -- if it needs to be cleaned by the owner of the car to function at all?

There are very good reasons most of the autonomous vehicle stuff right now is happening in arid desert climates like Nevada and Arizona. A generalized autonomous vehicle needs to be able to handle inclement weather, winter, etc, it needs to be capable of appropriate action after sensors have been disabled or the vehicle is hit by a more severe failure. The cameras need to be able to clean themselves or something and it needs backups, redundancies, and redundancies on top of redundancies not only in the sensors but in steering control motors and everything else. HW4 already looks to have more redundancies than HW3.

 

[EVNow]

Many interesting points - moved from market thread.

Absolutely right on the FSD before robots with full anthropomorphic features.

Of course the idea that there is any profit at all from Robotaxi is a complete guess- it is based on some assumption that Tesla can somehow capture and control a market (a very uncertain market) before others arrive. Since Waymo and others are going to be first launching in LA this year they've already lost the race to be first (LA is a top 5 ride sharing market- unlike Phoenix and San Fran) Waymo is partnered with Geely, an entity that can produce cars at volume, Tesla will have nothing to compete with except cost and perhaps a wider distribution.

Waymo "expanding" to LA is a misnomer. Its really more like testing - just like in SanFrancisco. Waymo had a grand average of 2 trips per day of driverless trips in SFO.

Ofcourse the question is - where will they be in 5 years, which Tesla may need to get to anywhere near robotaxi capability.

Waymo has not demonstrated any kind of agility or ability to scale up / out ... so, who knows.

You also have mobileye with a robust technology solution selling to OEMs

MobilEye apparently has good technology ... but no announced US OEMs. Thats a strange miss from a Intel company. They just have L3/L4 tie-ups with a small Chinese company now.

and Cruise that is actually pushing boundaries until things break (something interesting to observe in a company backed by GM-real silicon valley behavior and if GM were behaving like Cruise in auto production they'd be a fearsome competitor).

Cruise is really weird ... they keep getting stuck. I think cities will withdraw their license unless they can demonstrate reliable fall back options for stuck vehicles, as is required in L4.

I'm not even talking about the near run solutions like Zook - great robotaxi platform, or the myriad Chinese backed solutions. The point being that Tesla is not going to walk into a automated ride services business first not will they likely be 2nd or 3rd. Economic history shows that without some differentiator profits basically disappear. You hunt for a percent here or there, you don't get 500%.

This is definitely possible. Even Robotaxi vs Uber/Lyft is a competition that is yet to play out.

Could FSD have tremendous value as a solution to provide car owners with a chauffeur? I can absolutely see that. It could be a huge sales opportunity.

Yes - it will be a huge margin driver for cars or more likely a subscription model that brings in a lot of monthly revenue.

 

[ThisStockGood]

A little warm up to P&D tmr - this is great fun

TL;dw: Tesla FSD is not very nice /s

 

[Buckminster]

https://twitter.com/x/status/1642458136308224000

Let me know when 50%+ of you actually believe that Robotaxis will be a thing somewhere in the world pre 2026. At that point I will recommend that Superbulls need to believe in the Optimus Demi-God for entry tickets.

 

[EVNow]

Let me know when 50%+ of you actually believe that Robotaxis will be a thing somewhere in the world pre 2026. At that point I will recommend that Superbulls need to believe in the Optimus Demi-God for entry tickets.

Well, Robotaxis are already a reality in many cities - just not Teslas

I think to get to Robotaxi level Tesla has to
- Change camera positions / resolution. Currently it has to creep too far out that scares oncoming vehicles (and the Tesla driver).
- Figure out how to handle all the mapping issues. MobilEye has some good ideas ... where they use the fleet camera to enhance the map. Tesla should do something similar (they may be doing some of that according to @verygreen recent info).