Tesla, the investment

[White.Beard.Doc]

I have been a Tesla fan since 2010 when I ordered my first Model S. The stock has done very well for me over the past decade, but my enthusiasm is waning. I told a friend back in 2010 that it is inadvisable to bet against Musk's ability to execute. But my conviction about Musk is also waning.

Where does my current skepticism come from? The full self driving claims are just not credible. The cars can currently drive themselves well in simple situations, like on the limited access highways. So if we are at 75% self driving today, each stepwise increment towards 100% gets harder and harder to achieve. My current Teslas drive themselves using eyesight cameras, but their vision and depth perception is quite poor. The Tesla cannot even drive around a parked car on a local street if there is a car parked at the curb and the Tesla would have to cross the yellow center line to pass the parked car. This is a simple maneuver, yet the car does not have enough "brain power" to figure out what to do, and it does not have enough depth perception to know what spaces it can fit through.

On the plus side, the electric drive trains are awesome, and the touch screen interface between driver and vehicle is pretty slick. But the suspension and ride are subpar (although the ride on the S with air suspension is sweet, but not on the Y), and the full self driving is so far away from reality that it does not seem credible any time in the near future. When you drive a Tesla in self driving mode today, true robotaxis with their current technology feel like a bridge too far.

I continue long with my Tesla stock, just for fun, but I no longer feel all that optimistic about the upside from here.

[Nysoz]

Have you seen the fsd beta videos? It’s “orders of magnitude” better than what the consumers have.

still a long way from robot taxis but definitely making progress.

https://m.youtube.com/watch?v=J-EAaq6UATI

https://m.youtube.com/watch?v=dQG2IynmRf8

[Kamban]

Where does my current skepticism come from? The full self driving claims are just not credible. The cars can currently drive themselves well in simple situations, like on the limited access highways. So if we are at 75% self driving today, each stepwise increment towards 100% gets harder and harder to achieve. My current Teslas drive themselves using eyesight cameras, but their vision and depth perception is quite poor. The Tesla cannot even drive around a parked car on a local street if there is a car parked at the curb and the Tesla would have to cross the yellow center line to pass the parked car. This is a simple maneuver, yet the car does not have enough "brain power" to figure out what to do, and it does not have enough depth perception to know what spaces it can fit through.

On the plus side, the electric drive trains are awesome, and the touch screen interface between driver and vehicle is pretty slick. But the suspension and ride are subpar (although the ride on the S with air suspension is sweet, but not on the Y), and the full self driving is so far away from reality that it does not seem credible any time in the near future. When you drive a Tesla in self driving mode today, true robotaxis with their current technology feel like a bridge too far.

I think there is a max "FSD" limit that can be achieved with the current cameras and radar on Tesla with the current computer HW3 chip.

Waymo can achieve a higher magnitude of FSD because it has a bit more sharper cameras but more importantly, has Lidar.I think without Lidar Tesla cannot go much further, even with neural networks.

But Lidar mounted on top of the roof looks dorky. And Waymo has to reluctantly accept that its setup will cost close to a high end Mercedes - $100K in hardware / software alone. Add another 50-60K for the car and you are getting into supercar / real luxury car territory. Which few will want buy and few can afford it too.

Unless we get to pocket sized Lidar that can be mounted easily, true FSD will not become mainstream. And GM Cruise is not FSD.

[fatlittlepig]

I've been following the self driving technology over the past several years. I have watched the fsd beta videos recently posted, they are impressive to a point, but its abilities seem gimmicky in that it works pretty much OK (but no where nearly robustly as you will need) but you don't get a sense that it will be able to evolve to the point where it works and you can read a book or fall sleep. the hardware and sensor suite in it's current form seem to have a natural ceiling.

[Nysoz]

I don’t think it’ll ever be perfect. It just has to be better than the average human driver (which is pretty bad) to be a benefit and potentially save lives. The question is how much better before regulators allow it?

They're retraining the entire system now. Instead of labeling pictures they’re labeling video. What will help speed up the process and make it that much better is their dojo neural net supercomputer.

Humans are able to drive with 2 cameras and a brain. Will a car be able to do the same? I guess it depends on how well you can program AI to learn. AI can beat humans in complex tasks like Starcraft or go or chess. But all of those games are based off rules. Driving you have to know when it’s ok/necessary to break the rules and be prepared for anything like a plane landing on the road.

I agree with Tesla that Lidar and hd maps are crutches/training wheels for driving as humans can get by without those tools/that set of data. In some ways AI can be better than the average human as we get easily distracted by phones, texts, music, eating, applying makeup or whatever other crazy thing you’ve seen people do. AI can look using multiple cameras in all directions at all times while not getting distracted or fatigued.

I do think it’ll happen someday but with some limitations at first and definitely not on Elon’s time frame. It’ll also require municipalities to really maintain roads/lane markings to help these systems as much as possible in the future.

[Dont_know_mind]

I was looking at Arlo in 2019:

https://forum.whitecoatinvestor.com/general-investing/10400-arlo-the-stock

The way Tesla moved I thought the market had priced that they had found a likely solution to computer vision or something.

When I was researching Arlo I found this interesting computer vision video:

https://youtu.be/HKnY05QoSi0

It is not clear that deep learning/neural networks will be able to solve the computer vision problem.
Winter in the above video argues that there may be structural problems with deep learning.
This issue was raised by Yann LeCun and Gary Marcus in their AI debate in 2017.

It’s a fascinating area at the interface of philosophy of mind, psychology and computer science:

https://youtu.be/vdWPQ6iAkT4

[Tim]

“It’ll also require municipalities to really maintain roads/lane markings to help these systems as much as possible in the future.”

I really got a kick out of this one.Municipalities are currently really required to do this. I won’t hold my breath.

https://statutes.capitol.texas.gov/Docs/TN/htm/TN.251.htm

[Kamban]

I don’t think it’ll ever be perfect. It just has to be better than the average human driver (which is pretty bad) to be a benefit and potentially save lives. The question is how much better before regulators allow it?
.

A big elephant in the room is assignment of responsibility. In many accidents there is no clear cut way to find the responsible party and it becomes he said. she said. Unless you have all cameras and reconstruct the events you cannot assign blame and sometimes even after that it is still difficult ( two people simultaneously coming to middle lane from outer and inner lanes and both indicating)

The bigger issue is that FSD might be better, but will still make mistakes. So who do you give the ticket to - the person in the car reading that book or the maker of the car or the maker of the FSD software / hardware like Waymo?

[Zaphod]

It’s obviously coming, just not via Tesla.

And it will just take more time than people who are pumping it say, its basically true of all tech, its coming or possible, but just takes time.

[Zaphod]

I don’t think it’ll ever be perfect. It just has to be better than the average human driver (which is pretty bad) to be a benefit and potentially save lives. The question is how much better before regulators allow it?

They're retraining the entire system now. Instead of labeling pictures they’re labeling video. What will help speed up the process and make it that much better is their dojo neural net supercomputer.

Humans are able to drive with 2 cameras and a brain. Will a car be able to do the same? I guess it depends on how well you can program AI to learn. AI can beat humans in complex tasks like Starcraft or go or chess. But all of those games are based off rules. Driving you have to know when it’s ok/necessary to break the rules and be prepared for anything like a plane landing on the road.

I agree with Tesla that Lidar and hd maps are crutches/training wheels for driving as humans can get by without those tools/that set of data. In some ways AI can be better than the average human as we get easily distracted by phones, texts, music, eating, applying makeup or whatever other crazy thing you’ve seen people do. AI can look using multiple cameras in all directions at all times while not getting distracted or fatigued.

I do think it’ll happen someday but with some limitations at first and definitely not on Elon’s time frame. It’ll also require municipalities to really maintain roads/lane markings to help these systems as much as possible in the future.

Yes but AI also gets confused in situations where humans dont even blink. The self driving problem is much harder than it seems. Almost every real life situation is an edge case. Maps/lidar are just necessary components, people discounting them are usually talking their book. It'd be a lot easier if we had much better mapping since it would reduce where true self driving was necessary. You just need it to work, idc if its maps/lidar or ai, I'd just like it to get the job done.

Humans and AI are in no way similar. And calling what is essentially ML AI is also a huge stretch.

[Nysoz]

Yeah it’s going to be a new frontier for sure. Maybe tickets go to the driver until it gets so remarkably good that you can read a book or sleep. Because until that point the driver is supposed to monitor everything.

at least with the Tesla there are cameras recording all the time to help decrease he said she said. Then theoretically once it gets so good,Tesla insurance comes into play. Now it’s Tesla insurance with cameras documenting everything vs she said.

then with more Tesla’s on the road they can then pull up cameras from cars that area at the time of the accident to recreate things even more. Definitely potential big brother type of situation though.

[Nysoz]

Yes but AI also gets confused in situations where humans dont even blink. The self driving problem is much harder than it seems. Almost every real life situation is an edge case. Maps/lidar are just necessary components, people discounting them are usually talking their book. It'd be a lot easier if we had much better mapping since it would reduce where true self driving was necessary. You just need it to work, idc if its maps/lidar or ai, I'd just like it to get the job done.

Humans and AI are in no way similar. And calling what is essentially ML AI is also a huge stretch.

Yeah Elon definitely underestimates how hard it is. Especially the last few 9s.

humans and ai don’t have to be similar but ai just needs to get to a point where it can function adequately. There’s a lot of smart people out there working on it and having it things do I can’t even imagine.

Geofence it first, only in good weather, give the car a blinking light on top to let everyone else know it’s a robot car and to give it some leeway. This definitely won’t be an overnight flip of a switch and uncharted territory for everyone involved.

[Zaphod]

Yeah Elon definitely underestimates how hard it is. Especially the last few 9s.

humans and ai don’t have to be similar but ai just needs to get to a point where it can function adequately. There’s a lot of smart people out there working on it and having it things do I can’t even imagine.

Geofence it first, only in good weather, give the car a blinking light on top to let everyone else know it’s a robot car and to give it some leeway. This definitely won’t be an overnight flip of a switch and uncharted territory for everyone involved.

right, all these things are helpful and can allow the other stuff to do what it does best and not worry about other things.

[fatlittlepig]

I doubt we will even see level 4 freeway driving with the current sensor suite and software, not to mention local roads.

[billy]

I don’t think it’ll ever be perfect. It just has to be better than the average human driver (which is pretty bad) to be a benefit and potentially save lives. The question is how much better before regulators allow it?

They're retraining the entire system now. Instead of labeling pictures they’re labeling video. What will help speed up the process and make it that much better is their dojo neural net supercomputer.

Humans are able to drive with 2 cameras and a brain. Will a car be able to do the same? I guess it depends on how well you can program AI to learn. AI can beat humans in complex tasks like Starcraft or go or chess. But all of those games are based off rules. Driving you have to know when it’s ok/necessary to break the rules and be prepared for anything like a plane landing on the road.

I agree with Tesla that Lidar and hd maps are crutches/training wheels for driving as humans can get by without those tools/that set of data. In some ways AI can be better than the average human as we get easily distracted by phones, texts, music, eating, applying makeup or whatever other crazy thing you’ve seen people do. AI can look using multiple cameras in all directions at all times while not getting distracted or fatigued.

I do think it’ll happen someday but with some limitations at first and definitely not on Elon’s time frame. It’ll also require municipalities to really maintain roads/lane markings to help these systems as much as possible in the future.

LOL- as someone who lived in NY/NJ - how is the AI going to respond to a huge pothole that would break the car vs crossing a yellow line thats barely there while cars are double parked on both sides with pedestrians jaywalking across the street?
At least the cameras would prevent the fake lawsuits from people jumping into cars pretending the car ran into them