Tesla, TSLA & the Investment World: the Perpetual Investors' Roundtable

[PhaseWhite]

Actually the reason TSLA FSD is valued at zero is a little different.

Waymos and Cruises have been showing the demo for a few years now. Yet, they have not unveiled a robotaxi. So, now that Tesla demoed the FSD to investors now, they think it will be years before Tesla can release robotaxis.

There is also quite a bit of skepticism about robotaxi in general - whether it will ever be possible, mainly because of such slow progress by Waymo & Cruise.

It is possible Waymo & Cruise can release robotaxi in one city before Tesla. But, how fast can they release in others ? Afterall because of Lidar, their technology is worthless unless they achieve complete FSD. Ofcourse, there is the problem of test data for learning. They don't have 100k cars on the road in 4 continents.

For Tesla it is very different. Because this is how Tesla can do it ...
1. Keep releasing newer versions of FSD to the fleet and make the NN better & better. (Till oct '2020).
2. Get regulators to approve FSD without drivers or riders at low speeds. Stage the cars at various parking lots & let people include their cars in TN. When the rider calls, the car goes to the rider location and the rider will drive the car with the help of FSD that drives well 99% of the time. (Nov '2020 till regulators allow full FSD)
3. Full FSD in places regulators allow it (top urban areas)

Waymo & Cruise have to essentially wait till step 3 above to make any money - and they have confidence they have solved FSD (atleast for that geofenced area).

Objectively I think Tesla's autonomous strategy has a lot to like:

1. Incremental deployment with shadow mode feedback vs perfect L4 first means Tesla can move faster and leverage paying customers to help train and validate thier software.

2. Ability to fund the developement through customer software purchases vs burning investment capital. And I also agree its likely the Tesla network will be launched with a driver supervision requirement at first which allows more immediate revenue and test data generation.

3. The HW3 chip is really a game changer. Given what Tesla has already been able to do with the nVidia chip and 10x less effivtive throughput I think well see massive improvements in the next year. No other OEMs are anywhere close and the media is basically in denial of this major achievment. That will work to Tesla's advantage as other OEMs wont realize how f***ed they are until it's too late.

That said I'm not very confident we'll see Musks robotaxi dream realized fully in the next 24 months but because of Tesla's incremental approach it will be successful eventually and will continue to monitize things along the way.

 

[Electroman]

Please don't feel I'm picking on you, but this is exactly what I'm talking about. Amir Efrati (the anti-Waymo version of Anton Wahlman) once said Waymo struggled with unprotected left hand turns. This somehow morphed into "Waymo can't make left turns" here on TMC. Then a clip shows up with Tesla following a lane line on a protected left turn, and suddenly Tesla is "ahead" of Waymo.

It's pure wishful thinking. Tesla does not even attempt unprotected left turns yet.

Meanwhile, Waymo vans perform unprotected lefts every day. They are cautious when doing so. Perhaps too cautious. Last year Andrew J. Hawkins of The Verge rode in a Waymo van. It made several unprotected left turns, but he talks about one in particular that took a bit longer than he wanted:
Waymo Unprotected Left Turn

Is excess caution a problem? It is if it causes angry drivers behind you to riot. But insufficient caution is an even worse problem. You have to strike a balance, while also provide a smooth, non-jerky ride that doesn't scare your robotaxi customers. Waymo has worked on this and similar problems for years. They have better sensors than Tesla. They have more smart people than Tesla. Can someone tell me why it won't also take Telsa years to solve problems like unprotected left turns?

I agree. IMHO that video, is not even a left turn, per se. Tesla simply followed the car in front at slow speed with clearly marked guide lanes.

 

[Artful Dodger]

... Right click on Username, go to Profile and add to Ignore

That doesn't block the 150 REPLIES from other member to the 50 off-topic comments.

DELETE OFF TOPIC.

 

[Zhelko Dimic]

With lots of real world training data examples, which tesla has, it should be fairly trivial to train the neural net to handle any common driving case. The only thing to worry about with Tesla’s, or anyones, approach is the uncommon things that occur while driving. Tesla’s vision system, massive amount of training data, and now onboard processing power is a massive advantage.

Learn more about how neutral nets work and specifically how they are trained and you will understand Tesla’s advantage here. It is very very important. This isn’t something you throw man hours at to try and engineer your way through, or some algorithm you need to solve.

Disclaimer: I’m a software engineer

Driving itself is not in NNs, it's in heuristics layer still, with continuous strategy of carefully downloading particular tasks to NN, as tasks are better understood and training data can be extracted. They aren't there for general driving, let alone most complex tasks of driving. It was described during the autonomous demo day.

 

[AcesDealt]

Neither Waymo or Tesla have the technical solution yet and who is ahead is the wrong question. Tesla is the only one on the track that leads to rapid scaling. The technical tricks can usually be described in a few journal papers or carried in the mind of a single employee. The hard part is designing your entire business around the most infectious vector.

 

[K3jSe7Llp7]

I'm with Elon Musk. If, god forbid, the price went to zero I'd be there with him buying all the way. The ultimate privilege of having a few dollars to invest is being able to take a stand with the one person trying to fix what's wrong.

 

[EVNow]

Driving itself is not in NNs, it's in heuristics layer still, with continuous strategy of carefully downloading particular tasks to NN, as tasks are better understood and training data can be extracted. They aren't there for general driving, let alone most complex tasks of driving. It was described during the autonomous demo day.

Yes, some of those will continue to be in heuristics. But what they are doing is getting more and more information from NN - including possible intent of all the objects around the car. That's how they got the lane change working, for eg. Finally that is the difficult part which people think makes self-driving impossible - how will the car know what other cars and pedestrians are going to do.

 

[Tim S]

Thanks to all who recommended the "ignore" feature. I had forgotten about that. And thanks to the mods for banning the "Elon shouldn't tweet" member.

Back on topic, kind of... there is apparently now a new Tesla Service Center in Albany NY (in the suburb of Latham actually). Only 160 miles from @neroden! Latham - 326 Old Niskayuna Road | Tesla. 5 miles from me

Here's a pic from a member of our local FB group...

 

[Singuy]

People need to understand when Tesla claims they are "way ahead", they are way ahead solving vision. They are most likely not ahead today if you compare a waymo car with miles logged per disengagement vs a Tesla.

The point here is that lidar is not going to be the answer since it has many short comings like it doesn't recognize objects therefore it'll be too dumb to make a decision on what the intent of the object is (snow, rain, plastic bags can all be seen as a undrivable barrier), whichever company solves vision first will be the winner. Since no other companies believe vision is solvable in the way Elon and his team thinks it'll be solved(as in vision plus intelligence will make the car as smart as a human), no other company is actually trying. So autonomous driving via vision only is currently science fiction...but the rate at which is becoming science fact should not be understated.

No one is going to dispute "once vision is solved, then lidar is completely obsolete".

 

[MarcusMaximus]

Driving itself is not in NNs, it's in heuristics layer still, with continuous strategy of carefully downloading particular tasks to NN, as tasks are better understood and training data can be extracted. They aren't there for general driving, let alone most complex tasks of driving. It was described during the autonomous demo day.

I’d caution that that’s true for what’s in current builds, in the networks running on HW2. We have no idea what they’re running in their testing FSD/HW3 targeted builds.

 

[Causalien]

Porsche costs 3x to maintain compared to normal car, Ferrari is 10x.
Depreciation can be quite severe, brutal even, but if you choose rare car, it can be minuscule (my GT4), in a good economy.

I think Roadster will have severe depreciation, and modest maintenance costs, but if they choose to make low number of them, depreciation won't be severe, unless you drive a lot. Collector's cars must have low mileage, so it's a A->A vehicle for Sunday drive, not A->B vehicle.

As for ratio of assets, how important is truly fancy car to you? I bought my first new Porsche few years back, after my mortgage was paid off, RRSP in very good shape, and I had liquid assets to cash pay for it. But then, I'm sucker for track/driving events, so that played most important role, wouldn't have bothered otherwise, I arguable spend too much on cars.

I am just wondering about some simple ratio so I can be responsible for my finances. Like house net worth should be around 90- your age or stuff like at most 20% of your asset in debt, but am wondering if luxury car has a different rule like that. Ya, it'll be an A->A type of driving, but I mostly just want to enjoy life and do it without stressing my finances.

 

[Causalien]

People need to understand when Tesla claims they are "way ahead", they are way ahead solving vision. They are most likely not ahead today if you compare a waymo car with miles logged per disengagement vs a Tesla.

The point here is that lidar is not going to be the answer since it has many short comings like it doesn't recognize objects therefore it'll be too dumb to make a decision on what the intent of the object is (snow, rain, plastic bags can all be seen as a undrivable barrier), whichever company solves vision first will be the winner. Since no other companies believe vision is solvable in the way Elon and his team thinks it'll be solved(as in vision plus intelligence will make the car as smart as a human), no other company is actually trying. So autonomous driving via vision only is currently science fiction...but the rate at which is becoming science fact should not be understated.

No one is going to dispute "once vision is solved, then lidar is completely obsolete".

I am skeptical that no other company is trying vision. I don't think it's too much of a stretch to slap both Lidar and 8 cameras onto a car and then have the NN train based off of that. In fact, I think it will be a valid solution, but more expensive. Once that's solved, we are then talking about how much Mass Manufacturing can bring down Lidar's cost.

 

[Chunky Jr.]

I'm hoping TSLA does well tomorrow now that the bad news has been digested and the market is now absorbing the SEC settlement.

 

[Carl Raymond]

I am skeptical that no other company is trying vision. I don't think it's too much of a stretch to slap both Lidar and 8 cameras onto a car and then have the NN train based off of that. In fact, I think it will be a valid solution, but more expensive. Once that's solved, we are then talking about how much Mass Manufacturing can bring down Lidar's cost.

But remember that solving vision requires masses of vehicles reporting back with edge case data. If somebody else has tens of thousands of cars gathering data, I think we’d know about it.

 

[MarcusMaximus]

I am skeptical that no other company is trying vision. I don't think it's too much of a stretch to slap both Lidar and 8 cameras onto a car and then have the NN train based off of that. In fact, I think it will be a valid solution, but more expensive. Once that's solved, we are then talking about how much Mass Manufacturing can bring down Lidar's cost.

You could. You could also slap 16 cameras on. Or a backscatter X-ray. The question is what that gets you. If you already have depth from stereo vision, your LiDAR just becomes a very expensive, very crappy camera.

Oh, and, of course, you need vast quantities of data from cars with that exact layout. Tesla is aiming to accomplish it with 500k-1m cars reporting data over the course of years. The time it takes likely reduces with some <-1 exponent on the number of cars you’re using(as more cars gives not just more data, but more varied locations/situations within the data). So your LiDAR-equipped cars will need to be out in a similar sized fleet for a similar amount of time(or a much longer time if it’s smaller) to solve the same tasks.

 

[Fact Checking]

Driving itself is not in NNs, it's in heuristics layer still, with continuous strategy of carefully downloading particular tasks to NN, as tasks are better understood and training data can be extracted. They aren't there for general driving, let alone most complex tasks of driving. It was described during the autonomous demo day.

Yes, but note that the biggest part of FSD not handled by NNs isn't even just the high level policy decisions/heuristics, but physics:

• Newtonian mechanics: how fast other vehicles are moving (speed), how large they are (probable inertial mass),
• probable near-future position of other vehicles,
• how the Tesla is moving: current position and speed (can be integrated based on inertial sensors such as accelerometers and GPS), attitude (angle of the car),
• map position and expected road conditions based on the GIS data, selected route,
• road conditions and slippage based on wheel rotation sensors,
• path evaluation and collision detection, safe free road section ahead and on parallel lanes,
• state of battery charge versus distance to destination, battery temperature optimization for the next charging stop,
• path optimization and targeting: within the safe boundaries established by the above physics calculations, what speed and position should the car target, which lanes to navigate intersections - with safety, efficiency and comfort optimized at once,

This is a LOT of physics and very little heuristics, with a lot of physics calculations, fed by the 3D object detection of the vision NNs, and these are well understood algorithms that make little sense to offload to the NNs.

FSD is architecturally pretty close to a game engine: a 3D object engine, physics engine, high level game logic and a user interface - except that in FSD the 3D engine detects 3D objects via an NN - while in a game the 3D engine generates a 3D scene.

 

[Artful Dodger]

I am just wondering about some simple ratio so I can be responsible for my finances. Like house net worth should be around 90- your age or stuff like at most 20% of your asset in debt, but am wondering if luxury car has a different rule like that. Ya, it'll be an A->A type of driving, but I mostly just want to enjoy life and do it without stressing my finances.

So EASY! Get an original Roadster right now!. Pay the $29K for the bty upgrade.
Enjoy Vancouver Island all summer long! Tour the interior wine country in the Fall.

Cheers!

 

[Fact Checking]

People need to understand when Tesla claims they are "way ahead", they are way ahead solving vision. They are most likely not ahead today if you compare a waymo car with miles logged per disengagement vs a Tesla.

Tesla is many orders of magnitude ahead of Waymo based on disengagement events as well: Waymo has a fleet of a ~few hundred vehicles with millions of miles of FSD events.

Tesla has a fleet of 600,000+ vehicles that has been automatically reporting disengagements for years, for over a billion miles driven with their FSD test code (Autopilot and NoA).

Tesla decided to not voluntarily report their disengagement counts, which is probably why many think that Waymo is ahead in disengagement data - they aren't.

 

[Artful Dodger]

I am skeptical that no other company is trying vision. I don't think it's too much of a stretch to slap both Lidar and 8 cameras onto a car and then have the NN train based off of that. In fact, I think it will be a valid solution, but more expensive. Once that's solved, we are then talking about how much Mass Manufacturing can bring down Lidar's cost.

Yes, computer vision is the Mobileye/Intel approach. But their business model is to act as vendor to the larger auto industry, like DELCO or BOSCH or LG.

Telsa's vertical integration means they will ALWAYS have a lower cost basis in volume production, after R&D is amortized. Tesla's long view means they plan to eat the auto industry's breakfast, lunch, and dinner.

Sounds like the legacy auto industry would benefit from some computer eyes in the back of their heads right now...

Cheers!

 

[AudubonB]

I'd like to say a "plethora"...

Thank you. That means a lot.