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

[NicoV]

Another thing that Elon mentioned (and IIRC confirmed by Ashok) was that FSD is already a good driver in other countries.
That is something I didn’t expect. Looking at our local road infra, there are loads of situations that do not occur in the US. Like country roads that are so narrow that cars can’t pass each other. Every couple of hundred meters there’s a bit of extra asphalt on the side of the road where you have to wait until the car coming from the other side has passed. Or not even that, where one of the 2 cars has to drive a hundred meter in reverse before reaching such a place. Lets not even start with the myriad of traffic signs with a text (in the local language) below indicating in what circumstances the sign is valid. So a robot car would need to be equiped with OCR and an LLM to interpret the instructions (in the cases that info isn’t in the map data).
There have been reports of FSD test drivers in Europe (or at least job descriptions), so maybe Tesla already has some experience with this.
Anyway, if Elon and Ashok are right, introducing FSD in other countries may be much faster than expected.

 

[GatorMeat]

I meant, do you agree with what he said in that tweet about the Chancery court? You seem to be discounting the tweet due to the way he plays tesla stock.

I really don't care what he has to say about ..... anything. What are his qualifications? What is his experise? Living in your car, shoving all your funds into one stock, and losing your marraige over it and acting half your age doesn't make you smart or worth listening to.

 

[Knightshade]

Another thing that Elon mentioned (and IIRC confirmed by Ashok) was that FSD is already a good driver in other countries.
That is something I didn’t expect. Looking at our local road infra, there are loads of situations that do not occur in the US. Like country roads that are so narrow that cars can’t pass each other. Every couple of hundred meters there’s a bit of extra asphalt on the side of the road where you have to wait until the car coming from the other side has passed. Or not even that, where one of the 2 cars has to drive a hundred meter in reverse before reaching such a place. Lets not even start with the myriad of traffic signs with a text (in the local language) below indicating in what circumstances the sign is valid. So a robot car would need to be equiped with OCR and an LLM to interpret the instructions (in the cases that info isn’t in the map data).
There have been reports of FSD test drivers in Europe (or at least job descriptions), so maybe Tesla already has some experience with this.
Anyway, if Elon and Ashok are right, introducing FSD in other countries may be much faster than expected.

AFAIK the major obstacles to rolling it out in Teslas other two big markets are:

EU: Regulations make it impossible to roll out right now. Regs coming later this year/early next might make it POSSIBLE, but still not rapid (Rohan said some things about this soon before he left)- so this isn't coming SUPER soon.

China: Regulations for the service aren't a huge obstacle--- but regulations require car data to not leave china-- which makes local training a huge obstacle until Tesla has a lot of compute physically IN china---- made extra hard by it being illegal to import the more powerful GPUs. But you might see a less-optimal version launch there once it's "good enough" and then it'll improve steadily as they can ramp compute there with lesser hardware (but perhaps more of it).

 

[scaesare]

Free cash flow= cash flow from operations - capx
Capx directed at ai chip sets
Without investment in Ai compute chips there is no FSD

Yup.


‘We’re Going to Spend Staggering Amounts of Money in Capex’

-Elon Musk

 

[Usain]

Does anyone have any examples of vehicles with materially different wheelbases and chassis being built on the same physical line (like a Honda Civic and Accord or CRV and HRV).

I'm not sure if this counts, but I know that the Chevy Bolt and Chevy Sonic once shared an assembly line at the Orion plant in MI. They were both designed by Daewoo in South Korea, but the Bolt was an EV and the Sonic was ICE. They were of similar size, but I'm not sure if they really had that much in common as the Bolt was designed from the ground up to be an EV.

 

[Usain]

It got better at parking near sidewalk, especially when there is dedicated entrance. Does it impressively well at my drives to my local pool.

Yeah, 12.3 can do a little bit of pulling over. But it's still got a long way to go before it could do everything it needs to do for picking up passengers. I'm hoping that Tesla has already been working on this. We do know that Actually Smart Summon is being developed and that gives some of the needed functionality as well. I'm hoping we might get a demo of passenger pickup on 8/8.

 

[Zaddy Daddy]

FTFY

Changed this to read "Autos" rather than "EVs" as an "uncertain macroeconomic environment" is unlikely to target a tiny portion of the auto industry (EVs) in exclusion of the rest of it.

Everybody gets their portion of that macroeconomic pie. Though how they digest it may differ slightly, the effect will still be seen by all who are being served that dish.

Except everyone knows from the data that EVs have been hurt much worse than "Autos" in general. Not just Teslas, all EVs.

I'm ascribing that to "uncertain macroeconomic environment". If you want to separate it from "macro environment", then the CFO should have said "weakening BEV demand".

We cannot ignore that this has happened.

 

[Zaddy Daddy]

What, exactly, do we think the "new EVs" on the Model 3 / Y lines are going to be?

Are they / we counting Model Y refresh / Project Juniper as one of these?

Is there going to be a compact car?

Or a van? What do we think the breakdown and timelines are going to be?

This was too vague from the report.

 

[Usain]

So a robot car would need to be equiped with OCR and an LLM to interpret the instructions (in the cases that info isn’t in the map data).

As I understand it, FSD doesn't need to understand the semantics of language on road signs at all. It just mimics what a human does after seeing a particular sign.

For instance, it knows the design of a yield sign and the shape of the word "Yield" but it has no idea what words and letters are or what they actually mean. It just knows that when it sees that sign it needs to exhibit "yielding" behavior.

In other words, FSD is illiterate, but illiterate humans can learn to drive just fine. So it shouldn't be a major problem.

 

[MP3Mike]

China: Regulations for the service aren't a huge obstacle--- but regulations require car data to not leave china-- which makes local training a huge obstacle until Tesla has a lot of compute physically IN china---- made extra hard by it being illegal to import the more powerful GPUs. But you might see a less-optimal version launch there once it's "good enough" and then it'll improve steadily as they can ramp compute there with lesser hardware (but perhaps more of it).

Maybe they will do most of the early localization for China via simulation. They mentioned one difference was not changing lanes across a solid line, it seems like that would be easy enough to simulate and create videos to feed into the training. It would probably require more human work than using videos from the in-country vehicles, but until they get sufficient training clusters in China it is probably their best path forward.

 

[Knightshade]

As I understand it, FSD doesn't need to understand the semantics of language on road signs at all. It just mimics what a human does after seeing a particular sign.

For instance, it knows the design of a yield sign and the shape of the word "Yield" but it has no idea what words and letters are or what they actually mean. It just knows that when it sees that sign it needs to exhibit "yielding" behavior.

In other words, FSD is illiterate, but illiterate humans can learn to drive just fine. So it shouldn't be a major problem.

It's not a problem for universal only-means-one-thing signs like stop signs.

It's a major problem for signs where the words matter.

A lot of parking signs come immediately to mind where you have to know more than "that is a parking sign"-- you need to read the hours and days and lengths of time and other conditions parking is allowed or not.

Ditto with school zone signs that are also hour/day dependent and not at all universal.

Another one is signs like road closed to through traffic, but local traffic is ok... maybe your destination is in that local traffic only area-- if you just reject because you see a generic road-closed you'll never get to your destination.

Lots of other examples I'm sure (and even more in other countries)

Maybe they will do most of the early localization for China via simulation. They mentioned one difference was not changing lanes across a solid line, it seems like that would be easy enough to simulate and create videos to feed into the training. It would probably require more human work than using videos from the in-country vehicles, but until they get sufficient training clusters in China it is probably their best path forward.

I think that's pretty reasonable for good-enough functionality as far as following basic rules of the road and getting a functional L2 deployed... stuff like "drives and reacts like local humans do" stuff, one of the hallmarks of why people like 12.x so much, I think that's a lot harder to get in simulation vs real world data.

 

[Doggydogworld]

Does anyone have any examples of vehicles with materially different wheelbases and chassis being built on the same physical line (like a Honda Civic and Accord or CRV and HRV).

Shortened/lengthened wheelbases and different bodies are pretty common. Mercedes Factory 56 line builds S Class and Maybach in varying wheelbases. Honda and Toyota are especially known for multiple models on the same line. Here's a 5 year old article about a new type of line Toyota was playing with. It's not my field so I don't know how far they've taken this concept.

Tesla's magic was hard-wiring their lines to be inflexible but efficient. They don't even build 3 and Y on the same line. This works great because Musk is an amazing marketer able to drive 500k to 1m+ volumes for single models.

 

[dhrivnak]

And semi. But, as we learned from the call, the semi assembly line is still quite a ways off, so no need to ramp up 4680 any faster. As Elon explained, a big reason for 4680 was cost. Other manufacturers were placing HUGE battery orders, making cells from everyone so expensive. But, now that the legacy manufacturers have backed off EV production there is a glut of batteries to be bought on the cheap, so the cost reductions from the 4680 format aren't as important right now.

4680’s could go into the 3 effectively dropping the price 15-20% by getting the tax credit back. That would make a difference

 

[hobbes]

Another thing that Elon mentioned (and IIRC confirmed by Ashok) was that FSD is already a good driver in other countries.
That is something I didn’t expect. Looking at our local road infra, there are loads of situations that do not occur in the US. Like country roads that are so narrow that cars can’t pass each other. Every couple of hundred meters there’s a bit of extra asphalt on the side of the road where you have to wait until the car coming from the other side has passed. Or not even that, where one of the 2 cars has to drive a hundred meter in reverse before reaching such a place. Lets not even start with the myriad of traffic signs with a text (in the local language) below indicating in what circumstances the sign is valid. So a robot car would need to be equiped with OCR and an LLM to interpret the instructions (in the cases that info isn’t in the map data).
There have been reports of FSD test drivers in Europe (or at least job descriptions), so maybe Tesla already has some experience with this.
Anyway, if Elon and Ashok are right, introducing FSD in other countries may be much faster than expected.

I understood that more as a theoretical thing that in principle FSD should drive ok in other countries (as opposed to geofenced/HD maps based systems). As far as I understand (not an expert) they´d take the same NNs (NN architecture) and train them with driving data from another region which should not take very long as they have the data and their compute capacity is growing a lot.

The limiting factor IMHO will be regulatory (looking at the small carful steps UNECE is taking).

My guess is that when computing is scaled up it should take maybe months with some iterations to train FSD for another region, but likely still years until you´re allowed to use it the way it is possible in the US (that is without the driver confirming every single action it takes).

 

[mongo]

Something that I have to scratch my head a bit and would love some input

Why focusing on reducing the production cost of Robotaxi is so important? For a vehicle that will run 100's of thousand of miles, maybe over a million miles and bring revenues multiples of the cost it takes to build it, why going unboxed on it and not exclusively for the consumer one? This is the one low cost matters most no?

If a Robotaxi costs $50k or even $100k to build, with the projected revenues it will generate seems like a non issue

Maybe it's just a timing problem, making Robotaxi today makes no sense since the software isn't ready, and we need new consumer facing vehicles ASAP

Or I'm just overthinking since any vehicle can be a Robotaxi, specially removing the steering wheel and pedals, so the more expensive Robotaxi is just the current line up without driver interface

Robotaxi's cost isn't the driver, the architecture is along with timing as you mentioned.
Robo and NextGen were planned to both be unboxed, but the climate shifted so NextGen (now MidGen?) was pulled forward and hybridized to use existing infrastructure. Robo and NextGen² will continue development along with the unboxed assembly process.

 

[X Fan]

Long time investor and owner of 5 vehicles since '16 (currently have 3). Just sold all of my holdings.

 

[Unpilot]

Long time investor and owner of 5 vehicles since '16 (currently have 3). Just sold all of my holdings.

Ok

 

[jabloomf1230]

Robotaxi's cost isn't the driver, the architecture is along with timing as you mentioned.
Robo and NextGen were planned to both be unboxed, but the climate shifted so NextGen (now MidGen?) was pulled forward and hybridized to use existing infrastructure. Robo and NextGen² will continue development along with the unboxed assembly process.

Elon seems convinced that the only way to make a robotaxi profitable is to use creative manufacturing methods to reduce the cost of each unit. The present approach used by Waymo and others is to build a handcrafted robotaxi using an existing vehicle outfitted with a diverse array of sensors. Although this approach has proven to be technically feasible, the cost per unit produced is too high to presently be profitable. New technologies succeed because they provide a benefit, either new features, lower cost or in some cases both. Robotaxis do not provide any new features over that which existing rideshare options provide (other than the novelty), so a robotaxi must be less expensive in order to maintain profitability.

 

[Nissanman]

Long time investor and owner of 5 vehicles since '16 (currently have 3). Just sold all of my holdings.

Same! (Well, actually owner of 3 since 2020 (currently 2). Sold all my TSLA shares this morning. Fan of the company but believe I can find a better place to invest than the stock.

 

[keydiver]

I'd be very surprised to see this. Genuinely don't even know how they could pull it off. The demand for M3 and MY is not expected to increase significantly and there aren't any new models even announced, not to mention ramped up.

I think Tesla has a lot more data and foresight to project sales for this year, and I sincerely doubt they would mislead investors during an earnings call if they know otherwise.