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

[M|S|M MYP]

IT's RICH hearing GJ call anything "fluff".

I see GJ has not learned any manners by constantly trying to shout over the host. What a ...going back to his cherry picked data again, it was fun dunking on him in the beginning, but it has gotten to a point where he's just that 'annoying' kid in class.

 

[SPadival]

James Douma's explanation of the transition from V10->V12 tells an entirely different story. Karpathy wanted to do end to end since the beginning, but the signal is too weak using the training software they had at the time. So they developed in multiple steps, adding bits and pieces of NN to every version from 10->11 trying to boost the signal so eventually they can do end to end. So V12 is not only an additional attachment to but also came from the backs of V10 and V11. It was not some rouge team who wanted to try a different method that is not on the same path as FSD trajectory.

How come George was doing it 7 years ago with an abysmally low level of resources (hardware, human and software)!

We do not code in the rules of the driving problem.. we learn from hours and hours and hours of data

i.e. What Elon said last week during his FSD V12 demo, George was onto it 7 effing years ago!!! .. The licence plate of his car then was FUEL ON or FU ELON.. whichever way you want to interpret it

Let's give credit where it is due. The biggest mistake that Elon made was probably not going ahead with recruiting George for a couple of million bucks more. But hey.. we're finally there. (To be fair to Karpathy, he was focussed on Tesla Vision rather than planning & control).

 

[Cosmacelf]

Just finished the Lex interview of Walter Issacson. Was a good listen. Lex is starting to edit his interviews to cut out useless tangents and low quality content, which, frankly, was mostly his long winded questions. So this ended up being one of the most enjoyable Lex interviews for a long time.

 

[heltok]

How come George was doing it 7 years ago with an abysmally low level of resources (hardware, human and software)!

Lane Keep Assist on a highway is a lot less complex problem than city street FSD.

 

[skybluecgreen]

Yup, this is like an electric Volkswagen XL1, the car that VW should have build 10 years ago (their loss, soon their demise):

Volkswagen 1-litre car - Wikipedia

en.wikipedia.org

Doug DeMuro knows about the XL1:

Cheers!

Yeah that XL1 is pretty boss. And the Insight was also ahead of its time.

Future or past?

 

[AperiodicCoder]

So you all know this, but let me just reiterate what an advantage Tesla has with their online ordering system:

I'm up late, supposed to be working, but getting distracted with an Out of Spec Detailing video about Colton's new inventory Model 3 Performance that he got for a screaming deal (with federal and state tax credits factored in).

So I get the spark in my head to check out the inventory cars. Filtering for long range Model 3, I see one result: white with black interior and 19 inch wheels, exactly like my 2018 3 except AWD instead of RWD. And it's got a $3k discount on it. And I have FSD on my current 3 so I can transfer as my estimated delivery is 12-19 September.

About an hour later I had my order placed and all the paperwork filled out. I don't qualify for any tax credits but I'll have a brand new Model 3 LR with FSD for ~$48k -- I'll probably spend the money saved on acceleration boost as well.

So, bringing home the point I started with: it's so easy to just impulse purchase a Tesla when the logic strikes, this is yet another huge advantage Tesla continues to have over most of its competitors. I know that highland is coming out but this is going to be an amazing car for many years to come.

 

[MC3OZ]

Making money on new car sales is unusual.

In future car companies that can't make money in new EVs probably will not be in business for long.

Many opportunities to make additional money are already being tapped by Tesla, software, charging, insurance, FSD, mechanise, upgrades.

Finance is another potential opportunity, but fundamentally if a car company can't make money selling EVs it will be very hard to make enough money in other areas to survive.

To compete with Tesla other companies need to get the cost of parts, including batteries, in the same ballpark, and they need to be as good at manufacturing.

If the Chinese carmakers need to set up production in the US / EU to make avoid import duties then they lose some of the cost advantage they have in China.

The absolute best case scenario I can imagine for other carmakers in future is same COGS as Tesla, equally good product, same product volumes, same delivery / production efficiency and same brand appeal. I can only imagine it by being charitable to the point of being almost totally irrational, for most it will take at least a decade and will be a major miracle.

But assuming we have car markers somehow on margin parity with Tesla, why would they choose to sell at a loss, when there is no prospect of making up for that loss in other areas?

And do they have software, charging, insurance, Robotaxis and FSD?

 

[Cosmacelf]

Btw the Issacson book is now available via kindle, I’ve been reading it for the last 2 hours.

 

[SPadival]

Lane Keep Assist on a highway is a lot less complex problem than city street FSD.

Strawman.. Tesla didn’t immediately start developing city street FSD from the get go. They started with Autosteer aka lane keep assist. This video was recorded before Tesla even had lane keep assist tech (they were relying on mobile eye).

All that I am saying is if George Hotz could start lane keep assist with end to end NN, so could Tesla at that time. So this argument about weak signal is bull. 20/20 hindsight - We just have to accept the fact that we spent nearly a decade labelling, auto labelling, occupancy networking etc etc.

The good news is that we’re still ahead of George Hotz and his teeny company, simply by the virtue of having more resources and George Hotz is not selling out to OEMs (probably ever.)

 

[Dikkie Dik]

Has this been shared? Seba posted that the price of Li-ion batteries dropped below $100/kWh

https://twitter.com/x/status/1701251192037474756

According to

 

[Artful Dodger]

I can't think of a more qualified biographer for Musk than Isaacson.

Oh, weird AI... I thought you said weird Al...

 

[heltok]

Strawman.. Tesla didn’t immediately start developing city street FSD from the get go. They started with Autosteer aka lane keep assist. This video was recorded before Tesla even had lane keep assist tech (they were relying on mobile eye).

All that I am saying is if George Hotz could start lane keep assist with end to end NN, so could Tesla at that time. So this argument about weak signal is bull. 20/20 hindsight - We just have to accept the fact that we spent nearly a decade labelling, auto labelling, occupancy networking etc etc.

The good news is that we’re still ahead of George and his teeny company, simply by the virtue of having more resources.

George Hotz was pretty late to the game. Here is a presentation from 7 years ago talking about the then history of end2end:

Back then courses at Audacity, Coursera etc had students play computer games and then training end2end system to imitate their driving:

Even back in 1986 we had self driving cars doing lane keeping.

The hard part is doing it well, dealing with edge cases. If you look at the Hotz video, sure it was driving itself, but it was pretty shaky and not much better than some of the projects students did.

 

[Buckminster]

So if the $25k car is priced at err $25k then what will the smaller less complicated Robotaxi be priced at?
$15k? Nope $100k or not sold at all.

Strangely, the value of a Robotaxi might be worth more than an X or CT with FSD due to profitability (charging / cleaning costs).

 

[WarpedOne]

not sold at all.

 

[jeewee3000]

How come George was doing it 7 years ago with an abysmally low level of resources (hardware, human and software)!

We do not code in the rules of the driving problem.. we learn from hours and hours and hours of data

i.e. What Elon said last week during his FSD V12 demo, George was onto it 7 effing years ago!!! .. The licence plate of his car then was FUEL ON or FU ELON.. whichever way you want to interpret it

Let's give credit where it is due. The biggest mistake that Elon made was probably not going ahead with recruiting George for a couple of million bucks more. But hey.. we're finally there. (To be fair to Karpathy, he was focussed on Tesla Vision rather than planning & control).

I've been following George Hotz quite closely the last five years and can tell you your comment is off the mark.

- Yes, Hotz is a believer in "end to end" since a long time, but so were Tesla/Karpathy/Elon. It was just not practically viable yet;

- Hotz himself stated many times that Tesla has the right approach. Comma.AI is "2 years behind Tesla" at all times. Tesla will win Level 5 FSD. Etc. Hotz thinks the theory (E2E) is known for a long time, someone just has to "build it". (Meaning having the resources, compute and data to do so)

- Elon and Hotz both know getting a car to keep a lane is one of the more 'easy' parts of FSD (relatively speaking) and you can get there 80% of the way with 10-20% of the work/data. Look at early Comma.AI and early Tesla Autopilot. Both might have suggested "FSD is right around the corner" but both teams underestimated how complex the solution is.

- Recruiting Hotz is not possible. He has too free of a spirit (or dare I say 'ego') to be part of Tesla and therefore step in the shadow of Elon. Hotz himself explains he gets bored easily with stuff, he has to be intrinsically motivated and many times it's because he wants to crack a certain problem. But for example Karpathy's job (=90% data curation) would not be fitted for Hotz.
Hotz dropped out of Twitter soon after Elon "hired" him. Hotz his attention span or motivation or call it what you will, is not sufficient to be in it "for the long haul" for a big project like Tesla FSD or boots on Mars for SpaceX. Not dissing Hotz, he's a genius, but he has an entirely different philosophy and a lower "pain threshold".

So Hotz would only be an asset for a few months and then move on to the next thing. I don't think Tesla "missed out" on not hiring Hotz (if he even ever wanted to join, which I doubt).

 

[heltok]

I've been following George Hotz quite closely the last five years and can tell you your comment is off the mark.

- Yes, Hotz is a believer in "end to end" since a long time, but so were Tesla/Karpathy/Elon. It was just not practically viable yet;

- Hotz himself stated many times that Tesla has the right approach. Comma.AI is "2 years behind Tesla" at all times. Tesla will win Level 5 FSD. Etc. Hotz thinks the theory (E2E) is known for a long time, someone just has to "build it". (Meaning having the resources, compute and data to do so)

- Elon and Hotz both know getting a car to keep a lane is one of the more 'easy' parts of FSD (relatively speaking) and you can get there 80% of the way with 10-20% of the work/data. Look at early Comma.AI and early Tesla Autopilot. Both might have suggested "FSD is right around the corner" but both teams underestimated how complex the solution is.

- Recruiting Hotz is not possible. He has too free of a spirit (or dare I say 'ego') to be part of Tesla and therefore step in the shadow of Elon. Hotz himself explains he gets bored easily with stuff, he has to be intrinsically motivated and many times it's because he wants to crack a certain problem. But for example Karpathy's job (=90% data curation) would not be fitted for Hotz.
Hotz dropped out of Twitter soon after Elon "hired" him. Hotz his attention span or motivation or call it what you will, is not sufficient to be in it "for the long haul" for a big project like Tesla FSD or boots on Mars for SpaceX. Not dissing Hotz, he's a genius, but he has an entirely different philosophy and a lower "pain threshold".

So Hotz would only be an asset for a few months and then move on to the next thing. I don't think Tesla "missed out" on not hiring Hotz (if he even ever wanted to join, which I doubt).

Accurate. Elon offered Hotz millions if he could replace Mobileye but he declined. Hotz prefers working in smaller teams, when you run larger team you have go from being a developer to a manager and he is a lot better developer than he is a manager and prefers that.

Karpathy was great for his knowledge in training large neural networks, he set up the data labelling right, which is really hard. But I think he was mainly good at deep learning, still very intelligent and good at learning new stuff, but not really a roboticist like Ashok and Milan. Currently Tesla is all about rapid iteration across the entire stack and their team is battle hardened and experts at rapid iterations. With more compute coming in expect even more rapid iterations and rewrites. Once you have a good system made right from scratch it is so much easier to iterate than it will be for the software that VW has with their Cariad etc. Expect them to fall behind even more as they accumulate more technical debt that Elon, Karpathy etc have been so good at keeping away!

The book is out on Audible now also! Maybe we should do a thread with all our new insights from the book?
Audible version doesn't have any images, anyone know where you can find them?

 

[Buckminster]

Troy sticking to his demand limited message:
https://x.com/troyteslike/status/1700989901700010157?s=46
Poll topped out at 33% and received the most votes by a distance.

 

[FEERSUMENDJIN]

Yup, this is like an electric Volkswagen XL1, the car that VW should have build 10 years ago (their loss, soon their demise):

Volkswagen 1-litre car - Wikipedia

en.wikipedia.org

Doug DeMuro knows about the XL1:

Cheers!

Look. No Robotaxi is going to be "low profile" such as any of these concepts. It will be upright with good accessibility for people of all ages and physical attributes.

 

[Buckminster]

Here is an existing biography thread that I have already pushed more than a few times. Let me know if you want an investors section thread.
Isaacson Biography of Elon Musk

 

[Xepa777]

George Hotz was pretty late to the game. Here is a presentation from 7 years ago talking about the then history of end2end:

Back then courses at Audacity, Coursera etc had students play computer games and then training end2end system to imitate their driving:

Even back in 1986 we had self driving cars doing lane keeping.

The hard part is doing it well, dealing with edge cases. If you look at the Hotz video, sure it was driving itself, but it was pretty shaky and not much better than some of the projects students did.

1. George was years ahead of the curve saying end-to-end was the only way to do FSD vs Elon who publicly said at least twice that it was being worked on, but not necessary for FSD. And people are correcting those online claiming Tesla was first to this approach.

2. George sucks at taking anything to commercialization at scale due to lack of leadership, recruiting for proper talent, getting proper infrastructure, etc. reason. He’s a hacker, not someone who you can rely on to do something well enough to change the world.

Those two statements can both be true, and seeing arguments that are cross pointing at the two is tiresome and intellectually disingenuous.