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

[Xepa777]

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.

I don’t know how you can be so definitive unless you’re in the Tesla team. I know several engineers at tesla and even I’m not that definitive. Like three months ago if you ran a poll most people her would’ve said the Tesla robotaxi was going to be some shuttle*ish thing or modular design that can expand to a minivan. Not a two seater.


For the record I think the final design will conceptually be a two seater robotaxi and narrow 4 seater compact that can fit on those narrow streets globally.

 

[SPadival]

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

I have been following him even longer than that. I have followed his livestreams when there were barely 50 people watching. Used the Chffr app that he talked about in the video above and bought the dongle. But what has that got to do with anything??

It’s pretty clear from the video that he implemented openpilot via E-to-E NN while Tesla went in another direction with autopilot - even today, it gets iffy when the lane divides into two or broadens a bit.

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.

That’s not true. As per the Bloomberg article , Hotz offered to work for Elon in 2015, just not as an employee. Again, the offer came from Hotz.

There was a proposal that if Hotz could do better than Mobileye’s technology in a test, then Musk would reward him with a lucrative contract. Hotz, though, broke off the talks when he felt that Musk kept changing the terms. “Frankly, I think you should just work at Tesla,” Musk wrote to Hotz in an e-mail. “I’m happy to work out a multimillion-dollar bonus with a longer time horizon that pays out as soon as we discontinue Mobileye.”

“I appreciate the offer,” Hotz replied, “but like I’ve said, I’m not looking for a job. I’ll ping you when I crush Mobileye.”

Musk simply answered, “OK.”

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.

Again, this is not what happened. Elon didn’t recruit him to Twitter. Hotz volunteered his time as an unpaid “intern”. And he left because he wanted to refactor the whole code while Elon obviously didn’t want to do that while burning cash. He wanted to introduce new features ASAP and stem the cash burn. Elon took the right decision here, but he should probably revisit this in the future.

The fact that Hotz again volunteered on his own should indicate that he makes an exception for working with Elon.

Here is a presentation

I stopped reading after the word “presentation”. Elon has been a bad influence

 

[SPadival]

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.

Or it’s his ego/integrity that’s keeping him from scaling up faster. I wrote off George and comma ai a year ago and yet comma ai is still here and expanding into what can be called tiny dojo.

I like his approach. And the fact that he’s not selling out to legacy OEMs is an added bonus viz my Tesla investment. He has made it clear that he’s competing for the number 2 spot (Android of FSD) after Tesla (Apple of FSD).

 

[Buckminster]

Investor thread:
Elon Musk Biography By Walter Isaacson

 

[jeewee3000]

I have been following him even longer than that. I have followed his livestreams when there were barely 50 people watching. Used the Chffr app that he talked about in the video above and bought the dongle. But what has that got to do with anything??

It’s pretty clear from the video that he implemented openpilot via E-to-E NN while Tesla went in another direction with autopilot - even today, it gets iffy when the lane divides into two or broadens a bit.


That’s not true. As per the Bloomberg article , Hotz offered to work for Elon in 2015, just not as an employee. Again, the offer came from Hotz.





Again, this is not what happened. Elon didn’t recruit him to Twitter. Hotz volunteered his time as an unpaid “intern”. And he left because he wanted to refactor the whole code while Elon obviously didn’t want to do that while burning cash. He wanted to introduce new features ASAP and stem the cash burn. Elon took the right decision here, but he should probably revisit this in the future.

The fact that Hotz again volunteered on his own should indicate that he makes an exception for working with Elon.


I stopped reading after the word “presentation”. Elon has been a bad influence

Good additions of which I am aware. I was just summarising for the non-informed.

However, point stands that you cannot just recruit Hotz due to his character. And he has clearly shown to not follow through in his endeavours. Good enough is good enough for Hotz, and then he moves onto the next thing. I get why, but that wouldn't help Tesla or SpaceX.

 

[Todd Burch]

The problem with Hotz’s/Comma.ai’s approach is that they went basically straight for end-to-end and didn’t spend time curating more granular/modular neural networks. It may turn out that end-to-end will be what cracks this nut and Hotz will probably end up being right about that, but without more specific modular networks you don’t have a way to crank the data engine. Karpathy and Tesla recognized this quite a while ago—and the data engine was probably Karpathy’s greatest gift to Tesla. Comma.ai doesn’t have the funding or infrastructure to seek sufficient specific datasets from their fleet, or the compute power/storage to deal with it. It just requires a boatload of money.

Without these smaller NNs, you can’t send requests out to the fleet for intersections with stop signs, or examples of very curvy roads, or examples of intersections with more than three sets of traffic lights, etc.

Although Tesla has iterated and changed their approach several times, the time has not been wasted. The hardest part of FSD will likely be handling the many edge cases—and nobody has the ability to send searches out to their fleet to curate and collect certain edge cases like Tesla.

I think these modular NNs will be used for annotating/labelling the screen in the car too, to give drivers feedback.

Douma put it well in his recent interview with Better with Herbert: Tesla has a search engine, not a tape recorder. You just can’t physically scour through all the video you get if you just manually record it. You need a way to parse it, evaluate it, label it, seek it out, and figure out what to throw away and what to keep. And you need a way to automate as much of that as you can, because you will need several Boaty McBoatfaces-worth of data.

 

[ToTheStars]

The problem with Hotz’s/Comma.ai’s approach is that they went basically straight for end-to-end and didn’t spend time curating more granular/modular neural networks. It may turn out that end-to-end will be what cracks this nut and Hotz will probably end up being right about that, but without more specific modular networks you don’t have a way to crank the data engine. Karpathy and Tesla recognized this quite a while ago—and the data engine was probably Karpathy’s greatest gift to Tesla. Comma.ai doesn’t have the funding or infrastructure to seek sufficient specific datasets from their fleet, or the compute power/storage to deal with it. It just requires a boatload of money.

Without these smaller NNs, you can’t send requests out to the fleet for intersections with stop signs, or examples of very curvy roads, or examples of intersections with more than three sets of traffic lights, etc.

Although Tesla has iterated and changed their approach several times, the time has not been wasted. The hardest part of FSD will likely be handling the many edge cases—and nobody has the ability to send searches out to their fleet to curate and collect certain edge cases like Tesla.

I think these modular NNs will be used for annotating/labelling the screen in the car too, to give drivers feedback.

Douma put it well in his recent interview with Better with Herbert: Tesla has a search engine, not a tape recorder. You just can’t physically scour through all the video you get if you just manually record it. You need a way to parse it, evaluate it, label it, seek it out, and figure out what to throw away and what to keep. And you need a way to automate as much of that as you can, because you will need several Boaty McBoatfaces-worth of data.

I completely agree.
It looks like Gall's Law

"A complex system that works is invariably found to have evolved from a simple system that worked. A complex system designed from scratch never works and cannot be patched up to make it work. You have to start over with a working simple system."

 

[Singuy]

Or it’s his ego/integrity that’s keeping him from scaling up faster. I wrote off George and comma ai a year ago and yet comma ai is still here and expanding into what can be called tiny dojo.

I like his approach. And the fact that he’s not selling out to legacy OEMs is an added bonus viz my Tesla investment. He has made it clear that he’s competing for the number 2 spot (Android of FSD) after Tesla (Apple of FSD).

So 7 years ahead of Tesla on NtoN yield a result of a multiple attempt drive to taco bell late last year. You believe the only reason why it's worst than teslas v11 is because they don't have the resources.

Sounds like a contradiction here...doesn't exactly scream "the right approach" as comma picked the most compute intensive/data intensive way but having no money for compute and no infrastructure for massive data curation.

 

[SPadival]

Sounds like a contradiction here...doesn't exactly scream "the right approach" as comma picked the most compute intensive/data intensive way but having no money for compute and no infrastructure for massive data curation.

Huh? As Tesla has shown, the non E-to-E approach is also compute intensive/data intensive AND also human resource intensive (with all those manual labourers labellers toiling away in Giga NY). Also, as Elon said during the V12 livestream, E2E is also less processor intensive on the inference computer. George was working with a retail mobile phone to begin with (I think it was Oneplus 2) and still uses a mobile phone processor.

George is pretty happy being in the second place. He's developing Tinygrad and Tinybox to solve the backend processing bandwidth issue. He wants to bootstrap everything from a tiny scale and so be it... man in the arena and all. He's not a risk to my Tesla investment and I wish him all the best.

 

[Electroman]

Is there any other way for him to be? (Rhetorical). Honestly, if he could pull out the terminator et al excuses, fine by me. Just means a bigger mountain fortress.

Jonas is another variation of Gordon Johnson. Pretending to understand what he is talking and you add a dash of disingenuousness to it.

Was he the one that asked during an earnings call if Tesla cars can be used as missiles, and gave the stock price range from $60 to $600. ?

 

[Electroman]

Looks like Tesla setup/is setting up a second version of the prototype Semi production line, speculated to be more automated

https://twitter.com/x/status/1701235775415660990

Prototype line? What happened to the actual production line?

 

[MartinAustin]

Hmph... somewhat not-a-surprise that TSLA would fall this morning

 

[stealthyc]

I have been following him even longer than that. I have followed his livestreams when there were barely 50 people watching. Used the Chffr app that he talked about in the video above and bought the dongle. But what has that got to do with anything??

I have had interactions with George from far longer ago than most people, and I can say that he is incredibly intelligent and motivated but overly egotistical in such a manner that he would likely be an insufferable employee and also difficult to retain unless the spotlight was solely on him.

 

[bkp_duke]

Prototype line? What happened to the actual production line?

The original line was all hand work for the Semi. A semi-automated prototype line is the next step. Over time, that will get upgraded with robots, etc. and made as automated as possible. Pretty much how Tesla has done all their other lines in the past as well. The Semi is so different from everything else, and they have to drive so many more miles in their lifetime, that Tesla is appropriately taking a "slow and steady" approach.

 

[Big Time]

280 before mmd would be nice
To me it looks like someone wants this to go higher

 

[2daMoon]

Hmph... somewhat not-a-surprise that TSLA would fall this morning

Yeah, off to a slow start.

Maybe it will pick up the pace later.

 

[Singuy]

Huh? As Tesla has shown, the non E-to-E approach is also compute intensive/data intensive AND also human resource intensive (with all those manual labourers labellers toiling away in Giga NY). Also, as Elon said during the V12 livestream, E2E is also less processor intensive on the inference computer. George was working with a retail mobile phone to begin with (I think it was Oneplus 2) and still uses a mobile phone processor.

George is pretty happy being in the second place. He's developing Tinygrad and Tinybox to solve the backend processing bandwidth issue. He wants to bootstrap everything from a tiny scale and so be it... man in the arena and all. He's not a risk to my Tesla investment and I wish him all the best.

Perhaps you never heard of failed products that are "ahead of it's time". E2E was not viable 7 years ago. The amount of compute needed is intensive and that's why Tesla built smaller and less ambitious NN models before tackling E2E today after they have sufficient A100, H100, and Dojo. The Dojo project started back then and it usually takes 5+ years to go from design to silicon. You see how the labour intensive work Tesla was doing back then did not hit earnings as much as the cost of compute going forward today. Tesla could not afford the cost of gigafactories amount of compute/year unlike what ELon has guided this year.

Comma started 7 years ago and it took 6 months for Tesla to surpass them when all the stars lined up ready for prime time.

Just to be clear, Elon's vision was always end to end. He has said "vision to biological neural net to controls is how humans drive" to the point of ad nauseum.

 

[jerry33]

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).

Well, it could be sold for $15K without the ability to connect to the Robotaxi network. I'd think it would be just the thing for students.

 

[ItsNotAboutTheMoney]

Prototype line? What happened to the actual production line?

That will happen when they have an actual product.

 

[Singuy]

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).

Less complicated Robotaxi might just be a leased product. I have a feeling Tesla may not want to deal with the operating cost of running the hardware(cleaning the seats, dealing with vandals, fixing paint jobs). So people may end up paying Tesla a monthly fee plus a cut of their revenue from the network. I imagine Tesla will be employing the world to run their taxi service just like how they employed the world to train their FSDb and getting paid for it.