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

[JustMe]

I think some people are definitely confusing peaker plants and grid stabilisation.

When you have variable wind/solar energy over the course of a few hours (or even minutes!) you need grid stabilisation. Thats the thing that often requires sub-second response, or 10 minute response, and where batteries are the ultimate solution (because gas takes a while to spin up).
When you have a powerplant or interconnector FAIL or some major weather event knocks out, or down the power supply for hours or days... thats when you need a peaker plant.

There is an order of magnitude difference in the capital investment required, and number of batteries required to perform these tasks.

edit: great chart showing this on wikipedia for the hornsdale facility:

https://upload.wikimedia.org/wikipedia/commons/a/ab/Hornsdale_battery%2C_diagram.png

Tesla may well think that their batteries are more profitable put in new cars, or semis, before they put any in grid stabilisation facilities, and only when they are plentiful maybe build some peaker-plant scale facilities.

I don't know how Texas grid works, but here in the UK, commercial-scale battery owners can 'opt-in' to multiple schemes at once, so you can make your battery available for grid stabilisation, or for longer term (hours) storage or supply, at the discretion of the grid operator, with payments being very different for each scheme.

It’s already happening.

https://www.usnews.com/news/national-news/articles/2019-05-21/why-california-nixed-a-natural-gas-power-plant-in-favor-of-batteries

Work begins on 409MW/900MWh Florida battery project to ease natural gas plants into retirement

Construction work has begun in the US on what is claimed to be the world’s biggest solar-charged battery storage project, by utility company Florida Power & Light (FPL).

www.energy-storage.news

Elon Musk has said that he expected Tesla energy to grow to the size of their auto business. That’s not going to happen with grid stabilization services alone. Of course they need to make lots of batteries. His recent tweet suggests they won’t be so cell constrained By next year. Here’s hoping

 

[Maitri982]

Good points - I mostly agree with most of them.

One thing though is hard to quantify: The wide-spread belief that self-driving is a) possible at all and b) the arrival is imminent - within a very few years.

There is a cultural/psychological component that may be culturally related, maybe depends how strong your grounding in science and technology is. (Maybe also depends on how much science fiction you have read.)

People got used to Deep Blue beating Kasparov in chess, but it was fairly big then.
For a long time after, the game of Go was the next frontier, until suddenly that was not only conquered, but vanquished.

Go is *big* in the china cultural sphere - Koreans openly wept when their world champion Lee Se-dol was roundly beaten in 2016 by 4 games to 1.
Lee was optimistic, nay cocky before the game, thinking that he could hold a master class in go and teach these these arrogant (mostly) Americans programmers a lesson in humility.

He himself afterwords said that the go-program had a 'feel or deep intuition for playing go' that utterly stunned him. Given that, it is a testament to his mental strength that he recovered form this great shock, and actually won a subsequent game by playing perhaps the best game of his life.

Many people's minds were blown then.
Lee retired in 2019, saying that no matter how good he could still become he could never beat the new 'entity'.
Former Go champion beaten by DeepMind retires after declaring AI invincible

So, for option F) i would add this:
A large number of totally ordinary people waking up to the fact that the arrival of Self Driving might be pivotal moment, to be compared with other big technological leaps like steam engines, the assembly line, computing or the internet. And invest according to their newfound realization.

I think it will be an iPhone moment.

It will, but comparisons to Go are improper. FSD much much harder.

 

[powertoold]

So Karpathy and jedi engineers have been hard at work on FSD for 4 years, Tesla has their own highly trained labeling team, software workflow, training cluster, and FSD chip.

Competitors can just buy Teslas and Edit > Copy and then Edit > Paste FSD?

 

[MTL_HABS1909]

https://twitter.com/x/status/1375932607880380417

I assume Gary is a bit off-base with this, but AI is not in my wheelhouse.

He is. He doesn’t understand how machine learning and neural nets work. You can’t just “copy” it. This is laughable.

 

[EVWatcher]

So Karpathy and jedi engineers have been hard at work on FSD for 4 years, Tesla has their own highly trained labeling team, software workflow, training cluster, and FSD chip.

Competitors can just buy Teslas and Edit > Copy and then Edit > Paste FSD?

There's a lot going on that once Tesla achieves success where subsequent entrants will follow.

1. When the first version of a complicated computing solution is achieved by pushing the hardware to the brink, hardware improvements continue at a compounding rate. It might take Tesla's state of the art designed chip tailored for their system but 10 years later an average kids laptop will have more than enough general purpose power to do the trick.

edit3: just saw @KarenRei respond to this

https://twitter.com/x/status/1375453293057478672

2. A lot of the techniques Tesla is finding success with were grabbed from research papers, like the birds eye view composite and temporal labeling ideas. Once someone proves an idea gets you to the solution there will be 100 other companies going after the same techniques.

3. The autopilot team will not all stay at Tesla, maybe they are going to other places and can't work on self driving stuff. But they will have learned things developing FSD that can be translated to other goals and they will teach those things to people at other companies so as that knowledge spreads through the tech industry it will indirectly facilitate other people learning.

4. There are companies working in parallel to achieve what Tesla is trying to achieve, and maybe Tesla beats them to the punch, but they won't stop working on it.

The thing that Tesla has in their pocket at this point is they will already have 3 million+ plus RoboTaxi capable vehicles on the road by the end of 2022 when the next competitor has 100s. And you could say that if someone sells a plug and play package to Toyota that Toyota could put 20 million vehicles on the road in a year but:

1. Traditional automakers take years to plan for a model, whereas as Munro has pointed out Tesla can iterate a dozen times on an octovalve inside of a year. So even if they have a solution implementing it would be hard.

2. Customer acceptance of Robotaxi's probably wont be overnight. Pretty sure when Tesla launches Robotaxis they could have more Robotaxi hours on the road per week than Uber has uber driver hours on the road per week in less than a month. ( edit2: if you are a Tesla owner and a Tesla investor keep it in mind it is potentially in your investment benefit to send your car out for this reason ) They're might not be much more for Robotaxi utilization than Tesla can produce cars. X-factor here is that there's a massive overconcentration of Teslas in SF/LA and not too many in Chicago which leaves holes in regional markets potentially. (Same goes for the rest of the world, just using the US as an example)

edit: 3. It is also assumed that Teslas data moat will protect them for some time. Time well tell if thats true or not. But I think Gary was suggesting people could buy Teslas and mount hardware on it to collect data and then send them off driving 24 hours a day to collect data. Im curious how that will play out because you know people will try.

I realize that's more than you asked but I felt like typing alot

 

[Thumper]

4 years for Soviets to steal and copy A bomb.

 

[Maitri982]

There's a lot going on that once Tesla achieves success where subsequent entrants will follow.

1. When the first version of a complicated computing solution is achieved by pushing the hardware to the brink, hardware improvements continue at a compounding rate. It might take Tesla's state of the art designed chip tailored for their system but 10 years later an average kids laptop will have more than enough general purpose power to do the trick.

edit3: just saw @KarenRei respond to this

https://twitter.com/x/status/1375453293057478672

2. A lot of the techniques Tesla is finding success with were grabbed from research papers, like the birds eye view composite and temporal labeling ideas. Once someone proves an idea gets you to the solution there will be 100 other companies going after the same techniques.

3. The autopilot team will not all stay at Tesla, maybe they are going to other places and can't work on self driving stuff. But they will have learned things developing FSD that can be translated to other goals and they will teach those things to people at other companies so as that knowledge spreads through the tech industry it will indirectly facilitate other people learning.

4. There are companies working in parallel to achieve what Tesla is trying to achieve, and maybe Tesla beats them to the punch, but they won't stop working on it.

The thing that Tesla has in their pocket at this point is they will already have 3 million+ plus RoboTaxi capable vehicles on the road by the end of 2022 when the next competitor has 100s. And you could say that if someone sells a plug and play package to Toyota that Toyota could put 20 million vehicles on the road in a year but:

1. Traditional automakers take years to plan for a model, whereas as Munro has pointed out Tesla can iterate a dozen times on an octovalve inside of a year. So even if they have a solution implementing it would be hard.

2. Customer acceptance of Robotaxi's probably wont be overnight. Pretty sure when Tesla launches Robotaxis they could have more Robotaxi hours on the road per week than Uber has uber driver hours on the road per week in less than a month. ( edit2: if you are a Tesla owner and a Tesla investor keep it in mind it is potentially in your investment benefit to send your car out for this reason ) They're might not be much more for Robotaxi utilization than Tesla can produce cars. X-factor here is that there's a massive overconcentration of Teslas in SF/LA and not too many in Chicago which leaves holes in regional markets potentially. (Same goes for the rest of the world, just using the US as an example)

edit: 3. It is also assumed that Teslas data moat will protect them for some time. Time well tell if thats true or not. But I think Gary was suggesting people could buy Teslas and mount hardware on it to collect data and then send them off driving 24 hours a day to collect data. Im curious how that will play out because you know people will try.

I realize that's more than you asked but I felt like typing alot

It's still about the batteries. Even if GM, for example, gets FSD at some point, it will cost much more to run a network of gas guzzlers. They don't have enough battery capacity to build a ton of cars. Also, they would have to pack power hungry Nvidia or equivalent cards into their vehicles as they don't have their own low cost low power boards.

 

[Artful Dodger]

I only feel that way because it seems unlikely and incredible Tesla could get there this year, because if they do then the stock will probably explode upward lightning fast.

Yeah, one can imagine a number of ways: (each having some toehold in reality)

• First Robotaxi fleet goes operational in Dubai in 2021
• Tesla main contractor in $1T plan to rebuild U.S. Grid/Storage/EV infrastructure
• Dojo comes online in Q2 and solves cancer within the first 24 hrs; 3-D prints solution
• Tesla announces Giga LOE (space solar panels); 1st microwave pwr lnk is to ERCOT
• Tesla purchases SpaceX for $74B in stock; Announces nickel/platinum asteroid mine
• X.com begins taking deposits for black-hole powered Galactic transportation system

Remember, this is ELON we're talking about here (he doesn't think like you and I). If it's not explicity BANNED by the Laws of Physics, well, he's gonna swing for the fence.

Cheers!

 

[powertoold]

2. A lot of the techniques Tesla is finding success with were grabbed from research papers, like the birds eye view composite and temporal labeling ideas. Once someone proves an idea gets you to the solution there will be 100 other companies going after the same techniques.

Yes, Karpathy has said that the best vision NN algorithms are available in research papers, but I'm not sure if others can catch up quickly on data curation. It seems Tesla AP team spends much of their energy on data curation and the in-house labeling software workflow. Labeling itself requires a lot of discipline and standardization.

Perhaps others will copy the approach, but they need to first copy the sensor suite, the software workflow, and then train disciplined labelers, etc. Then they'd have to collect and programmatically address all the edge cases again. This sounds like a job for a startup, not lazy slow moving OEMs that still fail at OTA updates.

 

[22522]

Yes, Karpathy has said that the best vision NN algorithms are available in research papers, but I'm not sure if others can catch up quickly on data curation. It seems Tesla AP team spends much of their energy on data curation and the in-house labeling software workflow. Labeling itself requires a lot of discipline and standardization.

Perhaps others will copy the approach, but they need to first copy the sensor suite, the software workflow, and then train disciplined labelers, etc. Then they'd have to collect and programmatically address all the edge cases again. This sounds like a job for a startup, not lazy slow moving OEMs that still fail at OTA updates.

Does physics lead labels, or only shape? If physics leads labels the set converges a lot faster.

 

[powertoold]

Does physics lead labels, or only shape? If physics leads labels the set converges a lot faster.

Not sure I understand your question, but Karpathy has said that the labeling is very important. The labels must be done manually and done as accurately as possible. That's why Tesla invests a lot of resources into their labeling team.

 

[EVWatcher]

It's still about the batteries. Even if GM, for example, gets FSD at some point, it will cost much more to run a network of gas guzzlers. They don't have enough battery capacity to build a ton of cars. Also, they would have to pack power hungry Nvidia or equivalent cards into their vehicles as they don't have their own low cost low power boards.

I certainly didn’t mean to imply my list was exhaustive, this is certainly another very good point. I have a hard time envisioning any company both beating Tesla to the software solution and the practical rollout of the product.

 

[buttershrimp]

https://twitter.com/x/status/1375932607880380417

I assume Gary is a bit off-base with this, but AI is not in my wheelhouse.

Man this is good example of how very smart people can know nothing at all.

 

[MTL_HABS1909]

Man this is good example of how very smart people can know nothing at all.

Dunning-Kruger effect

 

[Cosmacelf]

I've really lost all respect for Berkshire and Buffett now. And all these years I was tricked into thinking he was such a kind, grandfatherly sort of figure. Stupid me.

He never was such. You don't become one of the richest people in the world through investing without being ruthless. Just for example, in the depths of the '08 recession, he was writing op-eds to make the situation sound even more dire in an effort to get legislation passed that would benefit Berkshire. His grandfather persona is carefully cultivated - it isn't real.

 

[Mo City]

Man this is good example of how very smart people can know nothing at all.

"Stay in your lane". Personally, I despise the phrase but it applies here.

Also, part of intelligence is being aware of what you don't know.

 

[EVWatcher]

"Stay in your lane". Personally, I despise the phrase but it applies here.

Also, part of intelligence is being aware of what you don't know.

I’ve found in life it’s rare for a person to know what they don’t know. Rarer still for someone who excels in a particular field.

 

[ZsoZso]

my only competence in AI is reading Life 3.0 book by Max Tegmark and from what I understood in deep learning neural nets integrating algorithms from all the datas explored and reviewed by computer engineers from problematic situation reported by all the fleet is what is giving the edge and the capability of programming an AI to respond adequately in every situation possible. Unless the other companies reverse engineer the coding and have an AI algorithm that improved itself without using all the real life data, you can’t reproduce all the problematic situations and make the AI improve itself without the computed data reviewed by engineers like they showed at Autonomy day.

I might be totally wrong, this is far from my field of expertise, and can’t wait for someone in the field to correct Gary Black or myself.

Yes, Karpathy has said that the best vision NN algorithms are available in research papers, but I'm not sure if others can catch up quickly on data curation. It seems Tesla AP team spends much of their energy on data curation and the in-house labeling software workflow. Labeling itself requires a lot of discipline and standardization.

Perhaps others will copy the approach, but they need to first copy the sensor suite, the software workflow, and then train disciplined labelers, etc. Then they'd have to collect and programmatically address all the edge cases again. This sounds like a job for a startup, not lazy slow moving OEMs that still fail at OTA updates.

You are both assuming the ethical and fully legal kind of "copy and replicate the methodology" performing a clean, independent implementation of it, but perhaps Gary was referring to a more literal "copy", i.e. reverse engineer the FSD CPU and board, copy the trained NN topology and weight set and replicate it. Mind you, such "copy" would be illegal, violating copyrights and other Tesla IP.
However, technically such copy would be possible as long as they intend to use equivalent sensor suite and hardware to execute the NN. That way they could bypass the need for large data collection and training. However, it would be a "static" copy of the given NN as trained at the moment, next time the Tesla downloads a new weight set via OTA update, they would have to repeat the copy and update their own NN other wise they start falling behind.

 

[Xd85]

https://twitter.com/x/status/1375932607880380417

I assume Gary is a bit off-base with this, but AI is not in my wheelhouse.

Deep Blue and chess supercomputers were not "learning" as much as computing. The threshold was to get enough threads computed in a reasonable amount of time, and the "learning" was needed only to be able to assign weights to outcomes. Outcomes bound by strict rules of chess, not stochastic behavior of humans and traffic. Even so, chess has not been "solved" (i.e. absolute prediction of all moves in game) unless there are <7 pieces remaining. ("Tablebases" ) For those not in the know, there are 32 pieces to start .

Agree..apples and oranges

 

[Big Time]

Innovators like Steve Jobs never got phased out by people copying. Either will Elon. They just continue to elevate what ever they are working on. When Steve Jobs did the iMac, suddenly flooded with all the copycats in the PC realm. Did any of that stop Apple? No they moved to Phones. Then a million androids. And just when everyone thinks Apple is finished being great they release the M1 chip. But all the while getting the lions share because everyone knows what’s best is best. People who get copied are already ahead of the pack, they just move to the next level why the people below pick up their table scraps. Get the lion share, then move on back to innovating. That’s what Apple and Tesla will do to stay on top.