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

[CLK350]

Further notes on AI day regarding Tesla Bot aka Optimus Subprime ( subprime as in subprime mortgages or not yet prime .. ?) beyond the hardware D1/ Dojo Clip and FSD advances noted here and elsewhere.

We were clearly being teased starting with Aug 4th's Dennis Hong's tweet (and subsequent deleted tweets) with the invite clearly saying " Attendees will be among the first to see our latest developments in supercomputing and neural network training. They’ll also get an inside look at what’s next for AI at Tesla beyond our vehicle fleet. "

Dennis Hong and his Romela lab at UCLA should now be viewed as the robotics equivalent of Jeff Dahn for batteries.

Tesla's approach for Optimus will be similar to that for cars: high value expensive applications first, with costs going down progressively.

The next upcoming Optimus incarnations are probably well advanced ( whence the Tweet mentioned by TMC's ZeApelido @WholeMarsBlog “There are some really crazy things coming” — Tesla employee ) - best guess in line with what Elon mentioned (repetitive, dangerous ..) most likely will be Disaster response robots Dennis Hong's lab* already has worked on prototypes for the DARPA Robotics Challenge back in 2015, then bomb squad robots (not to mention obvious military applications, armed sentinels, foot soldier buddies, and submarine/ tank crews).

Longer term the obvious last leg delivery for robotaxis, definitely a threat for Amazon; but even more of a threat, Optimus will help smaller Amazon competitors grow as the costs of systems go down. See small Amazon competitors like NewEgg who are already in place.

(*) Dennis also has a 2nd gen cooking assistant, funded by Wu Brothers, so under NDA, has many side jobs, like designing amusement park rides, and shares an interesting philosophy for his endeavors, " making people happy/ making the world a better place " A pretty good match for Elon (and brother Kimbal), see that recent July 2021 video (not a word there about Tesla, fun to watch knowing what we know now )

Other notes:
- hiring new blood and opening a new area for Tesla has, besides providing a helpful push for the tough final FSD release, also given great motivation for his existing AI engineers to stay onboard (vs leaving for other companies).
- the YouTube videos of David Lee and James Houma , Lex Fridman on AI day , are the most informative, next will be Rob Maurer's who was at AI day in person, so will have new info, and whose team already put out a 19 mins short version of Tesla's AI Day presentation. And of course whatever Cathie Wood and her Ark Invest team come up with (hi Tasha ;D )

Investor's viewpoint: I liked best the metaphor viewing Optimus as a new free long term option on TSLA - its incremental R&D costs is minuscule too.


RoMeLa | Robotics and Mechanisms Laboratory

 

[-=buzz=-]

Another good deep dive on the Dojo computer. Near the end they speculate how they can make Dojo 10x faster in which the system on wafer design could be key. This is more like Cerebras vs the currently innervation of D1 tiles. There are 4 links here, each to each part of the articule.

Sadly they show an apparent lack of knowledge, when talking about "unprecedented" full scale wafer chips in their 10x speculation, when Cerebras is already on Gen2..., what else are they hyping about without actual understanding it/knowing the state-of-the-art..?

Nevertheless there are some obvious improvements for the next generation:

- smaller process size should allow slightly higher clocks
- with smaller dies maybe they could place more chips per tile
- architectural fixes and improvements
- split cache and network router into its own chip and stack the cpu-part on top (as recently shown by AMD) guessing this is about 50% of the size, that's an instant double of cores per chip.. (with its own challenges of course)
- general density and throughput improvements on the tile packaging. It's their first iteration after all.

 

[lazybot]

CNBC is promoting this tweet with Gordon;'s false accusation. More proof they have this guy on only for the drama and clicks and views.

 

[redan]

View attachment 700016

Good news, the deer head (I mean “head” of the back office) just gave their approval.

 

[Lessmog]

CNBC is promoting this tweet with Gordon;'s false accusation. More proof they have this guy on only for the drama and clicks and views.
View attachment 700017

So, just chewing some fat here:
IF cnbs gets sued by Tesla for slander, libel, defamation or the like -- would they not be somewhat impeded in what they can report on while a court process is pending, open or active? IANAL, obv.

 

[StealthP3D]

View attachment 700016

"Back office" must equal "butt".

 

[Lessmog]

"Back office" must equal "butt".

Yeah, he just missed the spelling of "orifice".

And the deer head is not lying -- no lips moooving. On the other hand it has a dilemma ...

 

[Singuy]

Sadly they show an apparent lack of knowledge, when talking about "unprecedented" full scale wafer chips in their 10x speculation, when Cerebras is already on Gen2..., what else are they hyping about without actual understanding it/knowing the state-of-the-art..?

Nevertheless there are some obvious improvements for the next generation:

- smaller process size should allow slightly higher clocks
- with smaller dies maybe they could place more chips per tile
- architectural fixes and improvements
- split cache and network router into its own chip and stack the cpu-part on top (as recently shown by AMD) guessing this is about 50% of the size, that's an instant double of cores per chip.. (with its own challenges of course)
- general density and throughput improvements on the tile packaging. It's their first iteration after all.

There are so little performance benchmarks of Cerebras even for gen 1 out there besides "claims" from marketing materials. I am thinking it's having a very difficult time from the software side of thing to fully utilize its performance when Nvidia's software stack is the gold standard in generalized DL. On paper the specs are ridiculous, but most likely not so much in the real world. DOJO thankfully will not have this problem as it's vertically integrated.

 

[ninpb]

View attachment 700016

 

[capster]

Can someone please explain in simple words, what is the function of Dojo? How does Dojo help solve FSD? I keep hearing 'labelling', but I don't quite understand what that means. How are others solving this problem?

This is how I understood :

video labels
Cars ----------> Dojo -----------> Cars <cars can drive better now>

Let’s say I showed you a picture of a car and asked you what was in the picture. You’d say “a car."

When you want a system to recognize a car or other objects, you train the system by presenting it with an image and simultaneously telling it "this is a car." (There is complexity in how you process the image, extract objects for recognition, and much else that I don’t go into here).

Training requires lots of pictures of cars and other objects. Training also requires repeated presentations of each association to be made, e.g. picture of car/category "car".

Tesla’s vehicles take lots of pictures of cars and other things. However, those pictures cannot be used for training until the objects in them have been delineated and identified: “Here is car," "there is a traffic cone." That is what is meant by "labeling."

They said they have a thousand people doing labeling. Tesla’s trove of data is so vast that that is still nowhere near enough. So they also automate the labeling, called autolabeling, whereby their computers infer the labels from scenes by working backwards and forwards through prerecorded video streams (presumably bootstrapping from human labeled data and trained networks).

Autolabeling is a nifty trick. It is also requires huge amounts of compute power.

Training the very large networks requires truly massive amounts of compute power. This is because ’training’ involves tiny adjustments in a very large number of variables, aka weights or coefficients, each time an associative pairing is presented and each is presented many times during a single training run.

Dojo is intended to provide the massive amounts of compute power needed for training. I don’t recall if they explicitly said it also supports autolabeling, though it would make sense.

Once a network is trained it can be used. The network is essentially frozen. That is what runs on the car. It still needs some compute power in use but nothing like the scale needed to label the video data and train the network in the first place.

 

[UkNorthampton]

Here's an article on the boom in orders for EV/car manufacturing equipment. Some equipment suppliers are backed up quite a while which could delay plans of everyone trying to add new capacity. For example,

"We ran out of capacity for any additional work about a year and a half ago," said Mike LaRose, CEO of Kuka's (KU2G.DE) auto group in the Americas. "Everyone's so busy, there's no floor space."

The electric vehicle boom is pay-dirt for factory machinery makers

Validation that Tesla was right to buy Grohmann, Hibar, ATW etc - other OEMs buy brands, do joint ventures & spin off divisions, Tesla buys automation suppliers for exclusive use by Tesla (except Curevac commitment) - History of Tesla, Inc. - Wikipedia

 

[johnm6875]

The only thing keeping me grounded is market cap. I'd be thinking 100x if not for market cap. Tesla has everything they need to become the largest company with the largest market cap in history, I'm just struggling with how high I can realistically see that going....

Tesla is shaping up to be just as impactful to the world and will grow much faster than the Dutch East India Company. The VOC had estimated value, at its height, at over $7 Trillion (inflation adjusted). So, a Tesla 10X isn't out of the realm of historical reason.

 

[Jack6591]

I would contribute monies to an advertising campaign. A media campaign to advertise Gordon Johnson’s Tip Ranks profile—throughout the Wall Street area of New York City.

Of course it would have to highlight his appearances on CNBC.

 

[Lessmog]

I would contribute monies to an advertising campaign. A media campaign to advertise Gordon Johnson’s Tip Ranks profile—throughout the Wall Street area of New York City.View attachment 700082

Of course it would have to highlight his appearances on CNBC.

Of course the ad has to go in WSJ, NYT, LAT, Wapo and so on to reach its target audience. Can get expensive. Not to mention tv -- sorry, what the newfangled medium is nowadays.

 

[mongo]

CNBC is promoting this tweet with Gordon;'s false accusation. More proof they have this guy on only for the drama and clicks and views.
View attachment 700017

Here's the thing though: It is actually correct to say the the statement "everyday their cars were getting smarter" is not true. The cars do not learn on their own. The cars only improve when the mothership processes the trigger data, retrains the network, and then ships it out to the fleet.
The true inaccuracy is the first statement "everybody thought..." Which we all know is not the way it works.

So the flow is:
Assert a 'fact' they claim Tesla said (which they didn't) or that everyone knew is true (which isn't) .
Use actual data to show the previous statement was false.
Claiming that the original assumption was a falsehood perpetrated by Tesla as opposed to admitting they (GJ) had a wrong belief.

Since, "everbody thought" is more colloquial than factual, they (GJ) may sidestep slander/libel since the second statement (that the first statement is wrong) is actually true (unless he knew that wasn't how learning worked, then "we" would only refer to "everyone" which doesn't include "himself")

I mean for crying out loud, anyone who follows this at all knows that HW3 cannot train a network. If GJ is only learning this now, it shows a level of ineptitude that makes all his statements ever worthless.

 

[Marc_K_inNJ]

It will be pretty ironic when gas stations deploy TeslaBots as gas attendants, to pump gas...
It will also be very on mission....
The guy in the Honda pulls in, he is greeted by a Tesla. As the Bot hands him back his CC he bends down and whispers: time to get a Tesla...

 

[22522]

It would be nice if the robot could rebuild a mountain bike fork (standard maintenance after 200 hours).

 

[Unpilot]

I plead and beg you all, NO MORE PICTURES of GLJ!

 

[traxila]

Every time you post about this moron GJ, I guarantee he is getting a nickel deposited into his accounts by his employers.

Our nonstop fixation on him makes him yell ”Mission accomplished!” every time he wakes up in the morning and visits this board or its equivalent. His employer’s are probably celebrating as well looking at the traction and bandwidth they are occupying with the pittance they pay this guy. They are getting a tremendous return on their investment.

Can we create a GJ board for everyone who wants to post the daily blather from this self satisfied nincompoop who is bragging right now about how many people listen and react to him?

 

[dm28997]

EVs get higher speed limit in the EU since ICE cars spew out more pollution the faster they go. In Austria air quality improved where ICE limit is 100 while EV is at 130. US should consider this as an added incentive.
I am personally embarrassed by the mainstream media's incompetence; I think a grade school newspaper would have more factual reporting.
What the mainstream media doesn't comprehend is that small lies make all their reports suspect which harms our voter democracy.

Here is the details from Lars on EV speed limits and other info.