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

[jw934]

Tesla should modify their architecture to include memory of this type.

Sharing a data point…

When I first got my Tesla and drove up Brian Rd to make a left into Pondsview, I had to stop because of traffic. Same on the next two days. After that and for the next 5 months my Tesla always annoyingly stopped there for the phantom stop sign when I wanted to proceed through that intersection. After I got FSD beta my M3 was able to go through there without stopping.

 

[Discoducky]

We need to be clear about what Dojo is and what it is not. Dojo allows Tesla to iterate faster. It does not come up with new techniques to solve real world AI.

The speed of the latter is limited by the speed of human ingenuity.

Dojo is much more than iterating faster. GPUs have proven that they can do that. What Dojo buys you is scalability. The wafer fan out chip design is novel. And yes, GPUs have proven they can scale as well, but Dojo should be able to scale with higher value tradeoffs in TDP (power efficiencies), coding flexibility/agility, model sizes, memory distribution/utilization's and management.

Listen to Ganesh talk at the recent Hot Chips conference. Maybe another way to say it is that with GPUs you need nVidia to work with you, provide you engineering resources and you have to trust they will do that at all costs even when they are pulled by their other business partners. With Dojo, you just need your own engineering resources to move as fast as you can.

 

[PeterJA]

Might it be solved by code which presumes there may be hidden traffic control signage whenever sight line is obstructed while approaching an intersection? The AI could scan for things to build confirmation, like the base of a pole that might hold a stop sign, or, the back of a stop sign across the intersection for the oncoming lane and/or crossing lanes. (absence of crossing lane stop sign would weigh toward there being a hidden stop sign)...

Expanding on this helpful point...

AI researchers have discovered that neural nets can detect patterns that humans didn't know were there. This can seem like magic or "seeing in the dark." Such capability can emerge suddenly when the neural net gets big enough and has enough data to work on.

With Dojo and 3 million data collectors, Tesla's neural nets and data sets are getting very big.

 

[Coolio2000]

Double yikes.

https://twitter.com/x/status/1566444947112218629

"That has never happened before in this part of CA, and I am seriously concerned about human health & power grid integrity if it transpires."

Wow, who knew he had climate data for the last 4.5 billion years for California?

 

[Usain]

So you want people to recognize that FSD missed a stop sign and somehow tell Tesla about itand then Tesla manually adds it into a new stop sign database? Yes, that would work, but would also be kinda of a manual bandaid.

Actually, Tesla does not need crowdsourcing for this, the car eventually recognized the stop sign, just too late. So the car did know about it, eventually. My point was that Tesla should modify their architecture to include memory of this type.

In addition to the "memory" solution, there are two other approaches to solving this.
1. Teach the AI to recognize the stop sign sooner. Eventually, the AI should become better than a human at recognizing occluded stop signs from a distance.
2. Add a rule that says, "If I run a stop sign, hit the brakes."

These two can be done without introducing the "memory" approach. I agree that the system could be made more robust using memory. I just don't know if it is practical to implement.

 

[mongo]

2. Add a rule that says, "If I run a stop sign, hit the brakes."

Hard coded stopping is an attempt at law following; but, once the stop takes the car past the shoulder, may be the wrong action.

From a wider view, if the system is designed to be safe, there is no (non law based) need for an additional rule. The car would already be checking for cross traffic (to handle people running stop signs or lights) so it would either not have entered the intersection (futher) due to that, or accelerate to clear the intersection before collision (if cross traffic was previously occluded).

 

[CorneliusXX]

Double yikes.

https://twitter.com/x/status/1566444947112218629

I'm going to California for the first time in 20 years next week. Looks like I should us my Australian passport instead of my UK one.

 

[Big Time]

Thank heavens the market is closed on Monday. One less trading day till CPI & P and D

 

[Usain]

Hard coded stopping is an attempt at law following; but, once the stop takes the car past the shoulder, may be the wrong action.

From a wider view, if the system is designed to be safe, there is no (non law based) need for an additional rule. The car would already be checking for cross traffic (to handle people running stop signs or lights) so it would either not have entered the intersection (futher) due to that, or accelerate to clear the intersection before collision (if cross traffic was previously occluded).

I agree. There is a short window of opportunity where hard braking is the right thing to do. You don't want the car to end up stopped right in the middle of the intersection. And I agree, if you are past that window of opportunity, the right thing might be to accelerate instead.

I don't know if it would be best to implement this as a hard-coded rule or as a learned behavior. You could definitely train the AI to do this as a behavior.

@PeterJA seems to take the view that the AI will learn this on its own in the next few months (forgive me if I have mischaracterized your point of view). I find that to be unrealistic. I think we need supervised learning for this and for most other edge cases encountered.

Some day, we might be able to just feed the system enough data and it will learn everything about the driving task on its own. But even mighty Dojo won't give us that kind of capability any time soon.

 

[insaneoctane]

Kelvin Yang out with some juicy rumours around the next Shanghai factory. Supporting @Artful Dodger's thoughts on expansion in China. I wonder if this is the factory announcement that Elon has in mind or if the announcement relates to a location without an existing factory?

https://twitter.com/x/status/1566347304407375872

I can see this rumor going either way…on one hand EM has tirelessly expressed how amazing the China team is and their output and margins speak for themselves…on the other hand, COVID lockdowns and import /export restrictions...

 

[Bouboule]

The question isn't simply a "yes or no" for believing in fully autonomous driving -- at least not for me.
It comes down to timeline.

Do I believe Elon has a good chance to achieve it ? (without a specific deadline attached to it) Yes, I do !
He has thrown the best engineering talent at it and spares no expense to pursue it, so if any team can do it, his team will; and I do believe the problem is solvable.

On the other hand, I am skeptical about the "very soon" part. Sure, they are making great progress. I even believe that they will release some version of FSD to the wide public this year. But that version will still be a level 2 ADAS, not fully autonomous driving.
How many more iterations (software and hardware) will it take to make it true "FSD" (remember what the acronym stands for!) I have no clue, but my gut feeling is that it will take many more than what Elon currently projects. There will be many more gotchas just like there was many in the past 6 years of Autopilot development, don't forget his projection about "driving from coast to coast on autopilot" within a year from 2016.

FSD beta might already be fully autonomous on highways, so maybe we are already at level 3, we just don't know it yet. If Beta can do all of this in city driving, it must be pretty great on highways.

 

[Knightshade]

FSD beta might already be fully autonomous on highways, so maybe we are already at level 3, we just don't know it yet

Except, we do know that. Because Tesla told the CA DMV that the OEDR of the FSD Beta program is incapable of higher than L2 and they have no plans to change that in the future which is why, in Teslas own words,-(bold added)

a final release of City Streets will continue to be an SAE Level 2, advanced driver-assistance feature.

(city streets has long been the internal name for what is commonly called FSDBeta- Green was talking about it with that name even before the earliest public betas and it continues to be called that in internal code today)


Then they go on to say that any L3 or higher product they DO release would be a different thing, and have its own entirely separate development/test/validation/limited release/wide release cycle.

Nobody (outside Tesla at least) has seen that product yet.


I know it's the weekend, but this is yet another topic already exhaustively argued in the proper forum so maybe let's not scroll this one with the 378th repeat of already covered stuff?

AI, Autopilot, & Autonomous/FSD

Discussion about AI, Tesla Bot, Tesla Autopilot (AP), the promise of Full Self Driving (FSD), as well as other Autonomous Vehicles.

teslamotorsclub.com

 

[wipster]

Hard coded stopping is an attempt at law following; but, once the stop takes the car past the shoulder, may be the wrong action.

From a wider view, if the system is designed to be safe, there is no (non law based) need for an additional rule. The car would already be checking for cross traffic (to handle people running stop signs or lights) so it would either not have entered the intersection (futher) due to that, or accelerate to clear the intersection before collision (if cross traffic was previously occluded).

Along with sending a vid to the authorities so they can send you a ticket for your car's infraction... you can just send it to Tesla for payment!

/s

 

[unk45]

Like how Nokia in early 2012 had investment grade and Apple was still rated junk.

Nokia still is investment grade:

https://www.fitchratings.com/entity/nokia-corporation-80362398

S&P Global Ratings revised outlook on Nokia to stable and affirmed at

S&P Global Ratings affirmed the 'BB+' Local Currency LT credit rating of Nokia on February 25, 2021. At the same time the rating agency revised outlook to stable from negative.

cbonds.com

They have managed quite well since 1865.
it takes more than an ephemeral cellular telephone ‘flash in the pan” to do them in.
obviously Nokia is ‘marching to a different drummer’.

2012 was good for Apple too. By that time they should have had better ratings, so that much is clear. Nokia was an entirely non-analogous case, although the surface of their retail phone business created the image of failure, the reality was quite different. OT to go in details.

 

[insaneoctane]

It doesn’t seem like the EU electricity prices are impacted a lot by the NorthStream 1 indefinite closure. The spot baseload prices (for tomorrow) are less than 0.40 euro, and most hourly spot prices are also below 40 cents. Unlike the 1 euro per kWh panic at the end of august.

And yet I pay $0.54/kWh from 4pm to 9pm in California....prices set well before Russia invaded Ukraine, and yet the entire EU dealing with having significant amounts of their energy supply cut off by a nutty Russian president can do better. What's wrong with this picture?

EDIT: Technically the price was $0.49/kWh before Russia invaded Ukraine.

 

[insaneoctane]

So you want people to recognize that FSD missed a stop sign and somehow tell Tesla about itand then Tesla manually adds it into a new stop sign database? Yes, that would work, but would also be kinda of a manual bandaid.

Actually, Tesla does not need crowdsourcing for this, the car eventually recognized the stop sign, just too late. So the car did know about it, eventually. My point was that Tesla should modify their architecture to include memory of this type.

Agree with your analysis 100%. This is exactly how human drivers work, they do something wrong or sub-optimal, realize the "oops" (and potential dangerous consequences to that mistake) and behave differently next time. We've been told that FSD sees it's interactions for the first time, every time. Training the NN can improve things, if you identity the error and collect the necessary training data. These are the infinite edge cases.... I like the idea of memory (cloud based?) until the NN can train those mistakes away. Not sure how this would work when Tesla is bragging about FSDs continued independence from map data?

 

[Buckminster]

James is clever breaking this into smaller chunks. People can't hold their attention for long these da

 

[Artful Dodger]

Kelvin Yang out with some juicy rumours around the next Shanghai factory. Supporting @Artful Dodger's thoughts on expansion in China. I wonder if this is the factory announcement that Elon has in mind or if the announcement relates to a location without an existing factory?

https://twitter.com/x/status/1566347304407375872

Some background: (w. links)

• EXCLUSIVE Tesla plans new Shanghai plant to more than double China capacity - sources | Reuters {Feb 24, 2022)
• Tesla planning new Shanghai plant as soon as next month to double its China capacity: report {Feb 24, 2022)
• Tesla Adds Plant To Build 2 Million EVs In China; USPS Commits To Ugly Truck - Autoline Daily 3268 {Feb 24, 2022)
• Tesla to open a NEW Shanghai factory?!? (Ep. 594) Tesla Plans Huge Expansion In China (May 3, 2022)

I think this new Shanghai plant also has implications for the U.S. Inflation Reduction Act. Since imported EVs will not be eligible for the $45/KWh tax credit, Tesla has even less incentive to sell Robotaxi in the U.S.A. Indeed, it reinforces their interest in buying their own production, keeping those Robotaxi's (and their ongoing revenue stream) inside the Company.

Chairs!

 

[Electroman]

Then they are even more misinformed than I thought!

Not really. If you think about it, it makes absolute sense. More colors more credit worthy, less colors not so much. Just finance 101. Frist principles.

 

[thesmokingman]

Oof for arkk. Keep selling winners to buy losers. Even if their thesis is right, that's a lot of catching up to do.

View attachment 848812

I would link the Tropic Thunder video again but I think we all get the point with Ark, lol.