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

[UncaNed]

I'd be happy if all the FSD discussions all stayed on their special FSD thread for another year or two.

News Flash:
V12 is a bit different than the previous version. The same can be said of all FSD version changes.

(That said... love to be able to toggle back and forth between V11 and V12 to compare.)

Edit - Granted, it's radically new and amazing under the hood, I am genuinely in awe. Sorry I come across as jaded, it's just... endless FSD posts for so little practical advancement, and Robotaxi many years off yet...

 

[hobbes]

Yeah, if you took WholeMars's word for it, FSD has been basically perfect for a year now.

(I'm excited to see what other people experience over the coming weeks though.)

Yeah, we must assume the scenes are is cherry picked.

That said, what I liked (and I think is better than before):
- Quite in the beginning, it deals well with a fire engine and firefighters in the street, quite narrow (Omar might have helped it with the accelerator though, can´t tell)
- Later, it slows down more significantly for speed bumps than before (don´t think this behavior could be enhanced manually)
- Toward the end, lots of cars parking in the street which are passed with oncoming traffic, at least once it seems Omar is just about to stop it because it gets real tight but it works out fine
- Also many good pedestrian interactions towards the end

BTW this is the original speed version:

 

[elasalle]

+ Regarding FSD V12 .

outside is the same, but inside ...

.... single end-to-end neural network trained on millions of video clips, replacing over 300k lines of explicit C++ code.


... giant leap ...

 

[Drax7]

Yeah, we must assume the scenes are is cherry picked.

That said, what I liked (and I think is better than before):
- Quite in the beginning, it deals well with a fire engine and firefighters in the street, quite narrow (Omar might have helped it with the accelerator though, can´t tell)
- Later, it slows down more significantly for speed bumps than before (don´t think this behavior could be enhanced manually)
- Toward the end, lots of cars parking in the street which are passed with oncoming traffic, at least once it seems Omar is just about to stop it because it gets real tight but it works out fine
- Also many good pedestrian interactions towards the end

BTW this is the original speed version:

Impressive drive, that’s what can help sustain the p/e valuation, the expectation that
FSD will materialize.

 

[raffiniert]

Oh boy and Ben Shapiro is telling him not to worry about Climate Change because people in Miami can just sell property when the city is flooded.

Yeah, whatever he does, he‘s a bad human being. Not around the people you guys want, not your political opinion, yadayadayada

Elon visited Auschwitz to pay tribute. To me, that says all about him I need to know.

 

[Singuy]

Yeah, whatever he does, he‘s a bad human being. Not around the people you guys want, not your political opinion, yadayadayada

Elon visited Auschwitz to pay tribute. To me, that says all about him I need to know.

Elon visiting Auschwitz with a Jewish man to pay tribute is most antisemitic thing on I have ever seen!

---quote from an Elon Hater most likely

Then do a Bingo card with the words "photo op, damage control, nazi shill, Hitler" and you are guaranteed to win.

 

[cliff harris]

I think its significant that V12 FSD is end-to-end Neural Networks, and that makes it worthy of discussion in the main thread.
When FSD was seen as a 'we need better code' problem, then progress was gradual and incremental, and there was no guarantee that it would ever achieve level5.
The immediate quality difference between 11 and 12 suggests that end-to-end really IS the best solution. (TBH I thought this was the goal all along anyway). This is significant because if its really true that end-to-end is the best solution (I think the majority agree with that now?),then solving FSD becomes a data problem, not a hardware or coding problem.

That makes solving true FSD simply a matter of scale, and time. That means that its both inevitable, and something almost impossible to replicate by other companies who do not already have a vast fleet AND a massive data collection infrastructure.

US and European rivals have neither. BYD may have the fleet, but AFAIK not the data collection capability, or the AI talent or compute power.
regular drivers are getting their first taste of FSD12 now. As with everything Tesla, I think we need to wait for FSD13 to see the real potential, but I suspect 12->13 is simply going to be a matter of data collection and processing.
Timelines are always vague but I do suspect that 2024 may genuinely (at last) be the year of major FSD progress.

 

[Artful Dodger]

you "get lucky" and the stock you bought goes ballistic and you have $2m capital to work with

I think then there's also a very good case for selling very, very OTM calls and puts.

Please, please don't encourage folks here to gamble on Options. You may make money, but 99% of the unlucky folks here will just lose their bets.

Please take such talk to "the Wheel" thread.

Thanks!

 

[Knightshade]

I think its significant that V12 FSD is end-to-end Neural Networks, and that makes it worthy of discussion in the main thread.
When FSD was seen as a 'we need better code' problem, then progress was gradual and incremental, and there was no guarantee that it would ever achieve level5.
The immediate quality difference between 11 and 12 suggests that end-to-end really IS the best solution

FWIW that's the same argument we heard for the last couple of major re-writes, that THIS is the One True Solution and previous ones they THOUGHT were just hit local maximums (see for example going from still-frames to video)

. (TBH I thought this was the goal all along anyway). This is significant because if its really true that end-to-end is the best solution (I think the majority agree with that now?),then solving FSD becomes a data problem, not a hardware or coding problem.

That assumes current HW is sufficient both in terms of needed compute, and in terms of camera viewing angles (which esp. for the B-pillar view remains questionable)/ They thought HW2.0 was sufficient at one time- both the cameras and computer turned out inadequate... They thought HW2.5 was sufficient at one time- the computer turned out inadequate (jury remains out on the cameras)... They still insisting HW3 will be sufficient on both but that continues assuming facts not yet in evidence.


Anyway- long story short- it COULD be that THIS major rewrite is finally "the one"--- or it might not be. It COULD be they find a way to get >L2 on the current HW (both HW3s limited compute and the not-great-for-corners side cams) or they might not. In the end, nobody (including Tesla) knows how much of any of these things is "enough" for safe >L2 operation until they achieve it.

Either way, with any discussion of either still being largely speculative, it's probably best left in the dedicated FSD section.

When we've got something genuinely substantive (for example >L2 being launched on existing HW... or an admission they can't do that) it'd be more relevant here.

 

[Singuy]

FWIW that's the same argument we heard for the last couple of major re-writes, that THIS is the One True Solution and previous ones they THOUGHT were just hit local maximums (see for example going from still-frames to video)





That assumes current HW is sufficient both in terms of needed compute, and in terms of camera viewing angles (which esp. for the B-pillar view remains questionable)/ They thought HW2.0 was sufficient at one time- both the cameras and computer turned out inadequate... They thought HW2.5 was sufficient at one time- the computer turned out inadequate (jury remains out on the cameras)... They still insisting HW3 will be sufficient on both but that continues assuming facts not yet in evidence.


Anyway- long story short- it COULD be that THIS major rewrite is finally "the one"--- or it might not be. It COULD be they find a way to get >L2 on the current HW (both HW3s limited compute and the not-great-for-corners side cams) or they might not. In the end, nobody (including Tesla) knows how much of any of these things is "enough" for safe >L2 operation until they achieve it.

Either way, with any discussion of either still being largely speculative, it's probably best left in the dedicated FSD section.

When we've got something genuinely substantive (for example >L2 being launched on existing HW... or an admission they can't do that) it'd be more relevant here.

Honestly if end to end doesn't work then nothing will besides geo fencing and limited deployments.

End to end was always the original plan, however building the software and compute infrastructure around end to end was difficult.

This is why v12 is not beta. Not beta also doesn't mean "complete". It will be as work in progress as cyberpunk, which took almost 3 years since v1.0 to be the best it can be.

 

[GhostSkater]

"What if we made our truck have poor efficiency and a huge range, this way we can use all that excess of batteries we have and Uncle Sam will bail us out anyway"

For comparison, Cybertruck should be rated as follows:

• Dual Motor A/T - 88 MPGe | 38.4 kWh/100 mi
• Dual Motor A/S - 93 MPGe | 36.1 kWh/100 mi
• Beast A/T - 82 MPGe | 41 kWh/100 mi
• Beast A/S - 88 MPGe | 38.4 kWh/100 mi

Might even a bit better in summer tires

 

[Usain]

FWIW that's the same argument we heard for the last couple of major re-writes, that THIS is the One True Solution and previous ones they THOUGHT were just hit local maximums (see for example going from still-frames to video)

I don't think that's true at all. Anyone who followed closely knew that end-to-end was the ultimate goal.

Remember this slide from Karpathy in 2018?

We've always known the ultimate solution was that neural nets would replace heuristics. With FSD 12, Karpathy's dream has finally come true. This is a very significant leap forward.

There has never been an FSD release with this much promise.

Not only is it OK to get excited about FSD 12, it should be mandatory.

 

[Musskiah]

BEVs use a lot more chips than ICE but 2,000 is a stretch. More like 1000 including discrete components which is still a heck of a lot.

I don't have a problem being mistaken if you want to work off of real numbers, here is another fairly reliable source...

“The average electric vehicle has about 2,000 chips, roughly double the average number of chips in a non-electric car,” Raimondo, Commerce Secretary of the US

SOURCE: Chip Shortage Threatens Biden's Electric Vehicle Plans, Commerce Secretary Says

This is, of course, in reference to EVs in particular (as was the Indian article). I am not sure about the number of chips in Teslas in particular. If you have other sources I am open to learning.

 

[Knightshade]

This is why v12 is not beta.

Except, of course, it absolutely is still beta- right in the release notes for every release of V12 both internal and now this first external one.

“FSD Beta v12 upgrades the city streets’ driving stack to a single end-to-end neural network trained on millions of video clips, replacing over 300k lines of explicit C++ code

 

[hobbes]

For now all V4s have a V3 backend, meaning limited to the same 250 kW per stall, there aren't any true V4 Supercharger in the wild, which is interesting

Typically each cabinet supplies up to 387 kW to 4 plugs, but when there are multiple cabinets they are linked in groups of up to 7 cabinets / 28 plugs

What we've seen in Europe in at least one Supercharger (in France if I'm not mistaken) with a lower than 4:1 ratio of plugs to cabinets, likely related to opening up the Supercharger network since some other EVs maintain high power for longer than Teslas

To your question, in my opinion the lack of real V4s in the US isn't a good sign for Cybertruck charge curve improvements soon. Tesla knows there will be a need for likely a next 4680 revision for that to happen, and since that is still far away, no point in expending money in a likely more expensive charger that will not be used soon

I really doubt current Cybertruck being plugged in a real V4 and charging at 800 V will have any difference in the charging curve, if it was better there, it could be significantly better on current V3s

From my post from a while ago

View attachment 1010987

Not sure about the plugs to cabinets ratio, IMHO the main improvement of the V4 cabinets for now is longer cables to accomodate non-Tesla EVs.

 

[mongo]

Not sure about the plugs to cabinets ratio, IMHO the main improvement of the V4 cabinets for now is longer cables to accomodate non-Tesla EVs.

Those are V4 posts with immersed liquid cooled conductors. Cabinets are unchanged. It seems 1000V cabinets will look the same (based on Semi sites).

 

[Singuy]

Except, of course, it absolutely is still beta- right in the release notes for every release of V12 both internal and now this first external one.

V12 is not yet mass released, so until retail version comes out, it can be whatever it wants to be.

Imo v12 will be released as a different name. Fsd 1.0 perhaps. The back bone of v12 will have no more rewrites. This is how I take what Elon means by v12 not being beta.

 

[Krugerrand]

An interesting piece in a Norwegian newspaper:

The headline is: Can anyone dethrone Tesla Model Y in 2024?

They then asked four motor journalists four questions:

1) Do you think Tesla Model Y will be pushed off the sales top in 2024?

They all answer no. But they believe it won't be as far ahead.

2) How much of the market will the Model Y take this year?

Their answers range from 13-20% - depending on interest campaigns and price reductions. (TMY got almost 20% in 2023).

3) Will the Model Y get an upgrade and a facelift some time this year? And what impact can that have?

Maybe. It depends. Model 3 Highland did not take off. Perhaps people prefer SUVs. It may also osborne the current Y.
Exixting Model Ys will still be popular as a preowned car.

4) For those looking for a smaller and cheaper EV in 2024: What should you look for?

Volvo EX30 has lot's of preorders. It's twin Zeekr X is interesting.

Source may be paywalled: Kan noen vippe Tesla Model Y ned fra salgstronen i år? Ekspertene er ikke i tvil.

So why ask car journalists? Have these particular journalists hit the pavement and spoken to thousands of regular people on the street to get a good understanding of what ‘the people’ think?

I’m just trying to determine why these four are the best people to ask these questions of. Why should we take what these four people say in this regard over someone else perhaps not specifically a car journalist. Indeed, is anyone in the world actually qualified to answer these questions accurately?

I mean, these people don’t even actually know when a facelift Model Y will hit the market. Which means they don’t know anymore than the rest of us. So what does that say about the rest of their information and conclusions. Why when asked that question they didn’t answer, ‘We don’t know. Nobody but those working on it at Tesla would know.’?

How did they come by the conclusion that the refresh Model 3 didn’t take off? Where are their numbers to back that up and given various circumstances that have/could have affected those numbers? When did Tesla give them that data?

 

[hd-2]

V12 is not yet mass released, so until retail version comes out, it can be whatever it wants to be.

Imo v12 will be released as a different name. Fsd 1.0 perhaps.

So you're telling me I'm gonna have to specify yet another layer of software?

"I'm currently on software version v12.1 (2024.3.32.12), Full Self-Driving (previously known as Beta) Software v12.1.1, FSD 1.0"

I think we've got enough names in the current software menu

 

[Krugerrand]

I'd be happy if all the FSD discussions all stayed on their special FSD thread for another year or two.

News Flash:
V12 is a bit different than the previous version. The same can be said of all FSD version changes.

(That said... love to be able to toggle back and forth between V11 and V12 to compare.)

Considering how often FSD is discussed as an integral part of TSLA’s future valuation, SOME FSD information should be here. We don’t need play by play, but a general overview of what has changed is helpful, imo.