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

[Zaddy Daddy]

I'm not an expert here but James Douma seems to think they can squeeze efficiency out of the NN with software engineering optimization as they learn and gives examples of this. It's a long video but the core of his argument starts 20:24.

They can squeeze efficiency out, but not a > 100x reduction anytime soon.

 

[elasalle]

Any guesses on what the FSD subscription take rate going to be?
At 100K -150K take rate … it will be 30-45M $per quarter and none of it will need to be deferred



Got 12.3.4 last night like many here and eager to test it out

 

[elasalle]

Looks Ike FSD training is heating up

https://twitter.com/x/status/1778996541363216892

So they gonna identify edge cases with most disengagements and sent personnel to drive and collect data just like they did in Chuck Cooks unprotected left turns…

 

[mongo]

https://twitter.com/x/status/1779120619579482536

A glimmer of hope.

General note: The judicial process is still going and the share count has not been adjusted.

 

[Musskiah]

Who said zero?

Who said 10x better than human?

Right now, it's about 100x worse than human, and I'm inferring based on some data on how these models scale that getting to human equivalent may take more than 100x in data and model size. Compute too, but part of that can be solved by training for a longer period of time.

@Zaddy Daddy I first want to thank you for challenging the convention on TMC and for being contrarian without being adversarial. It's important to stress test beliefs, notions, and "facts" (I put this is quotes because I recently presented a "fact" about cost/KWh that you corrected and I appreciate). That being said...

This is a very important distinction:

V12 is certainly better than the average human at driving, albeit inconsistently. Don't believe me? Why don't you ask a random stoner 17 year old for a ride then? Or someone walking out of a bar? Or a retirement home? This is why people are calling it "human-like" - which it is. V11 and all prior versions were not. When I gently hold the wheel now, it moves effortlessly in a way that sometimes confuses me into thinking I took over. It's undeniably human-like.

Tesla can see where disengagements no longer occur. I can tell you several of my common routes where disengagements no longer occur - ever. I am certain there are hundreds of Teslas in my area that validate these locations. If Tesla geofenced off the few places where disengagements occur, it would already be possible to traverse a large swath of this greater metro area and ALL of my 100,000 person county where I live 80% of my life. I also believe even with just V11, ALL of the highways in Ohio would be far safer with FSD than humans. That's a big deal.

This is the point: If Tesla wanted to they could give me unsupervised long before they give you unsupervised because, apparently, there are a lot less edge cases here than the insanity you and others live in (no offense, everyone knows the west coast road systems are, well, insane).

 

[Usain]

They can squeeze efficiency out, but not a > 100x reduction anytime soon.

First, I question your earlier numbers. I don't think we are actually at 1000 miles per accident. We really don't have any idea because the data we have on critical disengagements is really, really flawed. I definitely wouldn't base investment decisions on that data.

Also, as I pointed out earlier, we don't know where we are on the improvement curve. From my experience 12.4 is definitely better than 12.3. But mine is a sample size that is much to small to know how much better. So we know the system got better in the short time between releases, but we just don't know much more than that.

It could easily be the case that in 6 months, FSD will be driving far better and safer than any human. It could also be the case that progress starts to slow and Tesla needs to drastically increase one or all of compute, model size, and data size.

We just don't know.

I'll say it again. Tesla knows far more than we do about the rate of progress. The best thing investors have to go on is the statements from Elon and Ashok. They think we are at the beginning of the end.

 

[Skipdd]

I agree I had my 2017 MS with 12.3.4 drive me to my colonoscopy ( ) and it did intervention free and smooth as glass.....then got lost in the parking lot one business over from Dr. office.

Did it drive you home ? Probably best not to answer. Great use case though.

 

[Mengy]

Right now, it's about 100x worse than human, and I'm inferring based on some data on how these models scale that getting to human equivalent may take more than 100x in data and model size.

This is a ridiculous statement, have you even driven FSD v12.3.3? It currently drives EXTREMELY well except for like 2% of the time, once those edge cases are solved for it's going to be superhuman 100%. It's so close right now.

To say FSD is currently 100x worse than a human driver is quite frankly preposterous. You either haven't experienced it or you are exaggerating to push some unrealistic agenda, I'm not sure which is the case here....

 

[Artful Dodger]

This is why my valuation assumes Tesla owns the entire fleet (except SEXY).

Have you priced the entire robotaxi fleet? Even at just $20,000 per car, 10 million cars per year is $200 billion dollars in capex to buy the whole fleet. That's every year. Tesla doesn't have that kind of capital. They need to sell enough robotaxis to pay for their portion of the fleet. This also mitigates the Monopoly and competition concerns which will undoubtedly arise.

 

[primedive]

Is this really how it works? Does FSD train on its own output? That seems...not great? It seems like it would be way better to train it on human driving.

I can see it training on negative output. I.e. where people are disengaging or sending feedback. More training data from FSD does help with that.

 

[nativewolf]

Have you priced the entire robotaxi fleet? Even at just $20,000 per car, 10 million cars per year is $200 billion dollars in capex to buy the whole fleet. That's every year. Tesla doesn't have that kind of capital. They need to sell enough robotaxis to pay for their portion of the fleet. This also mitigates the Monopoly and competition concerns which will undoubtedly arise.

Who would ever agree to finance tesla by buying RT that could be osborned tomorrow?

 

[Zaddy Daddy]

@Zaddy Daddy I first want to thank you for challenging the convention on TMC and for being contrarian without being adversarial. It's important to stress test beliefs, notions, and "facts" (I put this is quotes because I recently presented a "fact" about cost/KWh that you corrected and I appreciate). That being said...

This is a very important distinction:

V12 is certainly better than the average human at driving, albeit inconsistently. Don't believe me? Why don't you ask a random stoner 17 year old for a ride then? Or someone walking out of a bar? Or a retirement home? This is why people are calling it "human-like" - which it is. V11 and all prior versions were not. When I gently hold the wheel now, it moves effortlessly in a way that sometimes confuses me into thinking I took over. It's undeniably human-like.

Tesla can see where disengagements no longer occur. I can tell you several of my common routes where disengagements no longer occur - ever. I am certain there are hundreds of Teslas in my area that validate these locations. If Tesla geofenced off the few places where disengagements occur, it would already be possible to traverse a large swath of this greater metro area and ALL of my 100,000 person county where I live 80% of my life. I also believe even with just V11, ALL of the highways in Ohio would be far safer with FSD than humans. That's a big deal.

This is the point: If Tesla wanted to they could give me unsupervised long before they give you unsupervised because, apparently, there are a lot less edge cases here than the insanity you and others live in (no offense, everyone knows the west coast road systems are, well, insane).

First, I question your earlier numbers. I don't think we are actually at 1000 miles per accident. We really don't have any idea because the data we have on critical disengagements is really, really flawed. I definitely wouldn't base investment decisions on that data.

Also, as I pointed out earlier, we don't know where we are on the improvement curve. From my experience 12.4 is definitely better than 12.3. But mine is a sample size that is much to small to know how much better. So we know the system got better in the short time between releases, but we just don't know much more than that.

It could easily be the case that in 6 months, FSD will be driving far better and safer than any human. It could also be the case that progress starts to slow and Tesla needs to drastically increase one or all of compute, model size, and data size.

We just don't know.

I'll say it again. Tesla knows far more than we do about the rate of progress. The best thing investors have to go on is the statements from Elon and Ashok. They think we are at the beginning of the end.

This is a ridiculous statement, have you even driven FSD v12.3.3? It currently drives EXTREMELY well except for like 2% of the time, once those edge cases are solved for it's going to be superhuman 100%. It's so close right now.

To say FSD is currently 100x worse than a human driver is quite frankly preposterous. You either haven't experienced it or you are exaggerating to push some unrealistic agenda, I'm not sure which is the case here....

I have driven in FSD v12.3. It drives extremely natural, human like! I could totally imagine it already being smooth enough to drive as a robotaxi, as I was totally comfortable with how it was driving as I was a passenger. Lots nuanced tendencies that it picked up to drive like a competent human that in no way could be hard-coded.

But that "except for like 2% of the time" is the key. You cannot trivialize 2% as if it's easy to make that 2% 0.02%.

I agree with the statements that it assuredly works better in some areas than others. I bet it's closer to robotaxi level in the sunny, well organized suburbs of Phoenix than the hodgepodge of SF Bay Area.

I agree that we don't know what the true accident rate currently is, nor do we know what the true ratio of critical disengagements to (serious enough) accidents is. Tesla does but isn't saying. We are left to ascertain.

I get some people want ot trust Musk's bullishness, but I'm not because he's been bullish many times before. Ashok and other employees are more bullish and that's great, but that doesn't give real insight on timelines. If Ashok said he expects it to be robotaxi by 2025 or something, my ears would definitely perk up.

I'm totally willing to update what I think the current accident rate is, based on real discussions, data, and observations from say youtube videos in aggregate.

 

[primedive]

Any guesses on what the FSD subscription take rate going to be?
At 100K -150K take rate … it will be 30-45M $per quarter and none of it will need to be deferred



Got 12.3.4 last night like many here and eager to test it out

I would definitely not purchase at $16,000 CAD or subscribe at $200+, but at $99 CAD, I think it's worth it just to continue playing around it with and see how it improves. I'm not sure why people think that the price is going to go to the moon. Price != value and what matters is total revenue. Tesla isn't going to get there if they price most people out of the market.

Another poster mentioned personal vs commercial FSD and that approach makes some sense ( as well as Tesla possibly licensing the tech out to other manufacturers if it becomes more proven ). I can also see them approaching a geo-fencing approach as if you want to do commercial robotaxi, you really want to lock down a specific area and have it perform almost flawlessly, especially from a safety perspective. Entering the wrong parking lot is not the end of the world, but curbing the car or striking a median definitely is a no-go.

 

[wtlloyd]

Have you priced the entire robotaxi fleet? Even at just $20,000 per car, 10 million cars per year is $200 billion dollars in capex to buy the whole fleet. That's every year. Tesla doesn't have that kind of capital. They need to sell enough robotaxis to pay for their portion of the fleet. This also mitigates the Monopoly and competition concerns which will undoubtedly arise.

This is what I said earlier while you were gone. It seems obvious to me that Model 2 is needed to fund a Robotaxi fleet. $20 billion cash on hand is just a start, even if the cost could be as low as $10,000/vehicle produced. What about all the support, garaging, app development, servicing and maintenance?
If cost is $10K/unit for a $25K M2, every M2 sold buys a RT, with money left over for support.
It will be RT and M2, one for you and one for Tesla.

 

[Artful Dodger]

This is what I said earlier while you were gone. It seems obvious to me that Model 2 is needed to fund a Robotaxi fleet. $20 billion cash on hand is just a start, even if the cost could be as low as $10,000/vehicle produced. What about all the support, garaging, app development, servicing and maintenance?
If cost is $10K/unit for a $25K M2, every M2 sold buys a RT, with money left over for support.
It will be RT and M2, one for you and one for Tesla.

Indeed, for several years I've held that Tesla will be able to purchase robotaxis in the same proportion as the gross margin on the entire vehicle program. IE: if the hardware has a 30% GM then Tesla will be able to purchase about 30% of the fleet production. This self-funding business model is the way Tesla has always conducted business.

 

[Artful Dodger]

Who would ever agree to finance tesla by buying RT that could be osborned tomorrow?

Norges bank? Baillie Gifford? Four Shanghai banks? Black Rock?

Stare at the ditch, that's where you'll end up.

 

[Discoducky]

Had an occasion where it needed to take a right turn on to an on ramp that was the type where one lane from turning right and one from opposing traffic turning left merging into one ramp.

The right turn side closed marked by construction cones, and the remaining side was used by both directions of traffic. However this was unmarked

FSD correctly avoided the closed fork but drove right past the other one even though it would have had the right of way to enter

Hope that wasn’t too garbled explanation

Yeah, 12.x hands over to the 11 stack when transitioning to a class 4/5 road (aka highway). I'd assume this will get addressed in a month or so and the reason it is like this is most likely that the highway 12 stack is still training (yep, it might take over a month to train).

 

[insaneoctane]

I just downloaded FSD V12.4 and did my test loop a couple of times. My overall verdict is that V12.4 is definitely improved over V12.3. This was about one hour of driving with no disengagements and just one accelerator push to keep from annoying the person behind me.

There is a particular left turn lane that V12 refused to enter until now. It was really strange behavior. It was almost as if FSD thought that lane was not driveable space. With V12.4, it's a little hesitant, but it does get into the lane and complete the left turn. This was the one bad failure on earlier versions and now it's at least acceptable behavior with V12.4.

I'm quite happy with this result. It shows me that improvements are coming fast. It's getting to the point where I hate driving myself. I'd rather let FSD do it for me.

This is what I'm interest in most....not 12.4's absolute performance, but the rate of progress of 12.4 vs 12.3 which has been praised so well....I'll keep reading for these opinions.

 

[insaneoctane]

https://twitter.com/x/status/1779120619579482536

A glimmer of hope.

I hope we don't see this....imagine the flipside of this, when Tesla has to correct this ridiculous judgment and ensure that the world's hardest working CEO hasn't worked for years for free. I would hate to see what the reversal of the reversal would do to that quarter's EPS and, thus, feed the FUDsters. Not like we'd see a bump UP or positive articles from the first reversal.

 

[Rotte]

Any guesses on what the FSD subscription take rate going to be?
At 100K -150K take rate … it will be 30-45M $per quarter and none of it will need to be deferred



Got 12.3.4 last night like many here and eager to test it out

I hadn't gotten an invite for my new Model Y Performance and I really wanted to try Version 12.3, so I signed up for FSD for $99. However, I was given version 11 which is NOT why I signed up. Does anyone know a reason for this? How do I get version 12?