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

[Carl Raymond]

I'd like to go to the mall and let the car park itself in the multi-story parking lot.

The cost of the car park is a cross subsidy from non drivers to drivers, levied on every product by every store. We can look forward to the day most people arrive in a network car. While you shop the car takes somebody else home. The car park shrinks every year, till it’s gone and only drop off and pick up zones remain.

 

[MP3Mike]

But really, the point is valid. If you really want to get to 80 mph, why slow down one quarter only to speed up the next? Seems like part of the reason is the hubris not to do a capital raise in 2018. Had they done one, they would not have needed, or probably wanted, to fire a bunch of employees only to hire a bunch of employees a couple of months later. Unless of course all of these employees were terrible, which seems unlikely.

But Tesla probably isn't rehiring the same people into the same positions. They optimized some things so they didn't need as many people, now as things are ramping up they need more people in different areas.

 

[EVNow]

Only if we ignore the exponentially improving neural networks that scale in quality with computing power and fleet size.

I'm stuck by this exponential argument. EM mentioned it in AI podcast but was not challenged.

First of all - what do we really mean by exponential improvement ? Are we talking about # of scenarios solved, % of disengagements (reduction), # of edge cases taken care of ?

Since each edge case needs to be manually solved (yes they can collect data in an automated way, but lot of manual time is still involved in figuring out the edge case, coding it to send the filters to the fleet, once the examples come back - they need to be labelled, heuristics may need to change etc) - that can only be done linearly. The team may gain efficiency because of experience, better tooling etc, but still linearly.

So, how does the driving improve exponentially ? This is an important question because EM's confidence comes from this assumption.

 

[Krugerrand]

But really, the point is valid. If you really want to get to 80 mph, why slow down one quarter only to speed up the next? Seems like part of the reason is the hubris not to do a capital raise in 2018. Had they done one, they would not have needed, or probably wanted, to fire a bunch of employees only to hire a bunch of employees a couple of months later. Unless of course all of these employees were terrible, which seems unlikely.

You slowdown to avoid the mattress in the middle of your lane.

 

[JohnDinger]

Do you have any specifics you can refer to here?

In California, autonomous vehicles have to use "disengagement" as their key safety metric that they report back to the public.

That sucks for Tesla, because there are lots of instances where a disengagement might be totally normal, or even the driver's preference.

Ex: Driver disengages to make an aggressive lane change manually, then goes back on autopilot.

Or driver disengages to speed up to beat a yellow light when autopilot starts to slow down.



In Arizona, autonomous taxi services must be able to safely handle a laundry list of road signs, including odd ones like "No right on red during X hours," detours, and roundabouts. They also have a minimum mileage requirement with professional drivers.

This sucks for Tesla, because it hurts an incremental approach.




Tesla has to start the regulatory battle now.

Data isn't going to be enough. Vaccines are objectively safe and yet we still don't have laws requiring them. Smoking is objectively bad and yet we still don't hit it hard enough. Climate change is objectively happening and yet we aren't passing enough policy to stop it.

They need to open a massive PR effort and stop keeping everything behind closed doors.

The longer they stay behind closed doors, the more they allow Waymo/Google and Cruise/GM and their powerful lobbying arms to frame the policy discussions.

 

[EVNow]

The cost of the car park is a cross subsidy from non drivers to drivers, levied on every product by every store. We can look forward to the day most people arrive in a network car. While you shop the car takes somebody else home. The car park shrinks every year, till it’s gone and only drop off and pick up zones remain.

I don't think there are even 5% who are non-drivers in the malls here.

OT :

BTW, the real subsidy is from people all over the world who are too poor to own a car - to rich people who burn fossil fuel driving around.

 

[neroden]

Read up on INTJ or Mastermind types.

Yeah, I'm INTJ. And I think Musk needs to do his homework, because he's making dumb mistakes by not doing so. He's actually been showing some classic INTJ flaws; all correctable by doing his damn homework, of course.

INTJs screw up when we don't have sufficiently varied data, a bit like neural networks.

They always think as far ahead as they can see. If the whole concept of buying cars and driving them as personal vehicles is going to become obsolete and replaced with TaaS (Transportation as a Service),

It won't; not in Musk's lifetime anyway.

it makes no sense to invest into developing the next pretty BMW competitor with "killer features" of a car that people buy. Just zero sense. If the whole thing is going towards some ubiquitous boxes on wheels that are owned by the fleet operator that is only there to wash them and change tires once every 100K miles, all the fuzz that the current cars have will be fairly meaningless.

 

[EVNow]

It won't; not in Musk's lifetime anyway.

It definitely will - if not for FSD, it will because of climate change and attitude change.

The team I work in has a lot of young people just out of college. Half of them don't even know how to drive. This generation is going to do to cars what Millennials/Gen-X did to Block Buster.

 

[neroden]

But really, the point is valid. If you really want to get to 80 mph, why slow down one quarter only to speed up the next?

Because employees are not interchangeable widgets. Firing a sales employee in Los Angeles and hiring a service employee in Albany makes sense and cannot be replaced by reassigning the LA employee to a totally different job in Albany.

 

[neroden]

OT

It definitely will - if not for FSD, it will because of climate change and attitude change.

The team I work in has a lot of young people just out of college. Half of them don't even know how to drive. This generation is going to do to cars what Millennials/Gen-X did to Block Buster.

Oh, don't get me wrong -- the world went car-crazy in the 1950s and that's reversing. We may get back to the personal-vehicle-ownership rates of the 1890s. However, personally owned vehicles are certainly not becoming obsolete. In rural areas, no credible alternative to personally owned vehicles has EVER been proposed. Rural areas will continue to exist.

People talking about "transportation as a service" as if it's going to take over everything and replace personal vehicles have, uniformly, never lived in a rural area. It's an eyerollingly dumb idea for someone living on a farm. The logistics and economics of it don't work.

(In the old old days, your personal vehicle doubled as a work vehicle, and was called a "horse")

 

[AcesDealt]

The gaming side of me picked up on Friday 200 weekly contracts at 250 strike for the price of 1.39. If I unload those for 5+ this week I'll go ahead and get me a p3d . Someone has to keep them in business. Wish me luck.

Ironically this gamble is on hoping for some expression of financial maturity and derisking by the company this week. That drop last week was a bit like a 'get the Treasury secretary on the phone' moment. Not an advice.

Got my debt raise and I'm still down 50% on these. Had the debt raise come on monday or tuesday I'd be getting a car which will remain driven only by me for many years.

I'm like a Tesla predicting savant. Capable of leaping over most people by using simple spreadsheet cost math, a little bit of critical thinking about Elon claims, and a whole lot of pruning out tempting magical thoughts.

Here's what I believe. Model 3 is gonna sell like hot cakes this year but gross margins will remain 20% or below. The s and x will step up this quarter but get stuck at a 75k/year pace. Tesla energy and solar will continue to be costly distractions that they should have sold off already. Elon will realize at least one major project or line of business on the table should be cut completely or split with a major partner. Elon will have no further run ins with SEC about Twitter.

The biggest surprises will be around the thematic emphasis of Tesla network. The natural approach here is to isolate the real core value and eliminate the distractions. A core point of tension is that the tesla network does not need to be made out of Tesla cars. Someone may very well pay for the privilege of including their cars into that option.

 

[Singuy]

I'm stuck by this exponential argument. EM mentioned it in AI podcast but was not challenged.

First of all - what do we really mean by exponential improvement ? Are we talking about # of scenarios solved, % of disengagements (reduction), # of edge cases taken care of ?

Since each edge case needs to be manually solved (yes they can collect data in an automated way, but lot of manual time is still involved in figuring out the edge case, coding it to send the filters to the fleet, once the examples come back - they need to be labelled, heuristics may need to change etc) - that can only be done linearly. The team may gain efficiency because of experience, better tooling etc, but still linearly.

So, how does the driving improve exponentially ? This is an important question because EM's confidence comes from this assumption.

Not 100% sure I understood the part Karpathy was saying how NN start learning on their own. I assume means no more hardcoding edge cases with a programmer.

 

[neroden]

OT

Elon has started talking about Robotaxis more regularly recently because his confidence has increased and because he sees that the market is attaching far more value to ride hailing companies like Waymo and Uber than it is to EV companies.

The market is completely nuts in its overvaluation of Lyft, Uber, and Waymo, IMO.

 

[EVNow]

Read up on INTJ or Mastermind types.

Not big on psychological types.

Myers–Briggs Type Indicator - Wikipedia

 

[neroden]

Not 100% sure I understood the part Karpathy was saying how NN start learning on their own. I assume means no more hardcoding edge cases with a programmer.

Oh, no. All the real edge cases (flying cars, deer fighting, etc.) will still be done by hand.

What he's saying, IMO, is that they can auto-train the NN for simple stuff by using already-known detection methods, basically. For instance, line / edge detection is pretty close to a solved problem, so you can set up a machine which uses line detection to train the NN on lane tracking. Stuff like this. Saves the hand-labelling time for the more "edgy" edge cases.

 

[Singuy]

Sooo GF 3 is almost done..lol.

RayTesla on Twitter

 

[EVNow]

Not 100% sure I understood the part Karpathy was saying how NN start learning on their own. I assume means no more hardcoding edge cases with a programmer.

You don't have to hardcode. But you need to train the NN for the edge cases.

 

[anthonyj]

Moody's Investors Service on Twitter
View attachment 403460

With enemies like these who needs friends now

Literally prostitution

Sooo GF 3 is almost done..lol.

RayTesla on Twitter

Fake news/CGI/Photoshop/Fraud etc

 

[ReflexFunds]

I'm stuck by this exponential argument. EM mentioned it in AI podcast but was not challenged.

First of all - what do we really mean by exponential improvement ? Are we talking about # of scenarios solved, % of disengagements (reduction), # of edge cases taken care of ?

Since each edge case needs to be manually solved (yes they can collect data in an automated way, but lot of manual time is still involved in figuring out the edge case, coding it to send the filters to the fleet, once the examples come back - they need to be labelled, heuristics may need to change etc) - that can only be done linearly. The team may gain efficiency because of experience, better tooling etc, but still linearly.

So, how does the driving improve exponentially ? This is an important question because EM's confidence comes from this assumption.

Elon's calculation of the speed of FSD progress is definitely key to his confidence here. I've been thinking about the speed of Tesla's strategy for some time and its not very intuitive.

I think Tesla has set its system up for solving every 9 of accuracy to take slightly less staff time and slightly less annotated data than the previous 9. However, every 9 will require a 10x larger fleet to experience enough rare incidents to collect enough data to solve them. This is because for example if the first 9 required data collection, software and training to solve 1000 scenarios, you can solve the next 9 by solving another 1000 scenarios which are on average 10x less common. Every scenario should require a fixed number of annotated training examples and hence staff time to solve. The reason each 9 takes slightly less time than the last is because as the system builds up a larger and more varied set of scenarios, its solutions get more generalised and solving one problem is more likely to improve accuracy on a different problem.

One key bottleneck to progress is whether this system hits a dead end without an understanding of causation. The other bottleneck is if Tesla hits full fleet utilisation rates before solving enough 9s for robotaxi regulatory approval. Currently fleet utilisation rates are 0.1% for each data collection campaign, so Tesla should have about 2000x headroom vs current rates by the time the fleet has doubled next year ( potentially enough to solve another 3 to 4 9s without needing to improve the efficiency of their method).

Note the 9s that Tesla is solving for are not image recognition accuracy. It is something more like probability of avoiding a shadow mode discrepancy/disengagement/accident per mile. This accuracy will be far higher than individual image recognition accuracy because sensor fusion and driving policy both compensate for weakness in the more granular image accuracy rates (particularly if camera & radar accuracy probabilities are independent).

 

[Krugerrand]

Yeah, I'm INTJ. And I think Musk needs to do his homework, because he's making dumb mistakes by not doing so. He's actually been showing some classic INTJ flaws; all correctable by doing his damn homework, of course.

INTJs screw up when we don't have sufficiently varied data, a bit like neural networks.


It won't; not in Musk's lifetime anyway.

Have you considered that it is you who is missing data?