Market/Economic Impact of FSD and Tesla Robotaxi Network

[ggr]

Mod: Please use this thread for further discussion of this topic. --ggr

 

[ggr]

Mod: copied from main thread for context.

And once more pages and pages of discussions on FSD just to rehash the stale arguments both ways from the same usual suspects with no original thought that would be relevant for investors today. Please mods, make any discussion of FSD off topic and start handing out time outs for those who can’t seem to keep themselves out of those useless back and forths.

Mod:

Yes, I must agree. But here's the problem, there are (at least) three different kinds of posts about FSD:
1. posts about the economics of Robotaxis, Tesla Network, etc. I consider these on topic, although I think we would all prefer that they were in the right (that is other) thread: Market/Economic Impact of Tesla Robotaxi Network
2. Posts about timing and/or feasibility of FSD. As of this instant these are banned as off-limits.
3. Would you put your car in, how much should you make, how will it be run? Also banned as off-topic.

The problem for the moderators is that it's extremely difficult to winnow out the grain from the chaff and then move it when there is so much and such a mix. We just don't have that kind of bandwidth. So @schonelucht's worthy suggestion is adopted... take the good posts over there, don't bother rehashing the other issues, and we'll all be happy and unpunished.

--ggr.

 

[waterloo42]

Edit: I don't think people who have no experience in this sort of thing understand how difficult it is for a computer to recognize objects. Here's a seemingly simple case (with no life or death consequences if an error is made): take a single Landsat 8 image set and categorize all of the parts of the image. A person with minimal training can do it in a fraction of a second: here's a cloud, here's the cloud's shadow, here's water, here's land, here's a road, a city, farmland, a glacier, snow, etc. Now try to write code that does this automatically from a single image set. Impossible, even with many extra channels of non-visible wavelengths that indicate temperature. How do Google Maps and the other "cloudfree" image suppliers manage to overcome this problem and extract just the land portion of the imagery? Brute force. They basically take all the samples of a given 30 meter pixel over several years, sort them by the red channel, then take the median sample as the "real" state of that part of the Earth. I'm simplifying but that's the gist of it.

This is a solved problem for the past several years. At my company we had a side project to do exactly what you're describing -- taking satellite images and classifying them. Not only did we get > 92% accuracy (technically f1 score of 0.89) to detect clouds, roads, rivers, vegetation and to some extent buildings. This was loaded into a Raspberry Pi (i.e. cheap $35 computer) and $100 Intel Movidius stick. The team of about 4 people that did this took about 3 weeks working part-time and were relatively new at machine learning.

BTW, Karpathy was a pioneer in this area and ran the Imagenet competition since 2010. He is solving far more challenging problems than this right now at Tesla.

 

[abasile]

[Musk] imagines that robotaxis will reduce the total number of cars in service. That is just wrong.

Why do you feel that's wrong? Robotaxis, which we expect will be heavily utilized, should displace a larger number of mostly idle, personally-owned vehicles. The total number of vehicle miles traveled, however, is likely to increase.

Please consider Tesla's record on supporting customer service, and ask what the odds are that they can competently implement and operate something like AirBNB. I would say 0.01% odds. Turo already does a good job. Tesla does not know how.

Tesla, under Musk's leadership, has proven adept at learning quickly when it's essential. While Tesla's support infrastructure has been less than ideal for vehicles and home energy, it hasn't been bad enough to kill sales. When the time comes, I think they'll find a way to make Tesla Network work.

Mr Havlak's group now has a total ignoramus (about self driving) in charge, so it will continue to make little to no progress until Baglino moves to another job or is fired.

I'm assuming that Drew Baglino is very sharp and willing to learn quickly. Good technical management skills are certainly transferable across domains.

I do want to know how much in FSD deposits they have taken, because they will have to refund a lot of it.

As long as Tesla continues to make demonstrable FSD progress, and as long as they remain at least competitive with Waymo and Cruise, I think it's unlikely that they'll ever have to return deposits. To prevail against Tesla in an FSD refund lawsuit, I think a plaintiff would have to make a very strong case that Tesla acted in bad faith in accepting those deposits, which I firmly believe has not been the case.

I am beginning to think that the stock price is low not because of Q1, but because Autonomy Day demonstrated that Musk is actually delusional. And many investors don't want to invest in a stock with a delusional CEO.

Q1 was a huge factor, but I agree that Autonomy Day left many investors with the idea that Musk is delusional. To most people, the idea of FSD being ready in less than two years just seems so "out there". Personally, I see Elon as extremely optimistic, but not to the degree that I'd call him delusional. It seems to me that there is a small chance of Level 4 autonomy actually being achieved in 2020, though I wouldn't be surprised if FSD takes another five years. Elon's extreme optimism has been critical in getting us this far, both with Tesla and with SpaceX.

 

[neroden]

I’m so going to relish the day you’ll be eating crow for breakfast, lunch and dinner.

Never going to happen. I, however, will relish Musk's admission that he massively underestimated the difficulty of full self driving and that it's "really really hard" (like he says about manufacturing). I think Musk having this realization is a prerequisite for actually getting full self-driving, so once it happens, watch out, I will become a full-self-driving bull.

 

[tander]

The moon landing became a solved problem as soon as staged, engine-gimballed, liquid-fueled rockets became reality. It was simply a matter of scale.

FSD is not a solved problem. You could run Tesla's software on the world's fastest supercomputer, and it still wouldn't be FSD. The goal is to make it a solved problem.

Imo the moon landing was a much harder problem in the context of 1969 technology v FSD today. But the standards and stakes are a lot different. And unfortunately there hasn't really been a national or international effort for FSD, which is a shame if you consider how many lives it could have saved if Obama or Bush had made a "we choose to go to the moon..." type speech for FSD. I think you could probably consider FSD a solved problem today or even several years ago by moon landing standards if it was only going to risk the lives of a handful of men that had already signed their lives away. The difference is FSD puts millions of people instantly at risk, as well as whatever company is selling it, and it has to be signed off by insurance, regulators, municipalities. So the margin of error has to be tiny before FSD is released for public use, while I would guess the moon landings probably had a much lower standard.

 

[tander]

Full self-driving was being promised in the 1920s by the same fools who didn't understand what was hard and what wasn't.

Sure, it'll happen sometime, if humanity doesn't get wiped out by global warming first (which is actually perfectly likely given how long self-driving might take).

Landing on the moon was basically just a matter of getting accurate calculations and suitable materials, and then spending a lot of money.

That sounds a lot like how you solve FSD too. Not saying it's not hard but going to the moon fifty years ago was a pretty different ball game than FSD today.

 

[Bet TSLA]

Sure there is... just to name a few:

SOLID - Wikipedia

But more specifically with Tesla and FSD you should look to Karpathy and what he says about the topic:

A Recipe for Training Neural Networks

He has quite a bit more to say about it if you go down that rabbit hole... but none of this is unique to Tesla. Everyone trying to reach FSD are essentially using the same machine learning approach. It's the data advantage combined with a big lead in software and hardware which separates them from others aiming for FSD.

Um, no. You don't seem to know what a "first principle" is. It's not heuristics or best practices. It's stuff that if you fail to pay attention to you simply fail. Stuff like gravity and speed of light in a vacuum.

There is nothing like that in software. I can totally ignore the two things you pointed to and write working software. In fact, I have no doubt that 99.999% of all working software falls into that category.

There are some few things related to hardware which are first principles for software running on that given hardware -- things like how fast an interrupt can be serviced. But if that's a problem, in many cases you can just use different hardware.

 

[neroden]

Would we not be more wisely evaluating the impact on Tesla and our investments of a gradual entry in areas with congenial surroundings geographically, governmentally, societally and economically?

OK. I'll bite. Limited robotaxis in a few tightly-geographically-bound areas with particularly well-marked street grids... will have zero positive financial impact on Tesla.

Insignificant at best... financially dangerous if Tesla assumes the insurance risk. Since Tesla routinely overestimates how good their systems are (they're still the best on the market, but they have excess optimism), there *will* be accidents, even in the geofenced case (probably due to really weird stuff happening), and the liability *must* not be Tesla's if Tesla is to remain financially stable.

The availability of "best in class driver assistance features" worldwide for car *buyers* will be substantially more meaningful than some geographically limited (and still unreliable) robotaxi scheme. But all it really does is allow Tesla to raise prices; since Tesla's production limited, it doesn't increase car sales.

 

[neroden]

Somehow in this thread and elsewhere we have completely lost our grasp of reality. Neroden tells us it will be ten years and that the Tesla approach is wrong.

To set the record straight, I said it will take significantly more than 2 years, and with the wrong-background people they're currently hiring, probably 10. But I also said that the Tesla approach is *correct*. It's just going to take a long time.

 

[tander]

You have a major misconception about autonomy...91% percent of Americans are convinced they are better than average drivers. Even if that were possible, it would not be saying much. Because, statistically speaking, humans are terrible drivers. The bar to entry is very low indeed.

They've done studies you know, 60% of the time it works every time!

Afaik, there is no major new tech which needs to be invented for FSD, merely it is a matter of scaling the data collection for areas like traffic lights and parking lots, and then using it to train the NNs, so in a way you can think of it as similar to just scaling up rockets to reach the moon.

As a side note, what is the word for when a problem is solvable - its on the tip of my tongue and is something like tractable??

surmountable?

 

[neroden]

Why do you feel that's wrong?

FFS, we should have our own thread to explain transportation economics. This is textbook stuff.

Robotaxis, which we expect will be heavily utilized, should displace a larger number of mostly idle, personally-owned vehicles.

Common belief. Also totally false.

I really do not want to rehash this for the umpteenth time, it's actually *textbook* transportation modeling. It just flat out doesn't. Things like carshare sometimes displace at *less* than 1 for 1, and never much more than 1 for 1.

The two-word explanation for why is "rush hour". I do not want to repeat the long explanation why because I said it within the last week. If you get rid of rush hour (the Soviets tried, but failed -- but there are some communities which naturally lack rush hour for weird reasons), or serve it with high-capacity trains (which is often done), you can change this. However, in places with high-capacity trains, there are often a lot of people with no car at all, so when you introduce carsharing and they use it, it ends up increasing the number of cars on the road. Robotaxis would do the same.

The total number of vehicle miles traveled, however, is likely to increase.

Tesla, under Musk's leadership, has proven adept at learning quickly when it's essential.

Yeah. Tesla Network is thankfully not essential, which is probably why they haven't learned quickly.

While Tesla's support infrastructure has been less than ideal for vehicles and home energy, it hasn't been bad enough to kill sales. When the time comes, I think they'll find a way to make Tesla Network work.

Maybe it'll never come. I mean, demand for Tesla purchases will keep being very high indefinitely. Who needs Tesla Network? Certainly not Tesla.

I'm assuming that Drew Baglino is very sharp and willing to learn quickly. Good technical management skills are certainly transferable across domains.

Sure. Who knows. He'll have the learning curve of realizing how hard the problem is, though, and he's probably starting from zero there.

As long as Tesla continues to make demonstrable FSD progress, and as long as they remain at least competitive with Waymo and Cruise, I think it's unlikely that they'll ever have to return deposits. To prevail against Tesla in an FSD refund lawsuit, I think a plaintiff would have to make a very strong case that Tesla acted in bad faith in accepting those deposits, which I firmly believe has not been the case.

OK. That sounds wrong to me, though. Do you know the law in this area?

What I know is that you can get refunds from a company even if the company was acting in good faith, if they simply never delivered the promised product. (Generally without interest or penalties.) The question would be, what was the promised product, which would involve reading the marketing text carefully.

Q1 was a huge factor, but I agree that Autonomy Day left many investors with the idea that Musk is delusional. To most people, the idea of FSD being ready in less than two years just seems so "out there". Personally, I see Elon as extremely optimistic, but not to the degree that I'd call him delusional. It seems to me that there is a small chance of Level 4 autonomy actually being achieved in 2020, though I wouldn't be surprised if FSD takes another five years. Elon's extreme optimism has been critical in getting us this far, both with Tesla and with SpaceX.

I don't really see the value of delusion-level optimism, but other people tell me that it's motivating for them. Personally, I can be (and usually am) pessimistic and motivated simultaneously, but I am told that a lot of people don't work that way, so maybe the optimism is valuable for such people. Interesting Human Resources question there. You could be right.

 

[neroden]

OT

The best tech for sequestering carbon is planting more trees (from an economic perspective).

Personally, I think replacing our carbon-emitting concrete infrastructure with carbon-negative magnesium concrete would be the biggest bang for the buck.
Trees are great but with the climate changing it's hard to plant trees that will survive. :-(

 

[neroden]

The insistence that it's either level 5 "I can sleep in the trunk" FSD or nothing is confusing. In looking at autopilot build out, why wouldn't FSD follow the same patterns? Limited situations with monitoring at first, then rolling out more functionality.

Sure, yeah, certainly. But then none of this "Tesla Network" robotaxi financial modeling makes any sense. It's a driver assist system, not a driverless car.

 

[neroden]

At the risk of contradicting you today, this approach you talk about is heuristic.... the old way of doing AI with a bunch of rules.

No, it's absolutely not. Read what I wrote again. This is the *test suite* for whether the car is working properly. This is the instructions for how to *tell* whether the NN is giving the correct answer. This is how the people designing the system *label* the situations. Am I making myself clear?

 

[jhm]

How long will it take to solve FSD? Consider two scenarios:

FAST: FSD is solvable within few years
SLOW: FSD is solvable in a few decades

Under the SLOW scenario, Tesla make disappointing progress on FSD for many years, but continues to bring new models to market and scale up production. Tesla share prices grows toward $1000 over next 10 years.

Under FAST scenario, two possibilities arise:

FAST COMPETITOR: competitor beats Tesla to FSD
FAST TESLA: Tesla beats competitors to FSD

Under FAST TESLA scenario, Tesla dominates robotaxi business and disrupts the rest of the auto industry. Tesla share price goes to $10,000.

Under FAST COMPETITOR scenario: Other competitor dominates robotaxi business and disrupts Tesla. Share price goes to $1.

Given these scenarios and their consequences for Tesla shareholders, I conclude that Tesla must vigorously pursue any approach to FSD that could give rise to a FAST scenario. The stakes are simply too high to risk a FAST COMPETITOR scenario. Thus, Musk must lead the companies as if a FAST scenario has very high probability. Going slow carries an unacceptable level of risk.

Furthermore, the best way to know if potentially fast approach is unworkable is to actually pursue it with enthusiasm. Otherwise, stones left unturned become opportunities for competitors to learn things first that can best be learned by actually turning stones. The premium should be placed not on ideas with a higher probability of success but on ideas a promote a higher rate of learning when attempted.

As Musk has pointed out the only defense is to innovate faster than competitors. Go FAST or risk disruption. Let's hope that FSD is really, really hard to solve because that gives Tesla the greatest expanse of time to scale up.

 

[neroden]

How long will it take to solve FSD? Consider two scenarios:

FAST: FSD is solvable within few years
SLOW: FSD is solvable in a few decades

Under the SLOW scenario, Tesla make disappointing progress on FSD for many years, but continues to bring new models to market and scale up production. Tesla share prices grows toward $1000 over next 10 years.

Under FAST scenario, two possibilities arise:

FAST COMPETITOR: competitor beats Tesla to FSD
FAST TESLA: Tesla beats competitors to FSD

Under FAST TESLA scenario, Tesla dominates robotaxi business and disrupts the rest of the auto industry. Tesla share price goes to $10,000.

Under FAST COMPETITOR scenario: Other competitor dominates robotaxi business and disrupts Tesla. Share price goes to $1.

OK, fair enough.

I think the odds of FAST COMPETITOR are zero.
I think the odds of FAST are pretty close to zero.
Frankly, I think the odds of SLOW COMPETITOR are pretty low too.

 

[neroden]

Mods, could you please move my posts on robotaxi economics to the other thread :sigh:

 

[StealthP3D]

The only thing that matters in the short run for the stock price, is whether Tesla can benefit financially from whatever FSD capability they have.

Doesn't Tesla have a substantial amount of FSD revenue that can't be booked until they roll it out? So, as features become useable and are released, they will add revenue to each quarter in which the features are delivered? In other words, a lot of people have already paid for FSD but the money hasn't been recognized as income yet?

I've always avoided accounting so I'm no expert!

 

[tander]

Doesn't Tesla have a substantial amount of FSD revenue that can't be booked until they roll it out? So, as features become useable and are released, they will add revenue to each quarter in which the features are delivered? In other words, a lot of people have already paid for FSD but the money hasn't been recognized as income yet?

I've always avoided accounting so I'm no expert!

That would be interesting to know how much unaccounted revenue there is for that, they've been selling it for ? years now. It's called the revenue recognition principle, cash can come in whenever, but the revenue doesn't get booked until the product/service is delivered.What is the Revenue Recognition Principle? - Definition | Meaning | Example