Tesla reveals impressive new details on autonomy plans

Topic: 
Tags: 
Elon lays it out at presentation.

I have to hand it to Tesla. Their Autonomy day presentation was very good, and has boosted my estimation of the quality of their program. There is still lots to unpack -- and lots more everybody has to do, but I have the first of what will be several articles up with a summary of what Tesla talked about today, and where they are going.

Tesla bets the farm on autonomy with impressive presentation is on Forbes.com

Comments

I was really glad they let the tech guys talk. The chip is what I expected, but Bannon gave some helpful stats on it. Karpathy's examples of how they use the fleet to collect data were fascinating. Before this we just had some hints hackers had uncovered.

The best thing about Tesla being so aggressive is it will kick Waymo into gear. At least I hope it will. Waymo One is a joke and the whole effort is starting to look like Xerox Star.

Musk's anti-LIDAR stance is dumb. If each robotaxi will really generate $200k++, why not spend an extra $5k to handle edge cases (like night, ha) where vision+radar falls short? Might also help get regulatory approval faster. Doesn't make sense to risk losing a trillion dollar market over a personality quirk.

I'm glad (and somewhat surprised) to hear you were impressed.

The Tesla Network numbers are interesting. My first thought is that they are way off (even assuming that the software will be ready). However, a lot will depend on whether or not Tesla can (legally and practically) forbid people from using their Tesla in someone else's ride-hail network. If they can, this will be a very interesting business model: Sell the cars for little more than cost, and make pretty much all your profits on your 25-30% cut on ride-hail services.

Interestingly, if that plan works out, these Teslas you buy today possibly will go up in value. They could go up in value, because the cost of new Teslas goes up, because the cost of batteries skyrockets as demand outstrips supply.

Of course, as I alluded to in my third sentence, this is assuming that the software will be ready. It also assumes that the software is ready before there is any significant competition from other companies. Competition will take away most of those profits, both from Tesla owners and from Tesla itself. If batteries really do become the bottleneck, then competition will offer robotaxis with smaller (or no) batteries.

And competition is just a matter of time. Once the first company comes out with a successful robotaxi service, others will figure out how to do it too. If Tesla proves that its method works, many others will follow. I don't think they've got a patent on not using lidar. :D

Tesla claims others can't do it, because they don't have a large enough fleet. They see a virtuous circle, where having the largest fleet generates the most data which makes their FSD software the best and leads to them selling even more cars to make their fleet even larger.

There's a lot of fantasy in this, but that's their story.

Here's the real question. If their cars will generate $200k+ of profit over their useful life why would they sell them at cost? Heck, why sell them AT ALL? Not only is it financially stupid, but it delays the transition to EVs because many buyers won't participate in the network or will only partially participate. Tesla is wasting precious batteries on these cars that don't displace as many gascar miles as a dedicated robotaxi would.

Tesla makes a big deal out of changing their lease terms so people must return their Model 3s after 36 months to Tesla who will then use them in their self-driving fleet. Cheap cars for Tesla, hurrah! But why are their lease terms so horrific? If Tesla really wants to get cars back, shouldn't they make the lease irresistible?

Their actions speak a lot louder than their words.

Selling costs slightly above (not at) cost makes a lot of sense. There's no way Tesla could raise enough capital to build hundreds of thousands of cars, and drive them billions of miles, without selling those cars.

Once Tesla actually comes out with a car that can drive itself, with zero human intervention, that will change, of course. At that point maybe Tesla will stop selling cars to consumers. But not before then. Right now Tesla is getting customers to put up most of the capital *and* to act as safety drivers, all over the world, for free.

The lease is a gimmick. Tesla doesn't have enough cash to offer a good lease deal.

The yield on a Tesla bond is over 8%. They can't afford to not sell cars, and they can't afford to offer good lease terms.

Buy the SR+ lease down by 10k. Instead of ~3k upfront plus 500/month it becomes 0 upfront 300/month. If they build 300k Model 3s the next 12 months that's a 3b investment to capture:
200k NPV * 300k cars = 60 billion of value

Heck, they could get 50% participation at 1k upfront and 349/month. That's only 1.1b to capture 30b of value.

But as can be seen, the stock market didn't agree. I don't believe the market has this priced in already. The doubt is around whether Tesla can actually pull this off, or whether other competitors like Waymo, GM etc. have a better shot at it, or will limit how much Tesla can do.

What did the stock market not agree about?

If every car Tesla sold was worth $200K instead of $40K soon, investors who believe that would be massively pushing up the price.

It's hard to imagine how that would even happen. If Tesla is still selling new cars for under $50K, you won't be able to sell your used one for $200K. So is Musk saying that Tesla will stop selling equal or better cars for under $200K in the near future? Maybe you'll still be able to buy a Model 3 for $40-50K, but if you're not grandfathered in you'll have to pay an additional 150K to get access to the Tesla Network?

That particular claim I take to be puffery. But I do like the idea of the Tesla Network. It's a good option somewhere in-between owning your own robocar and relying exclusively on someone else's robotaxi service.

Well, one way they could go up in price is if Tesla says, "Teslas now cost $100K more than before -- BUT we promise you $150K in income as long as you put them in the network a certain amount of time."

But of course, in that case, just sell them for the same price, and take a larger percentage of the revenue.

They can prevent use in another network easily. The only way you will be able to give a destination to a car, to send it to a rider or take them to their destination, will be via Tesla's network. (Not "Tesla Network" but the actual connection to the car from Tesla HQ over the data networks.) It will be very easy to see if you are constantly sending your empty Tesla to pick people up. Sadly, you don't fully own your Tesla.

Why was I impressed? Because in the past they have mostly blown smoke and said how great they were while showing many of the signs of people being fooled by their own hype. Now they are ready to talk about real details. They are far from finished, but they showed some real working solutions to real issues, like training for distance measurement and real details on shadow testing. There are scores of companies all saying they can do stuff. Only the real companies are putting how they do it up for some scrutiny. But they still have much further to go than they claim.

"Teslas now cost $100K more than before -- BUT we promise you $150K in income as long as you put them in the network a certain amount of time."

Sounds like a security, which would require approval from the SEC (and they wouldn't get approval). (Frankly, the claims they're making already are probably going to get the SEC's attention. They're selling Teslas as an investment product.)

That said, if the car can actually generate $150K in income, without the owner doing very much work, prices will go up, or shortages will occur. That's the law of supply and demand from Economics 101.

The "law" of supply and demand is wrong more than it is right.

I'd love to hear why you believe that.

Because when demand increases, this causes technological development which increases the supply, often so much the price actually decreases as demand goes up. Especially when computers are involved, but also things like natural gas and many natural resources. More demand for robotaxi service causes more players to get in the game and invest in improving them and making them cheaper.

I guess that belief explains how you come up with such different estimates than I do. I don't think it's too hard to see why that couldn't be true, though.

Why bother conserving energy if as demand goes up, costs go down? Why induce carpooling if as demand for roads goes up, costs go down?

What about the cost of technological development? That isn't free.

I didn't say it happens every time. Just a lot. It's in a particular class of problems where the pattern can happen.

And indeed, with the cost of natural gas going down, people have indeed not be conserving it.

There are definitely intricacies to how supply and demand works. But if Tesla is selling a car for $50,000 that can generate $150,000 in net revenue, in just 11 years, people will keep buying them until they run out. And at some point there is necessarily a limit to how many Model 3s Tesla can make for $50,000. If nothing else, the world runs out of affordable cobalt at some point. (And yes, technology will eventually overcome whatever shortages occur. Eventually Tesla will be able to make cars without cobalt. But technological innovation takes time and isn't free.) (And cobalt was obviously just one example. Every resource that goes into making a Model 3 runs out, or gets prohibitively expensive, eventually.)

Another factor to consider is that the revenue per Model 3 eventually goes down as people buy more Model 3s. This could be because prices go down, or because miles per car goes down, or both.

The idea that Tesla could continue selling cars for $50K that can reliably and without effort generate $150K over 11 years is nonsense. It doesn't make sense.

(I suppose another way this could happen would be if interest rates went up to something like 20%.)

We have yet to learn what the demand for robotaxi rides will be per city. It will vary on the price of rides, the time of day and it will probably increase over time. Of course, we learn some things now from Uber (At $2/mile) and cabs (at $2.50/mile.)

There are also cities that are starting to limit the number of Ubers operating and they might limit the number of robotaxis. There isn't super strong reason to limit the robotaxis as long as they are legally parking and paying for parking when they do it, but there may be political reasons to do it.

I say this, though in the belief that generally there should not be free parking in any area with parking contention, certainly not for robotaxis. It might not be expensive -- pennies per hour -- but it should not be free. That can put a limit on how many robotaxis eat up parking. They will not cruise from rides, but they will do occasional repositionings.

When there are too many in an area, revenue per hour for all goes down. However, there could still be a problem. Even if a football game is getting out and people need 25,000 robotaxi rides, and there are 50,000 robotaxis, they might all want to go there because while only 50% will get a ride, that may be better than the odds elsewhere in a saturated city. Cities may want to regulate that, even beyond jacking up the parking. (Though the city may feel happy if cars will pay higher rates for parking to get no ride.)

I would like to experiment with spreadsheets but we must guess as to both demand and supply. But if demand is high, supply will keep increasing to meet it.

It's not clear to me what your last sentence means. Do you mean the curves will shift, or that we'll be on a different point on the curve, or do you have a completely different theory of what "supply" and "demand" even mean?

Maybe you're talking about quantity supplied and quantity demanded? But then, at what price?

Or are you shifting the price, and simply shifting along the supply curve?

Do you mean that the supply curve will shift over time?

The law of supply says that if you increase price, and hold everything else constant, that the quantity supplied will increase. Do you reject that law? Do you think there's an exception relevant to the supply of Teslas? If Tesla could get twice as much money per car, do you deny that they could produce more of them?

The point is that in most real world problems, it is meaningless to "hold everything constant" because changes in price and demand cause external factors to change, such as underlying technology, marketing methods and the very nature of the product and the rules governing its supply itself. The pure rule might still exist, but is meaningless, or very temporary.

In this very space, consider Uber surges. They start off looking like the traditional rule -- demand increases, supply is seemingly fixed, but then Uber declares a surge, raising prices. At first this matches demand to supply, but soon it increases supply as drivers respond to the surge pricing and go to the surge area. Then the demand drops and there are actually too many drivers and the surge price goes away and drivers start sitting idle so they take themselves out of operation. It's much more complex than the basic theory states. And it's different for robotaxis which don't mind sitting waiting if they don't have a task, though the owners slightly mind the small cost of parking and having an unused asset.

But the prediction that the price of a Tesla goes up because it can make a fortune as a robotaxi only can work for a short time, because more people buy Teslas to do that if this is the case, and more other cars also become available a robotaxis, until eventually there is a fleet large enough to meet demand with a modest but marginal profit when it becomes more like a commodity. Peak demand will probably still exceed the supply of full-time robotaxis, but that's OK because if demand goes over that threshold, the non-full-time robotaxis start coming into service (like Uber surge but without calling you away from dinner to drive) and there should be a huge supply of them. The one exception would be things like a disaster evacuation, which I have written about, where economic rules are usually not applied for moral reasons.

I agree with most of that.

The price of a Tesla will go up if you can reliably make a fortune with them. But it's unlikely that you will be able to do that.

The price of a used Tesla doesn't go up much unless Tesla increases the prices of new Teslas. They are not going to suddenly make a Model 3 cost $100K. They could make a model 3 that is not allowed to be in the network cost $40K while one that can be in the network costs $100K, I guess. I doubt that is the plan, though.

If they don't increase the price, and you can reliably make a fortune by buying one, then they'll run out.

But, again, it's unlikely you'll be able to reliably make a fortune by buying one. Among other things, 90,000 miles a year per car wouldn't last as the number of cars competing for passengers rises.

Well, they won't run out, because there are not a fixed quantity. There is a limit on how many they can make because they are (mostly) human but Tesla broke through everybody's predictions on what that was before. Much more modest price increases allow them to tool up new factories.

But you outline the real reason they don't run out. The world is not ready for anybody to deploy millions of $1/hour robotaxis quite yet. Regulations won't allow it in most places, and in the places they do, the market can reach saturation.

The funny thing is, if they really are fast guaranteed income, it is silly to Tesla to sell them at all. That's why the other companies like Waymo/Cruise/Zoox/etc. all plan to own their fleets. They also want to own their fleets so they maintain more control of them, especially in the early years when things move fast and you want that control for safety reasons.

Though it will be interesting the first day that Tesla finds a safety bug and decides to say, "Sorry, no autopilot or network operations until we have the fix ready to roll out." This is a nightmare for all the companies, but if you have partners, may open you for lawsuits from them.

There is a limit on how fast they can make them. So they can (and, in the recent past, have) run out.

As far as it being stupid for Tesla to sell them, I thought that too until I heard their plan to force owners to use their service exclusively. After hearing that, I think they're right to sell them. They can make just as much if not more money that way, and without as much in capital requirements, and without hardly any risk. Probably they'll make significantly more money that way. It's a great plan. As long as they can get the contracts to be legally enforcible. Lots of issues with them. Antitrust regulations might be the scariest.

As far as risk from lawsuits by owners, it's easy to both get owners to waive such lawsuits and to get them to agree to arbitration so as to avoid class action lawsuits altogether. One stumbling block would be if the SEC decides that they're selling access to the network as a security. That would also be relatively easy to avoid, but only if you can get Elon Musk to shut the heck up. Maybe. If the liberals take over, and it seems like they probably will, innovative companies are going to have a hell of a time making their well-deserved profits.

Interestingly, if any small- or medium-sized company figures out how to make a robocar using cameras, Tesla will be a great partner for a buy-out/merger. And the problem of getting jurisdictions to let the cars on the road is probably being overblown. These cars will be so much safer, and so much more efficient, than traditional taxis, that the majority of them will have no choice but to let them on the roads. Within less than a year of building the product Tesla will have billions of miles of real world data proving the safety. In the USA I wouldn't be surprised if the federal government steps in and preempts state laws that don't allow robocars. I'm not sure how the EU works but maybe the EU can do that too.

If Tesla is right and all you need is cameras, they won't be alone for long. There are others who are vision focused, including Intel/MobilEye who has a larger fleet than Tesla (though they have less control over it.)

And all the other companies, like Waymo, are going full bore on computer vision too. I would not be surprised to learn that Waymo has more people working on computer vision than Tesla, and that they are better people, and that they have use of Google's best-in-world neural network resources, and the TPUs which are, in spite of Musk's statement, more powerful neural network chips than the Tesla chip.

The difference is those teams are trying to make use of lidar data (using neural networks on it as well) to improve their results. Elon thinks that is making them sloppy, that it's a crutch. He had better hope he's right, but it won't take them long if they decide they agree.

Is there any serious doubt that you can solve the problem without lidar? Clearly you can. (Not "just cameras" though. To mimic humans driving you probably need microphones and you almost surely need accelerometers. Radar is probably, like lidar, a crutch.) The question is whether or not lidar helps you take a shortcut. I think Anthony Levandowski explained why that doesn't work (the problem is prediction). Maybe lidar was useful ten years ago, but it's getting less and less useful over time, as hardware and software advances are made.

No doubt Waymo and many others have better people at the software level. They are the Xerox PARC of autonomous vegicles. Brightest minds software-wise; terrible business plan.

As far as being wrong, it is much easier to add lidar to a system that needs it to get a little bit better perception than it is to remove it from a system that relies on it. It's not clear why you need it, though.

Yes, there is serious doubt in two ways. First of all, nobody has yet solved it without LIDAR today. So the core debate as to when that might happen.

The next question is how much better you are with vision plus LIDAR compared to vision alone. Nobody is doing LIDAR alone any more. So the real date to examine is when vision gets close enough to vision+lidar that it is no longer worth the expense of the LIDAR. Most are predicting that's too far away, or too uncertain, to gamble on not being ready until that happens. They want to be early in the game and win.

Lidar is vision, superhuman vision. Especially Doppler LIDAR which is even more superhuman (and helps greatly with prediction.)

At a very minimum, LIDAR lets you take your camera image and know, for each group of pixels in it, exactly how far away they are and possibly how fast they are moving. It gives 100% reliable segmentation from the background (at lower resolution.)

The question everybody else has answered differently from Tesla is, "why wouldn't you want that?" Tesla doesn't want it because of cost. There is no LIDAR they can deploy on production Teslas today.

Yes, there is a debate as to *when* the problem can be solved without lidar. For that matter, there's a debate as to when the problem can be solved with lidar. And there's also a debate on what exactly the problem is.

But I don't think there's a serious doubt that the problem *can* be solved without lidar.

I don't think it's accurate to say that lidar is superhuman vision. Lidar has both advantages and disadvantages to human vision. The disadvantages are such that you *can't* safely drive with lidar as your sole source of vision. One advantage is that it's easier to process than data from cameras. But hardware and software advances have largely closed this advantage already. The other advantages are almost exclusively in areas that simply aren't relevant to the kinds of driving situations that a robocar is going to encounter. Yeah, you can drive at night with no headlights. But it's illegal to drive at night with no headlights.

You don't need lidar to determine, within an acceptable margin of error, how far away something is. You can also do it with cheap cameras, good hardware, and good software. https://mileyan.github.io/pseudo_lidar/

Yes, Tesla doesn't want lidar because of the cost. They're not the only one making that choice, but they're by far the biggest and most successful one making that choice. Cost is very important. As you note, there is no lidar Tesla could deploy on production Teslas today.

And here, maybe, we should define what the problem is. Because if the problem is building a car that can drive unmanned in 99% of the places that 99% of drivers in first-world countries drive today, then I'd say the answer might very well be impossible to solve *with* lidar, and only possible to solve without it. Not for technical reasons, but for economic & business reasons.

If the problem is to drive unmanned in 50% of the places that 1% of drivers in first-world countries drive today, then yeah, lidar might be the way to go.

If you don't like the term superhuman, because it is lower resolution and range, pick your term of choice. While Google drove purely on LIDAR (except to see colours of street lights) for several years, nobody is planning to drive with just LIDAR or even just LIDAR+Radar.

I don't see a need to pick a single word for it, as my multiple sentences described it just fine.

If the only criterion for adding an expensive sensor that does things that humans can't do, there'd be no limit to what unnecessary sensors would be added to the cars.

released today. MIT robo-taxi service have two main economic flaws.

A new paper out Monday, written by researchers at the Massachusetts Institute of Technology and exclusively shared with FT Alphaville, agrees. It suggests that, at current prices, an automated hive of driverless taxis will actually be more expensive for a consumer to use than the old-world way of owning four wheels.

/"https://psyarxiv.com/6e94h"/

Add new comment