Articles/megaposts by DaveT

[DaveT]

Wouldn't it be that the neural net is already trained... but not validated. Tesla will deploy the trained neural net to the AP2 vehicles and run it in shadow mode, where the AP2 responses will be compared to human drivers. Where there is significant discrepancy, they can do further training with those situations. As the discrepancy rate versus correct human driving drops for particular driving situations, they can then enable those features?

Technically, the neural net's training has already been initiated by Tesla (with their own limited number of AP 2.0 cars) but it's not trained enough to be activated on cars. It basically needs more training and validation, and shadow mode is a way to train and validate.

 

[techmaven]

Technically, the neural net's training has already been initiated by Tesla (with their own limited number of AP 2.0 cars) but it's not trained enough to be activated on cars. It basically needs more training and validation, and shadow mode is a way to train and validate.

The reason I make the distinction is that the original description is that it sounds like Tesla is starting from scratch when the cars are deployed in the field with AP2. As if Tesla hasn't actually done anything, including its own testing yet. Also, I am skeptical that any actual training is being done in the field... instead, I think they do all the training on Tesla's servers, informed with collected data sets and discrepancies.

 

[DaveT]

The reason I make the distinction is that the original description is that it sounds like Tesla is starting from scratch when the cars are deployed in the field with AP2. As if Tesla hasn't actually done anything, including its own testing yet. Also, I am skeptical that any actual training is being done in the field... instead, I think they do all the training on Tesla's servers, informed with collected data sets and discrepancies.

I can see your point. I meant that the neural net needs additional training to be ready for deployment. And I agree that training likely takes place on Tesla's servers but with data from cars.

 

[Fallenone]

The reason I make the distinction is that the original description is that it sounds like Tesla is starting from scratch when the cars are deployed in the field with AP2. As if Tesla hasn't actually done anything, including its own testing yet. Also, I am skeptical that any actual training is being done in the field... instead, I think they do all the training on Tesla's servers, informed with collected data sets and discrepancies.

Neural network needs data to be trained with, a lot of data. Whatever Tesla has internally before this week would be too little. As we can see in the demo video, there's quite a few things need to be improved. And since AP2.0 is not using the same hardware as before, they need to have the data from all the additional sensors to train the new software.

 

[techmaven]

Neural network needs data to be trained with, a lot of data. Whatever Tesla has internally before this week would be too little. As we can see in the demo video, there's quite a few things need to be improved. And since AP2.0 is not using the same hardware as before, they need to have the data from all the additional sensors to train the new software.

True, but when discussing with non-neural and non-tech savvy audiences, it helps to be a bit more specific. The way it is being portrayed makes it appear in some cases that Tesla has assembled the hardware together and has essentially done nothing and will wait to have the cars teach themselves for the next few months for Autopilot 1.0 functionality and for level 5. If that is the case, then how can Tesla deliver on their promises or have any informed view of the future? Instead, the AP2 neural nets have gone through quite a bit of refinement... the setup has been mostly achieved and the initial learning necessary to get to the video we saw has been achieved. Now it has to be validated in the real world and refined, over and over and over. Remember, people freaked out over the beta designation, and computer people freaked out just as much if not more... but they haven't really thought through the development, testing, and deployment process and how a narrow AI system must be deployed.

 

[MitchJi]

The startup cost from Uber purchasing their own cars, plus the combined fleet learning mile advantage from Tesla would seem to make it extremely likely that they'll reach an approved Robotaxi before anyone else...

Five huge advantages for Tesla.

They've making a profit now for deploying their full autonomy hardware, which means:

1. The biggest hurdle for regulatory approval is accumulating a lot of data proving safety.

2. A huge development hurdle is the accumulation of data for training the system.

3. Tesla is making a big profit on that data accumulation process.

4. They're also making an overall profit on selling every car that they produce with the complete safety suite, which is the morally correct thing to do. Compared wiith VW for example who sells the components based on the trim package you purchase.

5. They are using OTA to roll features out incrementally as they validate the safety (using a huge set of data) and gain regulatory approval.

They actually will require almost no regulatory approval for most of the features, because they can all be part of advanced AP. The regulatory approval will be required for making the big change to using those features without a driver.

Compared with everyone else (VW for example) it's an absolutely brilliant strategy. They are getting paid to build a huge competitive moat, and simultaneously providing the safest possible cars at no additional cost to every customer.

 

[MitchJi]

Based on similar conclusions, but Dave has thought of some huge implications, that I entirely missed.

Thanks Dave!

3. Tesla Network
With Autopilot 2.0, Tesla has pulled a few years ahead of the competition. And this is going to help when Tesla releases their ride-sharing network, Tesla Network. I don't think it's going to be easy to take on Uber, but I think Tesla has a decent chance. If Tesla can get regulatory approval for driver-absent Level 5 autonomous driving up to 35mph (ie., let's say in 2018), then Tesla can effectively start scaling their ride-sharing network. Here's how. They can have cars (driver-absent) circle urban areas driving under 35mph. When they pick up a passenger, the passenger sits in the driver seat. If the passenger agrees to be the "backup driver" then the car can go faster than 35mph, and if there's a problem it can fall back to the passenger in the driver's seat who can take control of the car. If the passenger doesn't agree to be the "backup driver" then the car is limited to 35mph, which might be ok for certain urban areas.

They actually will require almost no regulatory approval for most of the features [of full autonomy], because they can all be part of advanced AP. The regulatory approval will be required for making the big change to using those features without a driver.

 

[pGo]

Based on similar conclusions, but Dave has thought of some huge implications, that I entirely missed.

Thanks Dave!

Looks like most of all AP 1.0 data are being reused.. guessing for validation and training the neural network.

 

[Alketi]

Some Risk Analysis (re: self driving)

One cautionary note regarding the new Model S as a participant in a future Tesla Robotaxi service: there's a difference between L5-capable and L5-approved.

Tesla is so far ahead, shipping what they hope will be L5-capable hardware, that they've moved faster than the regulators. The L5 guidelines posted recently state that some form of hardware redundancy is required, but it doesn't yet specify what must be redundant. Is overlapping camera coverage needed, or a mix of sensors (camera/radar/LiDAR), or does it mean deeper fail-safe hardware?

There's an Infineon presentation from their head of Automotive that highlights redundancy of processors and sensors, power supplies, and processors.
Page 12: https://dvcon-europe.org/sites/dvco...pe_2015_Keynote_Road_to_self_driving_cars.pdf

I'll grant that Infineon is hoping to sell as many chips as possible, but the truth is that we don't know what regulators will require.

So, factoring in all risk scenarios toward full autonomy, I see three outcomes:

1. The Model S, as described last week, with redundant forward-vision and radar/camera coverage, achieves L5-capability and is also ratified with enough redundancy for L5-approval. This means that a November-shipped Model S/X could someday join a Robotaxi fleet. This is the best scenario. Total jackpot. Tesla's lead is likely measured in years. Start printing money.

2. Current Model S achieves L5-capability, but, for true driverless operation, regulators require redundancy beyond what current Tesla's possess. Current Tesla vehicles become L4-approved/L5-capable, meaning that while sitting in the driver seat, a Tesla will drive you coast-to-coast, but Teslas won't be allowed to participate as Robotaxis. This is the 2nd best scenario and implies that Tesla solved the hard part (L5) and only needs to rev the hardware to include the required redundant pieces (power supplies, processors, systems, etc).

3. The Model S, as described last week, does not even achieve L5-capability due to corner cases, lack of side/rear radar coverage, etc. This is the worst scenario and would negates much, if not all, of Tesla's early-mover advantage. An even worse scenario would be if Tesla discovered that L5 capability cannot be achieved with just a radar/camera combination and that other technologies are required (LiDAR and/or vehicle-to-vehicle communication). This would basically be the disaster scenario where Tesla nearly moves back to the starting line.

All in all, I think each of these scenarios is in play.

I suspect, and certainly hope, that we land on either #1 or #2. However --

As for #3, it should be noted that Tesla is in the minority in favoring a radar/camera combination, while Google, Uber, Volvo, Ford and possibly others (too lazy to research) all favor incorporating LiDAR. Unfortunately, when dealing with an unknown outcome and one party heads in their own (cheaper) direction, that's not without some measure risk.

 

[DaveT]

"His company, Uber, could essentially be a replacement for owning a car for Kalanick, and also for the 40 million monthly active riders the company has. Kalanick revealed that number on stage at the Vanity Fair New Establishment Summit this year in San Francisco. Kalanick also said that drivers made somewhere between $1.5 billion and $2 billion last month. Those monthly active riders pay around $50 per month, he said."

Travis Kalanick says Uber has 40 million monthly active riders

$2 billion/month revenue is a $24 billion annual revenue run rate, and still growing rapidly. Uber has unlocked a true monster of a business.

 

[TMSE]

"His company, Uber, could essentially be a replacement for owning a car for Kalanick, and also for the 40 million monthly active riders the company has. Kalanick revealed that number on stage at the Vanity Fair New Establishment Summit this year in San Francisco. Kalanick also said that drivers made somewhere between $1.5 billion and $2 billion last month. Those monthly active riders pay around $50 per month, he said."

Travis Kalanick says Uber has 40 million monthly active riders

$2 billion/month revenue is a $24 billion annual revenue run rate, and still growing rapidly. Uber has unlocked a true monster of a business.

Dave,

I like to hear your take on Tesla vs. Uber in the ridesharing space.

With Uber in all the major cities that matter, I do not see much growth from now on. Total market size is no more than about $5 to $6 billion per month in the current manual driver model as labor is the major cost component and I do not see more customers embracing rideshare unless costs go down to say 2x the public transportation cost.

Self driving cars change everything. They will disrupt car ownership and even public transportation .

 

[DaveT]

Dave,

I like to hear your take on Tesla vs. Uber in the ridesharing space.

With Uber in all the major cities that matter, I do not see much growth from now on. Total market size is no more than about $5 to $6 billion per month in the current manual driver model as labor is the major cost component and I do not see more customers embracing rideshare unless costs go down to say 2x the public transportation cost.

Self driving cars change everything. They will disrupt car ownership and even public transportation .

Well, I've shared a lot in the past few months here. But overall, I think people (in general and on TMC) underestimate Uber and their competitive advantages. My advice to those who have the time and openness is to go to YouTube and listen to as many interviews and talks by Travis Kalanick (Uber's CEO). Go back several years, and hear him talk. Understand his perspective. See if his predictions come to fruition. And see over time how Uber executes. Uber, is not an ordinary company. They have very, very smart CEO and have a super lean and enviable business model. And their competitive advantage grows as they get scale and increasing network effects.

That said, it's great news that Autopilot 2.0 has hardware that Tesla expects to get to fully self-driving cars. Tesla needs a way to establish a foothold in the ride-sharing space, and probably only way they can do that is to provide better and cheaper rides for people. Self-driving cars are going to drastically reduce the cost per ride.

The question is how big of a lead does Tesla have over Uber, Google, and others with autonomous driving. Even, NVIDIA seems to be preparing their own turnkey solution of autonomous driving to provide to auto makers. Currently, they are putting the pieces together of hardware, and a software framework. But they appear to be building out that software framework more and more so that it becomes easier and easier for auto makers to quickly develop their own autonomous driving programs. Also, there are startups like Comma.ai, which I think has the right approach (very similar to Tesla's approach).

Before going on, I have to say that I think the two big prizes in transport that are in possible reach for Tesla are ride-sharing and trucking. I don't think manufacturing is that sexy... it's too low margins and too high liability. If Tesla is going to succeed in a big way, they need to win in ride-sharing and/or trucking. The other big prize is energy. All three of these markets are trillion+ dollar markets, namely ride-sharing, trucking, energy.

In order for Tesla to win in ride-sharing, they need to defeat Uber. Again, Uber is a well-oiled machine. Ruthless. Ambitious. And they want to win at all costs.

Tesla's advantage that they can leverage in a ride-sharing war is that of integration. No other auto-maker is integrating self-driving in a such an ambitious and comprehensive manner. In this regard, Tesla is at least a few years ahead of the competition.

However, the competition is growing, especially with self-driving cars. And one shouldn't look at what the auto makers are doing with self-driving tech because that's a distraction. The tech will come from tech and software companies. They will provide autonomous driving solutions to auto-makers. Companies like Uber, Google, Comma.ai, Baidu, NVIDIA, etc. The question is how quickly these companies can package together a compelling self-driving suite to provide to auto makers, and how quickly can automakers integrate it. Even if automakers don't integrate, there's always the possibility of aftermarket integration of self-driving features (ie., similar to what Comma.ai is doing with their $1000 add-on).

Regulatory approval for fully self-driving cars is not going to come easy. In the meantime, Tesla really has no competitive advantage to compete with Uber in ride-sharing. If Tesla debuts a ride-sharing network w/o autonomous driving, I don't think there's any way Tesla's network is going to do well. It will cost more and take more time. People will take Uber, almost every time.

Tesla needs regulatory approval for fully self-driving cars (driver absent), and then they have a chance to take on Uber. Since without a driver, Tesla can provide cheaper rides than Uber (assuming Uber cars have drivers).

One option is Tesla could pursue to get regulatory approval for driver-absent self-driving cars that go up to 35mph. In that case, Tesla can release their Tesla Network and have cars circle urban areas are up to 35mph. When passengers get in, they will sit in the driver seat. If the passenger is willing to be the "backup driver" (ie., willing to take over the driving if needed), then the car can go faster than 35mph to the destination. If the driver is not willing to be the "backup driver" then the car would be limited to 35mph.

This would allow Tesla to ramp their Tesla Network and provide CHEAPER rides than Uber. As they ramp, they can bring down the wait time as well. All while hoping that Uber doesn't get self-driving tech/approval soon.

If Tesla can succeed in pulling this off and win the ride-sharing market, then we're looking at literally hundreds of billions of dollars of market cap added on to the company. The big question is IF Tesla can do it.

Tesla pulling forward the expected timeframe of self-driving cars is advantageous to Tesla because it gives less time for others to react, plan and imitate what Tesla is doing. Nevertheless, companies will copy and follow Tesla... but the key is Tesla needs to be moving faster than their imitators. This needs to be Tesla's competitive advantage. Releasing AP 2.0 hardware and then refining software is a genius (albeit risky) way to push the envelope. If Tesla keeps this ethos, it'll be difficult for others to beat them in the autonomous driving wars. But again, even if Tesla wins the autonomous driving wars, it doesn't necessarily mean they'll win at ride-sharing. It all depends how much in front of Uber they can be when Tesla releases their own network. If Uber is 6-12 months behind w/autonomous driving tech, that might not be enough for Tesla to capitalize and build their network successfully.

In any regard, this is going to be a historic transformation of society, namely moving to self-driving cars. We've got the front seat. And it's going to be one amazing spectacle to see who wins the big prizes of ride-sharing and trucking.

 

[W84M3]

Say Tesla does not have a competitive advantage vs Uber on self-driving cars.

Tesla is still several years ahead of everyone else in the EV world + charging network which drastically reduces cost/km or mile.

The cost drivers of transport - AS IS - are fuel, maintenance of car and driver's pay check.

Tesla will dramatically reduce or nullify (if self-driving) all cost drivers.

Tesla is quickly becoming an established brand. Nothing stops the Tesla Network from being a huge success if they can pump out Model 3's and stay even on self-driving tech with competitors.

 

[WarpedOne]

Tesla Network does not need to wait nor rely on Model 3.

 

[W84M3]

Tesla Network does not need to wait nor rely on Model 3.

No, it does not. But it does to compete.

Mass production + much lower vehicle price makes Tesla Network if you want to compete with a much much bigger competitor.

You won't compete with Uber with a fleet of 1.000 expensive S/X's.

 

[Johan]

With regard to Uber, and its subsidiary Otto, as a competitor in the autonomous space, both passenger ride sharing as well as trucking/goods transport, this was an interesting article I think:

'Driverless' beer run; Bud makes shipment with self-driving truck

 

[Lump]

 

[callmesam]

Want to share a few thoughts on some of the comments above regarding Uber vs Tesla.

First, Uber has a very strong "moat", in my opinion because it's a two-sided market (see Two-sided market - Wikipedia, the free encyclopedia). When you bring two sides of the market together, you enjoy network effects and often a "winner takes all" position. Think about Ebay... is it just a website? No, it's a lot more than that. They bring together sellers and buyers, but they have more sellers and buyers than anywhere else... thus everyone goes there because it's significantly better than other sites.

In a similar way, Uber brings together both sides of the market and gives a very compelling value proposition to each side of the market. To the rider, they're given the most cars around them at the best price. To the driver, they're given the most rides (and thus most money). Anyone else that tries to get into the ride-sharing business will have a very difficult time overcoming this.

For example, let's say you open up ride-sharing app #1 and there's 100 cars within 5 minutes of you. Then, you open up ride-sharing app #2 and there's 3 cars within 15 minutes of you. Are you ever going to open up ride-sharing app #2 again? Probably not.

Let's take another example, but maybe more realistic. On day #1, you open up ride-sharing app #1 and it takes 5 minutes for your driver to arrive. On day #2, you open up ride-sharing app #2 and it takes you 9 minutes for your driver to show up. So, you try it again, on day 3 you try app #1 and it takes 3 minutes for your driver to show up. On day 4, you try app #2, and it takes 8 minutes for your driver to show up. After a while, you stop using app #2 altogether.

What makes app #1 give a higher value proposition (ie., quicker pick up times) is because they have more drivers, but the reason they have more drivers is because they have more riders as well. Both (number of drivers and riders) feed off of each other, and the service gets better and better, thus network effects. It's what makes Ebay the winner-take-all of their market.

Same thing goes for Facebook. Facebook doesn't own anything. They use other people's content. But they bring everything together.

Google as well. Think about Google Search, they use everybody else's content. They just deliver it to you in the most valuable way.

Uber doesn't need to "own" anything to become one of the most valuable companies in the world. Already, Google and Facebook have proved that you can make billions of dollars and not "own" the content... or in Uber's case, the cars.

Making cars, historically, has been a low-margin business. Ultra-competitive and not a very good business model... since you're also liable for future malfunctions of the car... and car manufacturers have to spend billions of dollars in recalls and lawsuits many years after they produce the car. Lot of people might think Tesla is in a great business making cars. But I don't see it that way. They're in a ultra-competitive business that tends to drive margins very low.

The real money maker is in what happens after the car is sold. If you can profit on every mile the car is driven for the life of the car, then you can make a consistent cash flow and it can be multiple times more profitable than being just a car manufacturer.

Now, what happens when autonomous driving comes to the scene? How does Uber compete?

Very simply, they buy autonomous cars. And they also let people "lend" their autonomous cars to their network.

Auto makers won't have much advantage over Uber in this regard but Uber can just buy the cars themselves. Auto makers don't make much margin anyway, so Uber will be paying cheap for these cars. They can then deploy them on their system. Or even cheaper, Uber let's people buy their own autonomous cars and then share them on the Uber network. Uber takes a cut and the car owner takes cut.

Just because Tesla makes autonomous cars themselves doesn't mean they have an inherent advantage in the ride-sharing market. The reason being is because Tesla doesn't have channel where millions of people can hail a ride in a super fast manner. Again, this channel is super difficult to create especially when you have a dominant market player. Think trying to create a social network when there's Facebook, or think about trying to create an auction site when there's Ebay, or a search engine when there's Google. People go to the best. They don't have time to waste. People are going to go to Uber because they have millions of cars ready to pick them up in minutes. They have the most cars and they're the fastest. Tesla can release their own app, but who would use it unless it is better (i.e., faster, quicker, etc) than Uber? And how can Tesla be faster than Uber if they don't have millions of drivers/cars like Uber?

Some might say that Tesla's solution would be cheaper to the end user. But I'm skeptical. Uber could work with a manufacturer and add autonomous driving (assuming that they develop the technology like they're planning) to low-cost cars. A Model 3 is still an expensive car, much more so than a Nissan Sentra or Toyota Corolla. I don't think Tesla would have much cost advantage.

Now, let's take the case where the Tesla/Model 3 releases autonomous driving in a few years and it's the best out there. Uber can simply buy Model 3 cars, or can have people buy them and "lend" them to the Uber network. So, Uber benefits. Sure, Tesla can try to release their own ride-sharing service, but the biggest problem is that they can't start with the scale that Uber has. Perhaps Tesla could start in on small urban area (ie., San Francisco) but it's likely that it would need to be heavily subsidized by Tesla until they can reach the scale that Uber has. In other words, until Tesla can pick up a rider as fast as Uber, Tesla is at a disadvantage.

Here's another thought. It's going to take a while for Tesla to release autonomous driving and get it approved by the regulators. Probably 4 years, best case scenario to get it approved by regulators. In that time, Uber is likely to only continue to grow and saturate the market. If you think Uber has a dominant position now, what until 4 years from now as the ride-sharing/hailing market grows and grows. At that time, their position might be so dominant that it might be futile for Tesla to even try to launch a ride-sharing service. Instead, Tesla might just give up that market and just let owners lend their cars to Uber (or whatever service exists). Actually, owners already will be able to lend their cars to Uber (whether Tesla offers their service or not) since they own the car.

Now, here's the problem if Tesla cedes the ride-sharing market to Uber. Then, Tesla becomes just an auto-maker. Sure, they have other branches, like Tesla Energy. But still in the auto market, they become just a manufacturer. And that's not a very sexy place to be (in terms of making money), since it's just a small part of the overall market of transporting people from point A to point B. And Uber is in likely position to capitalize on the bigger part of that market, and thus realize much more profits than Tesla (or even all the auto manufacturers put together).

There's another battle going on as well, which is transporting GOODS from point A to point B. Uber is already in the lead (over Tesla) as well... since they're already delivering food, etc. Amazon is probably one of the strongest competitors here, as they ramp up their own transportation services as well. But transporting goods from point A to B is one of the world's largest markets, and whoever can disrupt this will become one of if not the largest company in the world. So Uber getting into trucking (with the acquisition of Otto) is not good news for Tesla or Tesla investors. Sure, Tesla might win. Heck who knows who's going to win at this point, but it's going to be a heated battle with lots of casualties. And currently, the odds are against Tesla being successful in the ride-sharing market against Uber or in the transporting goods market against Uber/Amazon/etc.

Unless, Elon pulls something amazing out of his hat.

Care to revisit any of your thoughts in response to Tesla's self driving hardware, Tesla Big Rig and "Tesla Network" announcements?

 

[viKtor slamet]

"His company, Uber, could essentially be a replacement for owning a car for Kalanick, and also for the 40 million monthly active riders the company has. Kalanick revealed that number on stage at the Vanity Fair New Establishment Summit this year in San Francisco. Kalanick also said that drivers made somewhere between $1.5 billion and $2 billion last month. Those monthly active riders pay around $50 per month, he said."

Travis Kalanick says Uber has 40 million monthly active riders

$2 billion/month revenue is a $24 billion annual revenue run rate, and still growing rapidly. Uber has unlocked a true monster of a business.

DaveT, agreed that Uber has unlocked a true monster. However, the $2B/month in revenue might be subject to disruption with full self-driving car. Uber currently charges ~$2/mile (roughly, taken into account $/min, base fee, safety fee etc2). But that $2/mile rate was because there was a need for humans to operate the car, and we know how expensive labor cost is.

And truth be told, Uber ride is very expensive... Tried a few times commuting to work (~25 miles) and it charged me ~$45-48 per ride. Thankfully, I am on T-Mobile and it has T-Mobile Tuesday which give $15 coupon for Lyft. So I ditched Uber for Lyft because it was $15 cheaper.

With fully self-driving car, we can reduce/eliminate the need for a human to operate the car, thus Tesla Network could charge way less than $2/mile. if Tesla can reduce it and charge only $1/mile, imo, Uber is toast. Uber might as well buy every model 3, and help operate it over Tesla Network. But again this is an IF. Hopefully Tesla can execute well.

 

[palmer_md]

...
With fully self-driving car, we can reduce/eliminate the need for a human to operate the car, thus Tesla Network could charge way less than $2/mile. if Tesla can reduce it and charge only $1/mile, imo, Uber is toast. Uber might as well buy every model 3, and help operate it over Tesla Network. But again this is an IF. Hopefully Tesla can execute well.

Uber won't be able to use Tesla cars. The way the option for full self driving is written, it is only for pickup and drop off of friends and family, and not for revenue generation. Pretty much tells the Uber/Lyft and other companies that they will have to use a different source if they want to compete.