Welcome to Tesla Motors Club
Discuss Tesla's Model S, Model 3, Model X, Model Y, Cybertruck, Roadster and More.
Register

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

This site may earn commission on affiliate links.


Eh, Mobileye has patents on estimating distance from camera input going back to at least 2006.

Much like how everyone here was saying for years the "reason" Tesla could never read speed limits on AP2+ was because Mobileye held the patent on THAT, there's lots of ways to skin the same cat....

And a lot of very generally worded patents (especially ones that are doing variations of things already patented previously) aren't necessarily a roadblock to anyone- the USPTO folks are generally....very not technical.... so it's not uncommon for patents that shouldn't be granted due to prior art or being overly general but using words they don't understand getting granted anyway, Most commonly, if this results in lawsuits or anything at all, they end with either the newer patent being tossed if it's REALLY egregious, or if it's genuinely something different but not enough so it'd be easy to convince a jury it's different, in some kind of cross-licensing settlement that isn't usually super material to anything other than large bills from IP lawyers. (exceptions exist of course)



All that said- depending on the numbers, and how long Elon wants to keep it exclusive to Tesla- there's also always the option to just license the tech directly to other car makers.

Which is more in legacys wheelhouses anyway, and might be cheaper to quarterly results than a huge moonshot "copy FSD" program would.


Tesla can't build 80 million new cars a year.

Is RT revenue so great they'd rather keep it exclusive for only a few extra years- versus collect licensing fees on FSD tech in everyone elses cars for a LOT of years?
 
Basically it’s a patent for how to associate radar and camera objects as Karpathy showed here at 2:19:51
 
Buying big printers and just print them? And buy all the AA batteries in the world to power them millions of vehicles?
The reason I gave an example of VW is they already manufacture millions of vehicles each year, and as Elon Musk said, vast amounts of manufacturing are needed for robotaxis. And to be clear, specifically in regards to copying Tesla's approach to improving autonomous driving, the data collection can happen on non-electric vehicles, so there isn't a strict dependency on batteries.
And they could buy them all themselves... So no one will know if they lied about E-MPG or whatever.
You probably somewhat jokingly said that, but actually if VW or anyone else intends to build and potentially operate a fleet of robotaxis, they'll quickly realize only vehicles with low ongoing maintenance and operating costs will be competitive in the long term. So this will be even more incentive for others to transition sooner to sustainable energy.
 
If VW group really made the decision to copy Tesla's approach in installing FSD-capable hardware on all vehicles, they could deploy ~10 million future-robotaxis each year.
Probably not possible because they have stockholders that expect profits and dividends. If they invest enough to do that, goodby the 50 cent dividend and most or all of the profits. Remember, that unless they change to in-house most things the way Tesla does, they will have to program for the many different protocols from their suppliers. It just won't happen.
 
Eh, Mobileye has patents on estimating distance from camera input going back to at least 2006.

Much like how everyone here was saying for years the "reason" Tesla could never read speed limits on AP2+ was because Mobileye held the patent on THAT, there's lots of ways to skin the same cat....

And a lot of very generally worded patents (especially ones that are doing variations of things already patented previously) aren't necessarily a roadblock to anyone- the USPTO folks are generally....very not technical.... so it's not uncommon for patents that shouldn't be granted due to prior art or being overly general but using words they don't understand getting granted anyway, Most commonly, if this results in lawsuits or anything at all, they end with either the newer patent being tossed if it's REALLY egregious, or if it's genuinely something different but not enough so it'd be easy to convince a jury it's different, in some kind of cross-licensing settlement that isn't usually super material to anything other than large bills from IP lawyers. (exceptions exist of course)



All that said- depending on the numbers, and how long Elon wants to keep it exclusive to Tesla- there's also always the option to just license the tech directly to other car makers.

Which is more in legacys wheelhouses anyway, and might be cheaper to quarterly results than a huge moonshot "copy FSD" program would.


Tesla can't build 80 million new cars a year.

Is RT revenue so great they'd rather keep it exclusive for only a few extra years- versus collect licensing fees on FSD tech in everyone elses cars for a LOT of years
My guess is Tesla will sell FSD license to any manufacturer under the condition that it will be used in EV’s only. That’ll advance the mission.
 


I've just read the patent. Quite difficult (though not impossible) to get around. This is basically using radar/lidar to find distances and velocities of objects, associate them with optical, use optical to disambiguate (using time) then using those distances to objects to train a vision only model which is then used in the cars.

IIRC Tesla had a few cars with LIDAR, data collected by them were probably used by this algorithm.
 
It was great hearing Andrej Karpathy on that podcast using the same words I've been using when arguing with these Twitter clowns opining that Waymo/Cruise are gonna win because they have great simulators. Simulations are useful for varying characteristics of of data you've collected, but they only allow you to interpolate the space your real data collected has sampled. As Karpathy said, you need your real world data to cover as large of a space as possible (e.g. all the edge cases possible), so that your model can interpolate from that data and cover everything.

There is no extrapolation. It is basic brute-force model training.

Now for someone to copy it, well if they do so illegally I could see it not taking too long for them to use it. I have no idea if possible, but if they are actually able to copy the network architectures and the node weights, then they can take the backbone portions and use it with transfer learning and a little bit of their own new data to tune it for their camera setup. This will at least reduce the amount of data they would need to collect. But obviously illegal.

They could also try to learning on the input/output results coming from Tesla's system, but I imagine this would be too tedious and too black/box to be very useful.
Even if competitors copied the network topology and weights, they could not improve it without all the original labeled data along with their new senarios. Unless they froze the early layers and augmented it with new ones specific to the new data. However, without the test data set, they would lack the ability to validate regressive impacts. Umless they locked out the later copied layers also, which basically leaves a new NN in parallel with Tesla's (but that can use the Tesla nodes as-was).

Relatedly:
Regarding VW or others collecting miles. It's not just collecting, it's also labeling (including Karpathy's current job task of curating). This is part of the 4D 1,000 frame at a time support software Tesla is developing. Dojo and this software lead to 'Operation Vacation' where errors/ disengagements are automatically diagnosed and rolled into network updates. (With human oversight/ sanity checks).
 
And to be clear, specifically in regards to copying Tesla's approach to improving autonomous driving, the data collection can happen on non-electric vehicles, so there isn't a strict dependency on batteries.
Some municipalities are already moving toward mandating AV's be EV-only. I doubt any manufacturer would put $100's of AV hardware into vehicles that can never be autonomous.
 
Some days ago I posted a Youtube video from 'T-Study' - this account is methodical and clearly fluent in Chinese so worth keeping an eye on.
Here is a new video reporting on GigaShanghai, basically showing that the land being worked on to the East does _not_ currently belong to Tesla and that the current work there will _not_ support permanent buildings there.
They go on to argue that the land's current use is to support the work crew as they complete the last buildings in the GigaShanghai itself.
Finally they argue that with land being so expensive around Shanghai, Tesla may have succeeded in securing a land lease for that area, which they could then use as loading area for finished cars and similar logistics - something for which they do not need to own the land.

 
When Tesla proves their FSD approach works with fleet vision data collection from their ~2 million vehicles, a determined traditional auto manufacturer like VW could copy the approach and have multiple millions of vehicles within a year with similar sensors and compute to catch up. Add in someone like Mobileye with vision experience and dedication but lacking the high quality fleet data, it seems possible there will be plenty of partnerships when these companies realize this existential threat.

As others have pointed out, compute is constantly improving reducing the turnaround time to train neural networks that are also improving in architecture, techniques and efficiency; so there is some benefit in being a fast follower. So the big question is if competition will react to robotaxis like they did for electric vehicles. Maybe autonomy companies will be more convincing than battery companies in getting legacy auto to change their ways?

Elon Musk has mentioned winner-takes-a-quarter, so he thinks there will be at least 3 other competitors. But that tweet also highlighted "vast amounts of manufacturing" which again could be a potential advantage for existing large auto makers.
VW can't even copy OTA updates in a year let alone copying FSD within that window.
 
I still don't understand how anyone thinks that there will be fast followers in autonomy. Even if Tesla handed you their exact playbook it would still take years to train your own systems, not to mention deploy vehicles that could collect data or use the end product.

On the consumer side. Why on Earth would you not use the Tesla system at 99.999% safety and instead hop into a VW taxi that was only at 99%? It's not like VW could discount their service below Tesla's.
 
I've just read the patent. Quite difficult (though not impossible) to get around.
Is it? Firstly, it is a patent application, not a patent. Secondly, if you check out the Search report that comes with the publication, 3 publications are indicated, all with the label X for all the claims. The label X means that they destroy the Novelty of the claim. Now, a Search report is not the final word in the prosecution procedure and the patent attorney may be able to amend the independent claims to overcome the objections. this typically entails a reduction in scope, which may make it easier to get around the patent.

 
...something for which they do not need to own the land.
Hmm, sounds uninformed. Nobody owns land in China, including Tesla with Gigafactory Shanghai. Goverment owns all land, and leases it as they see fit. Tesla currently is operating on a 50 year lease for Giga Shanghai (they do not own the land that Giga Shanghai is built on).

Also, there is too much ground prep work going on SE of Giga Shanghai Phase II for it to be just for construction staging or logistics support for finished product. They are preparing to build something there.

Cheers!
 
On the consumer side. Why on Earth would you not use the Tesla system at 99.999% safety and instead hop into a VW taxi that was only at 99%? It's not like VW could discount their service below Tesla's.


No legacy OEM would put a 1% failure rate L5 system into service. Their tolerance for risk is FAR lower than Teslas.
 
Why wouldn't Tesla's FSD become like Bluetooth? No one even tries to create something to compete with it now. You just pay Bluetooth for the license to use that technology. And you can't even conceive someone marketing a non Bluetooth phone, speaker, headset or anything else wireless. Apple didn't invent Bluetooth and they pay through the nose to include that in their products.
 
No legacy OEM would put a 1% failure rate L5 system into service. Their tolerance for risk is FAR lower than Teslas.
My numbers are arbitrary of course but assuming that the end result is the same. Either they release it and Tesla is the far safer product (initially), or they wait even longer to release. Of course that assumes that we think Tesla can do it, and that another company can't do it better. If we are on this forum then we probably believe that.

How long did it take Barnes and Nobles to realize that Amazon's model was superior? Or for Nokia/Blackberry to realize that about Apple? Or for legacy cable providers, content creators (Hulu and Disney+), and video rental companies to see that about Netflix? Legacy companies tend to discount the threat at first, then dither about and try to keep a foot in each space, and then try to move aggressively after it's way too late. We still haven't seen the other players in autonomy ditch their models and try to copy Tesla. When that happens it's a minimum of several years before they can hope to catch up.

Why wouldn't Tesla's FSD become like Bluetooth? No one even tries to create something to compete with it now. You just pay Bluetooth for the license to use that technology. And you can't even conceive someone marketing a non Bluetooth phone, speaker, headset or anything else wireless. Apple didn't invent Bluetooth and they pay through the nose to include that in their products.
Licensing is possible but it's more complicated than something like bluetooth. FSD will depend on the specific hardware stack that Tesla uses so those items will need to be incorporated, chips, cameras, etc. I guess Tesla could offer a company like VW training services to train the core NN using VW's hardware suite. Tesla would need to be very production limited to spend that effort though.
 
We still haven't seen the other players in autonomy ditch their models and try to copy Tesla. When that happens it's a minimum of several years before they can hope to catch up.

Waymo and Mobileye actively disparage Tesla's approach. I'm not sure they'll ever swallow their pride and give up all their progress to follow Tesla (if it eventually works).
 
Part 3 of James Douma valuation.

I really don’t get it. Tesla would basically be handing over most of the profits to fleet operators. I cannot see Tesla doing that. Much more likely would be a high upfront cost (e.g. every Tesla sold has a mandatory > $30k FSD package) PLUS Tesla gets 30% of the revenue, or Tesla operates their own fleet and keeps ALL the profit. Tesla has never ceded that kind of profit to any supplier or partner.

in James’ model a new Tesla buyer could purchase a $40k Tesla and earn $77k revenue which is mostly profit, the 1st year alone. Those money making opportunities just DO NOT exist in the real world.

Would love other people’s take. One other small nit is that he showed Global production, but then the rest of the calculations were on U.S. alone.