Sign up for daily news updates from CleanTechnica on email. Or follow us on Google News!
In the world of Tesla Full Self Driving (FSD), I hear two main views. The Tesla bulls (of which I am a member) seem to pretend that it is a foregone conclusion that Tesla will solve FSD soon and the only thing left is details like the app to call a robotaxi or designing a car that is self-cleaning, leaving the elephant in the room (V12 was a huge advance, but little progress has been observed in the last 6 months).
Just finished a one-week trip to China. I’ve now “survived” all the major (~20) L2 self-driving and robotaxi vehicles in both the US and China. Some thoughts & observations:
▶️L2 self-driving
I tested major brands like $Huawei, $Li, $NIO, $Xpeng, and $Xiaomi. Overall, they… pic.twitter.com/2iTluT24SU
— Freda Duan (@FredaDuan) July 2, 2024
This recent post on X by Freda Duan of Altimeter Capital shows that Tesla has a lot of competition in China that is already deploying driverless vehicles. She sees Tesla’s “imitation learning/end-to-end as the only effective approach to self driving,” but I think that conclusion is premature, especially considering Waymo and Pony.ai have already solved self driving. Tesla bulls will frequently dismiss those that use LiDAR and mapping as not scalable for two main reasons:
- They state that LiDAR is expensive. That is true, but electric vehicles were expensive and the costs came down considerably. Many companies are working on less expensive sensors and they might lower the costs dramatically, especially as they scale up. Instead of researching the cost curves, they just wave their hands and dismiss it.
- They also state that only Tesla has the data to train FSD, so no one else can do it or do it before Tesla. The world of computing is full of companies that had an insurmountable lead and blew it. I like to bring up the example of my kids when they were 15 years old — they learned to drive without absorbing millions of miles a day of video from 8 cameras. They seemed to figure it out in a few weeks of training. Of course, I realize they had 15 years of training on “the real world” first, which gave them a great basis to learn to drive. I bring that up as a reminder that it is very hard to predict what the right algorithm to solve any problem is until you succeed.
An economic point they are missing is that if 90% of the value of FSD is in the cities and in 1% of the land mass (cities plus interstate highways), you may be able to map that 1% of the land mass and capture most of the money. If I can take a robotaxi to all the places I go regularly and it also takes me on my vacations, we shouldn’t dismiss it just because it can’t go the places I don’t want to go.
Bulls on FSD will say that the take rate on FSD will rise as the technology gets better, but I find that FSD has little value in the city as long as it must be supervised. I use FSD everyday and enjoy monitoring its progress and it is a joy to use on the highways and in stop-and-go traffic. But where it is making a lot of turns in the city, it is more work to carefully watch FSD to ensure it doesn’t make a mistake than to just drive myself. I continue to use FSD in the city just because I’m into technology, but other members of my family think it is more hassle than its worth.
The FSD take rate has been horrible over the last 8 years (except when a super sale is offered), showing the buying public agrees with me. Tesla just did a free trial of FSD and you can bet your bottom dollar that had it significantly increased the FSD take rate, Elon would have posted about that. The fact is that Autopilot is outstanding and free with every Tesla, so the only reason to buy FSD is if you like to play with the newest tech as toy (like I do). Of course, there is Elon’s promise that the car will be worth a fortune once FSD is solved, but that is dependent on a lot of assumptions. It assumes FSD will be solved before I sell the car. It assumes it will be solved for the FSD computer in my car or Tesla will provide an upgrade. It assumes the revenue split between Tesla and the owner will be decent. It assumes it will be worth the time and hassle to charge and clean my car between robotaxi rides.
FSD Training Costs Explode
My recent worry is that the costs of training FSD are rapidly expanding while we still aren’t even close to the level of safety necessary to remove the driver from the car. Some people will say that Tesla’s FSD is already safer than a human driver based on the figures Tesla releases. That is false. The figures that Tesla releases do prove that FSD being supervised by a Tesla owner is VERY safe and far safer than an average human driver, but using FSD every day shows me that I would have an accident every few days if I didn’t intervene for FSD when it needs some help.
This study by the American Automotive Association (AAA) agrees with the Tesla data that accidents occur only every 500,000 miles are so. This doesn’t agree at all with my experience where I have driven about 500,000 miles over 40 years and been in about 10 accidents (only one my fault). But I have a feeling that fender benders at 2 mph may not be included in those figures because no police report is filed and most of my accidents were collisions at 2 or 3 mph. Having said that, if FSD can go about 10 miles in Florida without an intervention that might cause an accident, we have to improve the quality of the driving about 50,000 times before it is safer than a human driving. Even if each release improves safety 5 times, that would be 7 releases to get that level of safety. If each release takes 3 months, that is almost 2 years, which is very reasonable, but if there is anything I’ve learned about FSD over the last 8 years, it is that progress can be going great and it can stall for a year at any time. Even if Elon posts that he is VERY confident that FSD will be solved this year, because he has driven the latest release and it is amazing, progress could be stalled and nothing could progress until a whole new design is developed by a large team of extremely expensive engineers using extremely expensive computer resources.
Conclusion
Of the roughly $10B in AI-related expenditures I said Tesla would make this year, about half is internal, primarily the Tesla-designed AI inference computer and sensors present in all of our cars, plus Dojo.
For building the AI training superclusters, NVidia hardware is about…
— Elon Musk (@elonmusk) June 4, 2024
The bottom line is that with about $30 billion in cash and significant positive cash flow adding to that stash, Tesla can afford several years of spending $10 billion a year on AI-related expenditures, but with expected earnings of about $8 billion a year in 2024 (none of it from AI), how long will the shareholders be willing to invest more than their annual earnings on software that has costs that keep rising but that doesn’t provide any financial return? The answer is as long as progress is impressive or until another company shows FSD has been solved and left Tesla in the dust.
Will Tesla solve FSD before the competition? The answer is no — they have already lost that race to Waymo and Pony.ai — but will they solve FSD before others find a way to scale level 4 self driving? I still think they will, but it is truly a race and Tesla engineers should not be overconfident that they cannot be beat just because they have some unique advantages.
Disclosure: I am a shareholder in Tesla [TSLA], BYD [BYDDY], Nio [NIO], XPeng [XPEV], NextEra Energy [NEP], and several ARK ETFs. But I offer no investment advice of any sort here.
Have a tip for CleanTechnica? Want to advertise? Want to suggest a guest for our CleanTech Talk podcast? Contact us here.
Latest CleanTechnica.TV Videos
CleanTechnica uses affiliate links. See our policy here.
CleanTechnica’s Comment Policy