I’ve been musing about shorting TSLA the last few days. Not bothering with numerical nuts and bolts, I wonder what happens if/when Elon announces that RoboTaxi isn’t yet ready for prime time, or that it simply doesn’t work as configured and “they’re going back to the drawing board.”
I note there are (at least) two companies already deploying robotaxi service, and while it’s not profitable I think that’s irrelevant. Musk has gone “all in” on using video cameras and AI to train, but so far (from every first-person account I have read) it doesn’t work. Suppose, just suppose, it will never work that way and that the more expensive Lidar system is the only way to get to the goal line, at least for the foreseeable future.
That would mean that not a single Tesla produced and on the road would be capable of being a RoboTaxi, exactly opposite of what has been promised. Further, that means no software sales to current owners, and the only way to put a robotaxi in service is to buy new metal and deploy. That seems a hard push, buying expensive capital to recoup $4 at a time - precisely why the companies already in the business aren’t making money at it.
Now I come back to the August announcement, which turns out to be the October announcement, and who will be surprised if it turns out to be the February announcement? On top of which Tesla’s numbers are not nearly so stellar as they were a year ago, so the stock is sitting on thin ice waiting for the pop which may never come. Musks continuing antics and distractions are another factor, but not nearly so important as this one moment in time: the acceptance or inability of a long promised feature to materialize.
The retrenchment of other automakers doesn’t bother me particularly, I’m not invested in any of them and it’s impossible to know at this point who survives and who doesn’t. Don’t get me wrong, Tesla survives, it has become a great and powerful company, but the stock is still lofty by any traditional measure and treading water is never a good look for a tech star.
Anybody have any insight in what might make TSLA go the other way?
But note that what is deployed is not only geo-fenced, but has remote drivers available to take over when an intervention is needed. Not only is that a huge expense, especially at their current rate of intervention, but it puts all kinds of limits on deployment.
Here are what I think would be the leading theories:
Late but right. Tesla falls behind Waymo and other LIDAR-based systems, because it ends up being vastly harder to make a “brain” that can drive with only visual inputs and without pre-mapping. However, it is solvable - and Tesla solves it. Tesla ends up with a cheaper solution, and one that can be scaled more quickly (presume here that it does not need the back-up remote drivers that other Level 4 systems do).
Moonshot lands. Tesla never solves FSD and it becomes a problem for them in automotive, but they hit the jackpot in AI or robots. So FSD becomes another (though larger) “Solar” for them - something they sunk money into and never panned out, but they expand into their market cap by getting ahead in AI or robotics.
Fundamentals get much better. Real autonomy doesn’t happen any time soon, so neither Tesla nor the other players get a good, inexpensive scalable AV system up and running. So Tesla pivots back to the current established business segments: human-driven cars and energy storage. Relieved from the distraction of FSD, Tesla re-focuses on their “dreadnought” factories and growing their energy systems, get back to rapid rates of growth (aided by a collapse of Chinese firms as overcapacity and protectionism take their toll).
I won’t try to defend the likelihood of any of these, but I think those are the main scenarios where Tesla scourges the doubters and skeptics.
Here is my first person account.
In the current configuration (version 12.4.x) on a Model 3 with HW3, it is not ready to be a robotaxi. I could make a big list of deficiencies, but the most important is that the first and last 50 feet of a trip are non-functional. It doesn’t allow enabling FSD until you get out of a driveway or parking space. And at the end of a trip it doesn’t park or really move to a safe place for passengers to exit. They could have this running on some internal cars, but that is not close to good enough.
Beyond that, a year ago I would have said just the driving from point A (+50 ft) to B (-50 ft) still needs years of work. But version 12 really did improve a lot.I very rarely made a trip over a year ago without an intervention of some kind – mostly minor non-safety related issues, but certainly deal breakers. But in the last 6 or 8 (?) months with version 12 installed I estimate at least 4 out of 5 trips have no interventions except taking over the last 50 ft because I know where I want to park and the car doesn’t and there currently is no way to tell it (unless it happens to go into a parking lot and there are cars parked and a spot shows up that you can tap and make it park there).
The biggest failures I see, while in the middle of a drive, it that the navigation says to take a freeway interchange and the car just doesn’t do it. This may be some mapping error on lanes, I don’t know, but it is very repeatable on a couple of specific roads.
Other than that, I find that the car is too timid in taking a turn from stopped. For example a right turn on red. Sometimes I give up and tap the pedal to force it to go while it was waiting for a extra large distance between oncoming cars, missing 2 or 3 chances. (If someone is waiting behind me I’ll always give it a nudge sooner)
For anyone interested in how far things have gone with the most developed release (12.5.1.3) of FSD, and I am sure the most up to date hardware, there is a video worth watching that Smorgasbord1 posted a link to on the paid side. The driver picks up his father at the airport (LAX) and has FSD take them home. The first part, getting through and away from the airport, is particularly interesting, but the whole 40+ minute trip is informative. No hand touched the steering, no foot touched a peddle.
If they have no driver in the car, they definitely need some sort of failsafe like back up remote drivers. If the car can’t move on its own, what else could they do???
LIDAR isn’t much different than vision. Input-wise that is. A few years ago, I worked on a system that used LIDAR to detect pedestrians entering and walking across a street crossing and it essentially used the same kind of software that video input would use. And this particular system was spectacularly bad at determining individuals, it would routinely spit out 5, 10, or more instances of a single individual instead of properly tracing the single individual as they made their way across the street.
In this scenario, Tesla’s AI - though late - is good enough that the need for a failsafe is eliminated completely (like driver interventions), or nearly so. Tesla’s gamble is taking on a tougher challenge on the “brain” side of the AV, rather than using the “crutches” of LIDAR, mapping, and a robust system of remote operators. By making the software smarter, all the other aspects of the system are cheaper and can scale more quickly.
That won’t work for Level 5 autonomy. Or even Level 4. The car has to be able to handle safety issues entirely by itself - or nearly so. It has to all-but-eliminate the need for any real-time human interventions. The car has to be able to get to a safe condition without any help, if that’s necessary. Which is where Waymo is, for the limited areas and circumstances in which it operates.
The remote operators are used when the car doesn’t know how to handle a non-safety dilemma. It’s already in a safe condition/location, but it’s unable to proceed from there to its destination. For example, if it’s encountered something unfamiliar that it can’t process and can’t adjust to, it might stop and call for human assistance - the passenger is safe, but the car can’t figure out on its own what to do next. The remote operator doesn’t drive the car - it just tells the AV what to do.
Honestly, it seems like it would be more dangerous than helpful, on the whole. There are few scenarios where pushing it might help avoid an accident (if your AI can’t always avoid a head on collision, it shouldn’t be operating without a driver in the first place). And lots of scenarios where pushing it causes an accident.
The selfdriving vehicle incidents that make the news are usually collisions. Usually with other vehicles; sometimes with pedestrians or bicycles, etc. A stop button would help.
“Late but right”. Entirely possible, but I would be making a short, short bet, that is, only expecting to hold it some months not years. If he’s right, it won’t be soon. If he’s wrong, it won’t matter because I’ll be out anyway.
“Moonshot”. Again, it seems unlikely that any of the other businesses will come true within, say, a few months. They may be great and wonderful someday, but I don’t think it’s going to hold the stock up within the short time frame.
‘Fundamental get much better”. The only way for that to happen, again within my time frame, would be for the car business to rebound smartly. Given the increased competition, his antics alienating his most likely customers, and extending current trends into the future (always a fraught exercise in investing) a rebound of fundamentals also seems unlikely.
Gotcha. Well, if it’s a short term bet, I think you’re just gambling, if you’re making a guess on how the stock will respond after the October robotaxi event.
We don’t know what the substance of the event will be, but unveiling the actual launch of the robotaxi service was never in the cards. Everyone knows that FSD isn’t ready for that. Instead, we’re looking at possibly a vehicle reveal, discussion/demonstration of the customer-facing ridehail app, possibly some information about future milestones or testing of ridehail in some states, and the like. Nothing that is likely to be delayed.
You’re really just betting on how shareholders will feel about what’s revealed. Will they be excited about the progress towards robotaxi, or deflated by the details? It seems like you’re placing a wager with very low information.
And then what exactly? Do they send a tow truck to pick up the car?
LIDAR is not a crutch per se, it’s just another way of “seeing”. In fact, one of the reasons Tesla presented when they eliminated it was “What if LIDAR sees something one way and vision sees something another way, how do you react?” and the answer was that vision always trumps LIDAR.
Mapping is not a crutch, it is used all the time for driving. Detailed visual street maps are a crutch, and are generally ineffective because they change all the time. Even something as simple as a truck/car winging a signpost changes it somewhat. Construction changes it. Temporary cones change it. Etc.
This is a true crutch. Might be needed for the first 2 or 3 generations of AVs. Might not. I can’t decide yet, but likely yes for now.
The level 1 remote supervision will almost surely be AI. It will hand off to level 2 human supervision in the rare cases that it is necessary. There’s no way Tesla will have hundreds of thousands of AVs around the world with all-human supervision, no way!
In the short-term, Tesla will go down. I guarantee it. It will also go up. And it may also stagnate. Guaranteed. How can I guarantee that? Because it is true for all stocks in the short-term. Using short term option spreads are best for this kind of bet because they allow you to choose ANY level of risk and potential return you choose. Literally “any”.
No, they would have some remote operators. But that particular scenario (“Late but better”) involves Tesla getting to an AV package that scales better than competitors. If your AI relies heavily on human intervention in too many cases (hi, Just Walk Out), it doesn’t scale well.
Right - but that’s the point (and why Musk uses that term, I think). If you ratchet up the inputs to the AI driver with things other than vision (like LIDAR and mapping data), the brain doesn’t need to be as powerful. Because it’s supplemented by the extra inputs. But the extra inputs are brittle and don’t scale as well - LIDAR requires more expensive sensors, and 3D mapping starts getting broken over time if it’s not maintained.
So the Tesla theory is that it’s better to put more money and time into getting a better software package into the car, rather than use these other assistive measures. Because software scales.
That assumes, of course, that it is possible to create a Level 5 vision-only software driver using the technology available to us today (or over a relevantly short horizon). That doesn’t work if we’re a few decades away from having the skills and tech to do this.
SEPTA regional rail runs on overhead electricity. Their solution to a power failure is a power car with a generator on it. Apparently it can move a stalled train.
