Tesla (NASDAQ:TSLA) likes to make problems real hard. For instance, while Tesla has lots of cameras, it doesn’t even try to have pairs of same-spec (same lens, same field of view) cameras pointing the same way. They all have different fields of view, so a problem that’s already hard (reliably re-creating the world in 3D from camera vision alone) is made harder still.
A while ago, I covered the emerging CounterGate saga in my articles titled “A New Tesla Scandal Is Brewing” and “Tesla’s Countergate II – The Sequel.” In that saga, Tesla tried to covertly nerf customers’ car performance. Tesla did this, quite obviously, to limit possible warranty claims down the road stemming from the usage of the cars’ full power.
The CounterGate scandal had two acts:
- First Tesla tried to limit the power on Performance models after a certain number of full-power uses. Then, having had that come to light, it relented saying it would no longer limit power.
- However, it still did – just in another fashion. The second act was thus Tesla trying to limit the use of full power to certain restrictive instances (launch mode usage).
Well, this scandal is now over. Tesla got sued by one of its customers. Tesla still tried (includes news of the lawsuit, attempted NDA resolution, etc) to use its old trick of offering to fix only that customer’s car and then putting him under a NDA (Non-Disclosure Agreement). The customer did not accept the solution, and finally Tesla quit and simply accepted to restore the performance to all its customers’ cars (as demanded by the heroic owner). So that was the end of the CounterGate scandal.
With this scandal being over, today I am going to talk about yet another emerging scandal. That will be the APGate scandal.
APGate, The Origins
The APGate scandal begins in October 2016, with Tesla’s introduction of HW2.0 (Hardware 2.0) to implement AutoPilot 2.0. This hardware, as per Tesla, was ready to provide full self-driving.
A bit of context. Some might already be qualifying what this “full self-driving” ultimately means. Elon is quoted as it being Level 5, so fully autonomous under all conditions. But what Elon says and what Tesla sells are sometimes different things, so what does Tesla’s order page actually say? Here’s a screen grab (highlight is mine):
Notice the following: There are two instances (when the car seeks parking, and when it is used in a supposed Tesla Network) where the car self-drives while empty. This removes all doubt. When empty, the car needs to be able to face any possible circumstance. What this means, is that the level being implicitly promised is SAE Level 5 self-driving. This claim was present at all times starting in October 2016.
Now, Tesla doesn’t have self-driving technology. So Tesla is selling something it neither has, nor can be certain of having in the future. Tesla never promised any timeframe for having such a technology, but did mislead customers both with videos of Teslas supposedly self-driving, and promising an autonomous trip from California to New York during 2017– which I already explained wouldn’t prove anything.
Furthermore, testing data shows Tesla lags severely behind other companies seeking to deliver self-driving technology. This lag isn’t surprising. Tesla started late, and is trying to solve a more difficult problem, because while self-driving leaders use a combination of LIDAR+cameras+radar, Tesla chose to leave LIDAR out of its hardware suite (given the present cost).
In removing LIDAR, Tesla removed the most reliable way to currently detect the environment around the car. A car equipped with LIDAR can get a high degree of certainty regarding all relevant objects around it (even before recognizing those objects). A car equipped with cameras, can’t. It sees the world, but that’s different from actually detecting and measuring the distance to all objects around it.
Tesla, obviously, is trying to recreate the information given directly by LIDAR using cameras and computation. This task has been researched, and is possible to be performed (and indeed, used in some instances). However, the reliability in performing this task isn’t necessarily high enough for self-driving duties. The problems emerge both because some objects won’t be reliably detected, and because objects will be detected which don’t exist or aren’t in the car’s path (false positives).
In self-driving, false positives are a large problem. A self-driving car needs to turn conservative when faced with a possible solid object it might hit. If this object is “imaginary” (a false positive), the car will brake – sometimes brake hard – for something that isn’t there. In many instances, this will create a safety hazard for other cars in the road. Unexpected braking will be met by slower reaction times from human drivers, and crashes resulting from it are to be expected.
Now, on the issue of false positives, we already have at least two reasons to believe they’re affecting Tesla:
- The fact that AP2 cars often brake for underpasses and trees (among other flaws) when using TACC (Traffic Aware Cruise Control) or AutoPilot (TACC+lane keeping).
- And the fact that Teslas, when subjected to pedal misapplication, can happily run from a standstill into a building (see my prior Tesla article for examples).
In the first case, the car is seeing a “false positive” object in its path, and not ignoring it. Thus, it unexpectedly brakes. In the second case, the car is certainly getting a large radar signature from the building. However, in trying to ignore all those false positives which happen elsewhere, the car ignores the giant building straight ahead and plows into it.
Now, both of these cases also highlight something else: They highlight that at present, Tesla hasn’t yet solved the problem of using cameras to reliably map the surroundings. Had Tesla done that, and neither of those cases would have happened.
That Tesla hasn’t yet solved how to fully detect its environment using cameras, tells us just how far from FSD Tesla still is. Detecting the environment isn’t the end point of a self-driving effort. Instead, it’s the starting point. When using technologies like LIDAR or Flash LIDAR, the very first thing a self-driving developer gets is a reliable 3D picture of its (solid) environment. Using this picture alone, a startup self-driving effort would neither brake for overhead signage, nor accelerate into giant buildings from a standstill.
It’s not a coincidence that Google cars haven’t yet hit any other cars. Even when Google’s car got in an accident where it was at fault, the Google car was hit (scraped) by a bus, not the other way around. This happens because the Google car has a 3D picture of its solid environment, and thus can avoid hitting (but not being hit) literally anything almost by definition (there are exceptions with transparencies, small objects, etc).
Now, with this description up to here, we already have a good clue about where Tesla stands on FSD, and how far it is from actually delivering on its promises. It’s already a scandal that Tesla is selling something it cannot deliver on any specific schedule. Indeed, Tesla is selling something it might not be able to deliver using its present hardware suite at all. It’s also rather interesting that even Elon Musk himself puts the timeframe to Level 5 about 2 years away (counting from late April 2017…), and Elon is usually late. It’s interesting because Tesla cars often get sold on 2-year and 3-year leases.
But the scandal I’m bringing you today is rather different. Instead, the scandal has to do with the fact that all the FSD promises above were made on a hardware suite (HW2.0), which Tesla has already obsoleted by launching HW2.5. “Oh but that was already known,” you’ll say. Indeed it was, Electrek broke the story a month ago.
At the time, to soothe fears, Tesla said it would upgrade HW2.0 customers to the improved computer on HW2.5, if need be (Source: Electrek article linked above):
“However, we still expect to achieve full self-driving capability with safety more than twice as good as the average human driver without making any hardware changes to HW 2.0. If this does not turn out to be the case, which we think is highly unlikely, we will upgrade customers to the 2.5 computer at no cost.”
Then, more recently, HW2.5 customers got notice that their AEB (Automatic Emergency Braking) feature would stop working for a few weeks, as it re-calibrates the new hardware.
It was the AEB going out which got me digging a little. You see, computers, even more powerful computers built on the same architecture, don’t need any re-calibration. So something else had to be amiss. And indeed it was.
At the time of the HW2.5 introduction, Tesla downplayed the changes HW2.5 brought (Source: Electrek article linked above):
“The internal name HW 2.5 is an overstatement, and instead it should be called something more like HW 2.1. This hardware set has some added computing and wiring redundancy, which very slightly improves reliability, but it does not have an additional Pascal GPU.”
With this description, it seemed that all that was changed was a bit of cabling for redundancy and reliability, and then the computer unit which could be upgraded for free down the line (as explained by Tesla in the first quote). But this is misleading, because (as discovered by enterprising Tesla owners):
- Redundancy itself might be a regulatory requirement for FSD in the future.
- And most importantly, wiring and the computer were not the only changes. Along with those changes, Tesla also added a new radar. Previously Tesla used a Bosch MRREvo14f, and now Tesla has changed this radar for a more capable Continental unit. This new radar, is the reason why HW2.5 cars now need re-calibration. And once the re-calibration ends, a surprise will emerge: the new radar is more capable, so these cars will perform better than the HW2.0 cars already.
As a result of these changes, the car is already getting more capable computers and sensors. Also, still more hardware changes still are oncoming. For instance, the Model 3, using mostly the same HW2.5 suite, already includes an interior camera. It won’t stop there.
The official Tesla line is that the original hardware will still be capable of the promises. However, faster and more capable hardware would always make for an easier FSD problem. As a result, at the very least a FSD solution would arrive on the more capable hardware first.
Now, given what we saw regarding the state of Tesla’s FSD development (both from testing data and from the fact it doesn’t have a reliable 3D view of the world), for now neither hardware solution (HW2.0 or HW2.5) is likely to attain FSD. But Tesla continuing to add to its hardware solution will, at some point, fully obsolete HW2.0 (and even HW2.5). In the meantime, tens of thousands of cars will be carrying that obsolete hardware together with a promise that it will someday be enough for FSD Level 5.
This development is an even larger liability than CounterGate ever was. This touches every car Tesla sold since October 2016. This touches even the cars which did not pay for the FSD option, because the capability to have FSD, even if bought later, could arguably have been at the core of the buyer’s decision. In the end, what’s happening to Tesla with the HW2.0 APGate, is no different from what happened with Volkswagen (OTCPK:VLKAY) and the emissions scandal. VW (and Tesla) sold cars based on capabilities and specs they did not have. And VW ultimately had to pay compensation and even buy back cars from owners.
A Side Note
As expected, Tesla has been on a tremendous discounting drive to improve Q3 2017 deliveries. This discounting drive included up to $30,000-$40,000 discounts on Model S P90DLs, P100Ds, as well as other discounts across the range.
There are also customers which managed to buy new 90Ds and P90Ds with leases as low as $600 per month. To put this in context, a new base Model S75 RWD, costing $69,500 in cash, on the cheapest terms possible (36 months, 10,000 miles/year) goes for $790/month.
This discounting wave also introduced something that’s remarkable. You see, Tesla discounts “discontinued” models more. Thing is, many of discounts were on “discontinued” 90Ds which were coming new off the line, being built during the quarter!
Tesla guided for a 20% drop in automotive gross margins (from 25% to 20%) for Q3 2017. Tesla put it down to starting up Model 3 production. However, it’s evident that a lot of the gross margin drop is coming both from including more equipment in base models, and extreme discounting (as predicted).
The CounterGate, so many times denied by Tesla fans, was not only true but has been solved by Tesla capitulating on it. As a result of that scandal, Tesla ended up spending some of its best customers’ goodwill. It will also be eating higher warranty claims in the future, because of the now unrestricted use of those cars’ performance.
The new APGate stands to be an even larger scandal. Already Autopilot hardware is evolving beyond the initial specification, promised to be enough to attain self-driving SAE Level 5. Ultimately, and with this continuing hardware improvement process, HW2.0 (and even HW2.5) will be fully obsoleted.
At that point, Tesla will have created a gigantic liability for itself, since it sold tens of thousands of cars which it said would be able to perform to SAE Level 5 self-driving duties, and this will never be attained on the original hardware. Neither will the hardware be retrofit-able, as what’s being changed goes well beyond just the CPU box.
On top of that, this development will see further customer goodwill being wasted even before the lack of FSD capability realization takes place. For instance, right now AP2 customers have cars which don’t perform to AP1 levels, even though such parity was promised for as early as December 2016. Moreover, with the new hardware improvements it’s highly likely that Autopilot features on HW2.5 cars will quickly surpass those on HW2.0 while HW2.0 cars don’t yet get their promises fulfilled. This will be more evident on TACC, because of the new radar. It’s my estimate that the new radar will quickly reduce the false positives which lead to severe braking out of the blue on HW2.0 cars. At that point current customers will still be stuck with the worse-performing TACC on their “FSD cars” while HW2.5 customers will no longer have to endure it.
Tesla’s most important asset is the brand it built. Yet, with these scandals and unfulfilled promises it’s burning a bit of that brand every day. This is already visible in Tesla Motor Club’s forums, where many former Tesla fans are growing every more skeptical. Tesla is slowly destroying its most valuable asset.
Disclosure: I am/we are short TSLA.
I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.