Latest comedic fantasy from Tesla Q2 2025 earnings call

Stocks A to Z
,
1

Tesla’s earnings call was full of entertaining content, part fantasy, part comedy, with several hyperbolic statements from people who should really know better, especially after literally years and years of making statements very divorced from reality.

The two most revealing statements (7k miles driven in Austin, unsupervised driving is not ready) came from the VP of AI, not the CEO, unsurprisingly.

The best piece of actual data from the call (aside from financials in SEC filings) was that in Austin their 10-20 human-supervised, geofenced, limited hours, limited weather, taxis with invite-only ridership have achieved about 7,000 miles in one month.

At that rate it will take a very long time for Tesla to estimate an accident rate near 1 per million miles (or better) which is what Waymo has shared in a publicly available analysis of about 55 million unsupervised miles through 2024 (and because Tesla is doing supervised driving, they will have to estimate which disengagements were otherwise likely to have resulted in an accident).

But Tesla says they want to scale “hyper-exponentially,” so maybe their mileage rate will really pick up. Maybe they will relax the many limitations on their service which also limits the generality of the service and the generality of the data they collect with the service.

This begs a question.

What about all of the miles from their customer retail fleet that were touted as being such a data advantage?

Can’t these miles be used to prove safety? After all, aren’t all of these retail miles the same supervised driving as their small fleet of Austin taxis? (maybe not “same enough”?)

How did/will this data advantage actually manifest itself?

No one can answer these kinds of questions about the advantage of the retail fleet data.

Here are some quotes from the earnings call focused on autonomous driving followed by my comments.

Quote from earnings call

No, you won’t have autonomous (unsupervised) ride-hailing available to half the US population by the end of 2025, not technically because you are not technically able to do it at a sufficiently safe level.

Quote from earnings call

No mention of expanding the number of vehicles or expanding availability to the broader public. Expanding the region has limited value without expanding vehicles and ridership - unless your confidence in your system is low.

Quote from earnings call

Nope. The person in the driver seat is not just to expedite something about regulations per se but because Tesla has not yet demonstrated capability of unsupervised driving with a reasonable safety rate. And some of those regulations might also require demonstrating a certain level of capability, so technical achievement may not be separate from regulatory compliance.

Quote from earnings call

Nope. Won’t happen by end of 2025. This is probably the most outrageous, clearly not achievable “prediction.” Although maybe they will attempt something with a ton of limitations, like the geography of a parking lot.

Quote from earnings call

Clear acknowledgement that their unsupervised driving product is not ready, from the AI VP.

Quote from earnings call

Seven thousand miles in Austin is the most useful data point, again from the AI VP. They do mention expanding the number of vehicles, which will be required at some point.

7 Likes
2

Nice summary and rebuttals, ML.
I have no dog in the fight ( I do own some TSLA via index funds ), but I am interested in EV’s, so it’s nice to read the con to the very, very pro commentary that comes from TSLA fans.
I think one of the reasons I dislike people like Musk and Trump is that my parents brought us all up to not lie and make outlandish claims. But making outlandish, pollyanish claims is a daily habit of those guys.
I am not anti EV, I do think I will own one before I die. FSD would be fantastic, if safe and reliable. So I hope Tesla succeeds. Same for Waymo.
Appreciate your commentary on the current disconnect between what is, and what is claimed to be “just ahead”.

5 Likes
3

Yawnnnnn…

You must hate most politicians.

The Captain

1 Like
4

Apparently Tesla is expanding their taxi testing to the SF area, also mentioned on the earnings call.

It looks very similar to their Austin taxi service, except they have to dance a little more carefully with the CA regulators.

From what I’ve read, in SF, looks like Tesla might have (also posted in other thread):

  • done lidar and/or other sensor calibration in advance
  • geofence
  • human-supervised FSD from driver’s seat
  • passengers who are employee and employee family and friends
  • pass this off to regulators as human taxi/charter service so it doesn’t fall under AV testing permit and won’t have to report associated data, but they will be effectively AV testing except without bothersome transparency of disclosure

Tesla might achieve:

  • marketing the effort as “robotaxi” “service” “expansion” (as compared to reality)
  • opportunity to get test data, including lidar calibration (but, as I ask repeatedly, if they need to do further geography-specific supervised testing and lidar calibration, where is the value-add of their enormous retail fleet supervised data set? where is the scale advantage in their retail data?)
  • avoid general public rider scrutiny, especially SF customer base which might be much less friendly than Austin, and instead have a very captive ridership of employee-connected people
  • avoid data disclosure to public and regulators that would reveal system performance

I wonder why the even tighter restriction on riders. It seems they want to avoid transparency even beyond Austin’s rider limits.

I wonder if this has to do with CA’s stricter regulatory regime?

We will have to see if the number of vehicles and miles gets disclosed eventually. With such thin ridership, seems like miles will be relatively low relative to a service open to the general public.

There were - and still are - all of the big predictions and proclamations, and then there is the above much, much smaller reality.

1 Like
5

Meanwhile, I completed 750 mile trip across multiple cities in various conditions on FSD, HW3. 3 interventions. 2 years ago, this would have not been possible.

FSD is getting better every day.

2 Likes
6

Out of curiousity - what were the circumstances surrounding those three interventions? What was the car about to do that you needed to intervene?

1 Like
7

  1. I exited a parking lot and drove in a one way stree in wrong direction (by mistake). Then turned on FSD by habit. It did not know how to recover. It just stood there.

  2. On the border between US and Canada, it needed to take a U turn. FSD tried to cross the border (Niagara). It does not understand the concept of borders. I had to take over and do an “illegal” U turn.

  3. On flashing light, it slowed down and stopped. I had to push on the accelerator to nudge it.

FSD navigates flawlessly through construction zones, rotaries, right / left turns, lane merges, pedestrians, bikers, small lanes, highways etc. It is especially very useful in complex areas in cities where I am not familiar with the roads/turns. It figures out everything.

I have HW3. New cars are on HW4. I suspect Robotaxis are on HW5.
In 2 years, models trained with HW6 will be much much better.

5 Likes
8

The hardware in the car does not do the training, it does the inference based on the training in the datacenter.

AI confirms

That’s correct. In the context of self-driving cars or other AI-powered systems, the training of the AI model typically happens in a powerful data center, not within the car’s hardware itself. The car’s onboard hardware is then used for inference, which is the process of using the pre-trained model to make predictions or decisions based on real-time data it receives.

Here’s a more detailed explanation:

  • Training:

This is where the AI model learns from vast amounts of data. For example, in a self-driving car, the model might be trained on millions of images and videos of roads, traffic, pedestrians, etc., to learn how to recognize these objects and make appropriate driving decisions.This training process requires significant computational power and is usually done in large data centers equipped with powerful GPUs.

  • Inference:

Once the model is trained, it’s deployed to the car’s hardware. The car’s sensors (cameras, lidar, etc.) capture real-time data, and the pre-trained model uses this data to make predictions and decisions, such as when to brake, steer, or accelerate. This inference process needs to be fast and efficient to ensure the car can react in real-time to its environment.

  • Edge Computing:

While training is centralized in data centers, there’s a trend towards “inference at the edge,” meaning more inference tasks are being moved to devices like cars and smartphones. This allows for lower latency and better performance, especially in applications where real-time responsiveness is crucial.

  • Example:

Imagine a self-driving car approaching an intersection.The training phase would have taught the model how to recognize traffic signals and other vehicles. The inference phase on the car’s hardware would then use that knowledge to interpret the current traffic situation and decide whether to proceed or stop.

The Captain

3 Likes
9

Thanks for the information. One follow-up: why did you wait until after having left the parking lot to activate FSD? Do you not use it in parking lots?

10

No. The AI model for each hardware type and each car model is different. These AI models are trained with specific camera resolutions and car dimensions etc. Much of the data and annotation are same but may have subtle SFT elements for each student model from the teacher model.

The inference chips are the same but that does change when a new model gets uploaded. My inference chip is same as 7 years ago.

4 Likes
11

I don’t use it parking lots because FSD is overcautios and slow in parking lots.

2 Likes
12

Yet, later on the same post you quote them more than mentioning that:

1 Like
13

Which I noted in my comment immediately after the earnings quote.

14

So Tesla has now launched a service in San Francisco…but apparently Musk is referring to it as a “ride-hailing” service, rather than calling it “robotaxi”:

His actual post is a little odd, because I’m pretty sure you’ve been able to ride-hail a Tesla for well more than a decade in the Bay Area. The only difference is now you’d be using a Tesla app to do it, rather than Uber or Lyft or any other existing ride-hail service.

4 Likes
15

Don’t think day by day. The new revenue stream has just started and expanding already.

In a few years, this will be massive.

2 Likes
16

I read somewhere that they aren’t permitted to call it a robotaxi yet. First they need 50,000 hrs of supervised driving, then they can apply for a robotaxi permit, and when the permit is granted they can call it a robotaxi. CA almost always has more regulations and requirements than other places. And that extra regulation tends to delay things somewhat.

1 Like
17

Not thinking day by day - just looking at what today’s news tells us about where the puck will be “in a few years.”

Right now, it doesn’t look like it will be massive, any more than a massive revenue stream was “a few years” away back in…I don’t know, 2021, when Musk was also predicting that actual self-driving was coming soon.

Robotaxis with a driver (or safety monitor or what have you) have the same 1:1 ratio of people to cars as conventional ride-hailing or taxis. You need at least one person for every car in service. That’s the same labor cost structure as ordinary driving for hire.

You see savings (possibly) if you get the people out of the cars, and can reduce that ratio. At least, if you can get them out of the cars faster than you have to ramp up your out-of-car staff.

But it’s been about five years since Waymo first started offering driverless rides (in Phoenix). That’s five years that Tesla’s had to work on having the software work without someone in the car doing part of the driving (in addition to the years of work on FSD before then). Years and years of testing after Waymo managed to do that on the reg. And they’re still not ready to pull the human from the car?

3 Likes
18

Waymo is a dead end. They don’t have their own cars and the tech is too expensive. They cannot expand fast. It is a loss making machine research experiment funded by Google big daddy’s money.

1 Like
19

They don’t need their own cars, and the cost of the tech can fall.

Meanwhile, Tesla’s tech hasn’t been shown to work yet - the software stack still hasn’t been demonstrated to function with a low enough failure rate that it can run without a human in the car performing critical driving functions. You could just as easily label Tesla’s self-driving efforts as a loss making machine research experiment funded by Tesla big daddy’s money, if you were so inclined.

It may be that neither system ever “works” as a viable enterprise. We might just be too far away from the point where computers and AI are powerful enough to self-drive. You couldn’t have built a self-driving car in 2010 with all the money in the world - we were still more than a decade away from the basic tech and software design knowledge to even try. We were still more than a decade away even in 2016, when Musk (mistakenly) thought we were a few years away from coast-to-coast self-driving.

Nothing Tesla’s done with Robotaxi should give you confidence that this will make them a material amount of money in only “a few years.”

3 Likes
20

today’s news is 99% noise on a lucky day.

The Captain

2 Likes