We were talking Waymo, too, and you painted both with the same “dangerous” brush.
Nevermind that Tesla is now collecting suitable data with their safety monitors and safety drivers in their trials in SF and Austin.
First, it isn’t a “wrong way” as in “going the wrong way on a one-way street,” it’s changing its mind about which way to go. No traffic laws were broken that I can tell.
And, yes, I’ve seen humans do similar things, especially as Nav systems in even non-autonomous cars catch up after being screened from cell coverage inside a parking garage. You can see on the screen that the navigation on the screen hasn’t been updated with the car still in the multi-level garage.
1 Like
I’ve always thought I was human. Guess I was wrong again.
1 Like
If the board with the driving computers running the software fails, how exactly does the Tesla software perform a DDT fallback?
We don’t have to guess. The answer is it can’t.
- You’re also guessing that regulators will come to the same conclusion as you.
Again, no guessing required. The regulators have already come to the same conclusion as me. If you disagree, feel free to name all the jurisdictions that have issued L4 operating permits to Tesla.
Remember, only one year ago on the Q4 2024 earnings call Musk said the only limitations to unsupervised FSD being available in most jurisdictions by the end of the year were regulatory–not technical.
Since the end of the year has passed and unsupervised FSD is not available in any jurisdiction. Since the issue is not technical, it must be regulatory.
Unless of course you have information Elon Musk doesn’t have concluding the problem is technical, not regulatory. In that case, you may wish to send him an email explaining the situation to him.
You’re also assuming that Tesla needs current vehicles to be L4 approved by regulators to be a good investment, and that upgrading some subset of “current vehicles” is either not practical or too costly to solve the issue in time for robotaxi deployments.
Is that a bad assumption? Again, same earnings call. Here is the direct quote (emphasis added):
That same asset, the thing that – these things that already exist with no incremental cost change, just a software update, now have five times or more the utility than they currently have. I think this will be the largest asset value increase in human history. Maybe there’s something bigger but I just don’t know what it is. And so, people who would look in the rearview mirror are looking for past precedent, except I don’t think there is one.
Elon Musk plainly doesn’t think additional hardware is needed. But if it is, the need to retrofit existing vehicles changes the value proposition a lot.
Instead of simply flipping a switch and unlocking all this high-margin software subscription revenue. Instead, new hardware needs to be fabricated, tested, and installed. All that requires time and money.
As investors, we’re aware the NPV decreases exponentially with time as cash flows are pushed into the future.
If unsupervised FSD takes three years with hardware updates, that’s a much different investment case than one year without. I do not understand why this point is at all controversial.
I hear this a lot and it baffles me how people come to this conclusion. It takes a long time to get AV ride hailing permits.
1 Like
We already addressed this, why are you repeating yourself? On the board failure that happened a while back, Tesla was able to patch around it with software that actually enabled far more than just a DDT Fallback. That software, or its equivalent, is probably in all Tesla vehicles today.
They have done no such thing.
Tesla in California got regulators to approve their AV Testing Permit, which is step 1 of the CA PUC’s process, which is:
- AV Testing Permit with Driver (no paid fares)
- Driverless AV Testing Permit (no paid fares)
- Driverless AV Deployment Permit (paid fares allowed)
Step 1 is a data gathering process. If Tesla is unable to move to step 2 or step 3, then we can investigate why and discuss. Until then, your insistence that regulators have already told Tesla they need to modify their hardware is unfounded.
We don’t know the design/architecture of CyberCab. We don’t know what CC has that Tesla’s POVs don’t, and since the production line isn’t complete, there’s still time to modify its design as Tesla moves through the process. It’s unclear to me at this point whether Tesla will do its first paid non-driver on-board rides in CyberCab or Model Ys.
You may be right that Tesla will need retrofits for existing vehicles to gain L4 robotaxi approval. We don’t know. But even if you’re right, that doesn’t mean Tesla can’t have UFSD for POVs with a standy-by driver, in an L3 fashion, for existing vehicles with the then-new vehicles having whatever new hardware is needed.
Mercedes, Lucid, and others are jumping on the Nvidia Drive AGX platform to get L2++ capabilities. Tesla is already in that market. Once the first Mercedes production vehicle with Drive AGX enabled ships (promised end of this year), we’ll be able to do some side by side comparisons, and that will be interesting.
Because robotaxis will happen well before 3 years. UFSD is not the only stock price driver, and is not even first on the list.
Depending on your definition of “long time,” sure it’s a process, as I outlined generally above. But Waymo’s scaling problem isn’t permits, it’s vehicles. They’re already at about 3,000 vehicles today in 5 cities, with only maybe another 500 donor iPaces available to them. The Zeekr thing, thanks to tariffs, is unlikely to save them money, and the Hyundai versions are further out. In some cities, Waymo has had to partner with Uber since otherwise the wait times for their vehicles would be too long, again due to vehicle availability.
Here is a quote from the article:
Human vision is so much more capable than the vision of a car equipped with seven 5-megapixel cameras, only one of which is narrow-view, while all the others are wide-view. So you’re basically dispersing those 5 megapixels in a way that makes the actual effective vision more like 20/60 or 20/70. The rest of the cameras in a car like that wouldn’t even pass a DMV vision test.
I also found this comment on the article interesting:
This premise of humans only use vision and neural processing for driving is fundamentally flawed. We use many other systems, such as our vestibular system, to drive. So even if we had camera systems that were as high fidelity as the human eye (not even a great eye in the animal kingdom), it still wouldn’t be a comparable system to what an actual human uses. Also, why wouldn’t you want more sensors to make the system BETTER than human, even if the vision only premise were true?
I have been using FSD as long as it has been available, and it still consistently tries to murder me every so often. I absolute cannot see this system hitting level 4. Ever.
3 Likes
What do you mean?
Unsupervised FSD was deployed in Austin in December.
I saw it on video.
No driver!
Here’s the source, which is a video interview:
First Teslas have 8 cameras, not 7, but I guess he’s not keeping up with the latest.
Second, no LiDAR or Radar system in the world would pass any DMV / Stellen vision test, as they can’t read signs at all.
Third, his calculations appear to be wrong. This analysis indicates that a 1.7MP camera at 50º would pass the test. Whether the wide-field 120º camera needs to pass that test given the camera overlaps is an open question. I couldn’t find data on the FOV for the other Tesla cameras.
Rivian is upgrading from the current 5MP cameras in the R1 series (11 cameras for a “total 55MP”) to a “total 65MP.” Whether that’s distributed among all 11 cameras equally (5.9MP per), or concentrated on certain cameras is unknown - at least I couldn’t find that detailed information. I doubt a 5.9MP camera is that much better than a 5MP camera, and things like dynamic range, lenses, as well as subsequent image processing are certainly just as important as raw number of pixels.
At any rate, not surprising to hear Krafcik talk up his latest company’s (Rivian’s) book. It’s too bad he wasn’t asked about the Nvidia Drive AGX system. Rivian has been using Nvidia hardware for autonomy for years, but we don’t know if they’re using Nvidia’s AGX Drive OS software, and where their development stands compared to Mercedes or Lucid, for instance.
Also interesting that Krafcik still places a huge emphasis on the amount of data collected, but claims Waymo’s advantage is all the data from its sensors, not mentioning that Tesla has more data from more roads and conditions than Waymos experience.
1 Like
Here’s another quote from the article above:
Now, Krafcik is suggesting that the hardware itself is the reason for that failure. He notes that without the precise depth perception of LiDAR or the velocity data from radar, Tesla’s “cameras-only” system struggles in edge cases, like blinding sunlight, heavy rain, or low-contrast environments—that wouldn’t phase a sensor-fused system.
I think what he is saying is that it doesn’t matter how many miles TSLA has if the data collected from its cameras-only solution is compromised.
1 Like
That’s what Fred is saying, not Krafcik. Did you catch the “not mentioning” part? Did you catch that Fred was talking about Tesla “failures,” but Krafcik did not mention any failures on Tesla’s part at all? Fred is extending himself here - that’s fine as an opinion piece, but not a proper summary of the interview.
That’s why I like to go to the source, not someone’s interpretation of the actual interview. Did you watch the actual interview?
You didn’t address it. And you might be exaggerating what the software fix actually did. The Japlopnik article you linked to didn’t suggest the software fix could enable “far more than a DDT fallback” at all. In fact, it sounded like it caused a different set of problems.
Anyway, what happened was a hardware/software conflict caused a “reverse current” during the vehicle’s power-up sequence, which could short-circuit the computer board and brick the system. Tesla changed the power-up sequence to prevent the reverse current from occurring. Great. Prevented the problem from happening. I’ve never seen anything about a DDT fallback sequence regarding this issue.
However, that’s not what I’m talking about. I’m talking about what happens if the board with the driving computers fails. When that happens, the ADS system goes off line. The car is still drivable, but because the ADS system is completely non-functional, it cannot perform a DDT fallback.
In a previous post, I listed some examples of when the board can fail. The most common one (that I’m aware of) is a failure of the cooling system. This problem cannot be solved with software. If the board fails, there is no condition where the vehicle can perform a DDT fallback.
I was going by this:
Tesla’s fix was to replace the computer completely, but sources also mentioned a temporary software fix to enable some of the features in the meantime.
If some features were enabled, it would seem that a DDT fallback could be performed. I’ll grant you that’s a “maybe.”
The question as to how far redundancy or failure tolerance has to good isn’t specified by J3016. Other items in FMVSS that require redundancy, like braking with two master cylinders, allow those two cylinders to be in the same casting, so a crack in the casting could affect both. And then there’s the pedal to cylinder interface that is singular and common.
Does the ADS need better redundancy than that? Maybe, but maybe not. And if does need more, do we know if CyberCab as that or not? For instance, CC already has rear camera washers and a wiper that the rest of Tesla’s POVs don’t have.
I listened to the interview in the background while doing other stuff. I’ll take your word for it that Krafcik didn’t say that.
I just want to add on “huge data” requirements from Krafcik that Waymos have driven about 160 million miles while Teslas have driven about 7 billion - more than an order of magnitude more. And again, on more streets in more areas and certainly under more varying conditions.
Right.
Which leads me to the unanswered question that no one will/can answer.
How did Waymo get to L4 autonomy years ago with so much fewer miles?
And now Waymo is generalizing across cities and onto highways and to international locales and to northern cities with worse weather while Tesla is not accruing autonomous mileage even in temperate California and southern Texas and saying they need even more miles to solve the long tail and the march of 9s to sufficient safety.
What’s Waymo doing with vastly fewer miles that Tesla cannot do without 10 billion miles?
Seems like Waymo has a legitimate and demonstrated capability advantage. The technology is more capable and the team is more capable.
That could affect the business case.
1 Like
Not only fewer miles, but geofenced and some routes far more popular than others and pre-mapped in greater detail … oh look, a simpler problem!
Just ran a Google search on this and Waymo tested on billions of simulated miles. That probably helped.
1 Like
Really, you can’t think about what the answer is?
How about:
- Waymo didn’t have to worry about any use cases involving freeways, rural roads, and could even exclude some roads in cities (like Lombard St in San Francisco).
- Waymo didn’t have to worry about speeds greater than 35 mph.
- Waymo only had to worry about a very limited geographical area (1 city at first, 5 cities in 5 years).
- Waymos have put themselves in situations where a human FSD driver would intervene, but Waymos either struggle through it or get remote assistance that Teslas don’t yet have, being POVs.
- Waymo has a less-scalable system involving pre-mapping and more expensive vehicles.
Again, doing things in different orders means you can’t compare “who’s ahead” in terms of technology or overall capability.
For instance, Waymo can’t drive coast to coast. FSD can, and has, multiple times. So, exactly who has more “overall capability?”
I’m not saying the Waymo system is bad. I’m saying it’s different, was designed to do different things. Tesla does the things FSD was designed to do better than Waymo; Waymo does the things it was designed to do better than Tesla. The question still remains, what the end state will be, which includes whether Waymo can scale up faster than 1 city in over a year for the past 5 years. They’re claiming 40 cities by end of the year, but I’ll take a gentlemen’s bet they won’t be charging for driverless rides in 40 or more different cities by Dec 21, 2026.
2 Likes
Waymo rides can now take freeways in Phoenix, Los Angeles, and San Francisco.
2 Likes