Tesla is betting that $200 worth of cameras can beat Waymo’s $75,000 sensor array. The difference is the 10 billion miles of user data that Tesla has collected.
If Waymo wins, FSD is a bespoke, geofenced taxi service in a few large cities. If Tesla wins, FSD is everywhere at low cost.
That’s probably why the stock hasn’t cratered with the years of “delay”. The upside is significant.
intercst
I didn’t watch the video but that’s a typical Tesla bull argument. However, it is also a false dilemma fallacy. Another option is that if Waymo wins their version of L4 ADAS will be everywhere at low cost.
In various interviews, Waymo co-CEO Dmitri Dolgov has said the Waymo Driver doesn’t strictly need Lidar or geomapping to operate, but they use both for safety and performance reasons.
I mention that because right now the Tesla robotaxis without human operators only operate in geomapped areas. Unlike Waymo however, Teslas robotaxis cannot operate in bad weather or at night. So right now, the Waymo tech is closer to a generalized solution than Tesla’s.
Probably several errors here, but I’ll just cover this one.
Why do you believe this 10 billion miles is an advantage when Waymo is 200 million miles (and accelerating) ahead of Tesla in actual autonomous miles with customer rides available in 11 metro areas, including some highways and airports?
Don’t take seriously someone who has been materially wrong about many thngs for at least a decade, in no area more so than AI.
Look at what is actually happening, not what someone claims.
Use your skim, scam and fraud detector.
Alphabet is a leader in many things AI, including some foundational machine learning methods that form the basis for AI processes like object perception.
Waymo can simulate sensor data from as many miles and scenarios (including edge cases) as they need to train their AI driver.
They are clever and Alphabet can spend more in one quarter on capex than all Tesla-related companies can spend in an entire year (and still print $40 billion in net income in a single quarter).
This thread has more:
I am. Everyone thought Elon was over promising on the rockets until he landed the boosters in tandem. That made the experts at Boeing and NASA look very foolish.
Is highly likely you’ll see the same on FSD.
intercst
Maybe, but the evidence says no so far and Waymo is accelerating ahead from Tesla every day.
I would also note that Waymo is iterating on their hardware, with next generation versions with fewer sensors and materially lower cost equipment.
If Tesla was alone and didn’t have a Waymo to catch, and now also Zoox, I might think differently.
I believe it’s a mistake to extrapolate some success in one area (rockets) to another (autonomous vehicles).
I don’t know much about rockets, but I’d bet AI is a materially newer and different technology with less societal knowledge than what was required to build EVs and land rockets.
Meta didn’t make the metaverse happen even though they made facebook, but they got a new company name.
The Captain
I would change “highly likely” to “possible”.
We all remember the triumph Apple had with the Lisa, that Microsoft had with the Zune and the Windows Phone, that Google had with Google+ and Buzz, Sega with the Dreamcast, Facebook with VR, Amazon with the Fire Phone, New Coke, the Arch Deluxe, the Edsel, and the famous RJR Tobacco smokeless cigarette.
These were all highly successful, some even entrepreneurial companies, yet managed to launch dramatically unsuccessful products even with the brightest minds and marketing muscle available to them.
If Tesla is right, it could be a big win, but I have to wonder how they are going to navigate the (not very) edge cases of rain, snow, and fog. (If they can’t, it’s game over, I would think.) The theory is “Well if a human can do it, then a camera can do it too” is simply wrong, in my view (and I have a bit of experience with video camera abilities.) The costs of Lidar have cratered: from $75,000 per unit 10 years ago to $150-$1000 per today, and the price continues to drop. (Heck, they’re putting them on self-mowing lawn mowers, fer cripes sake.)
If, in fact, “camera only” turns out to be the wrong alley, then all of Tesla’s miles are worthless, and they will have to start the training all over again (or mostly, anyway.) I suspect the “camera only” approach was decided back when Lidar was prohibitively expensive, and now that Tesla is locked into that path (by dint of not having any camera+Lidar miles) they are forced to proclaim Lidar is unnecessary, lest they have to start over, well behind several other contenders.
Everyone thought Elon was overpromising on the solar roof, until…well, it turns out he was.
Everyone thought Elon was overpromising when he said Tesla’s sales would grow 50% per year until it reached 20 million units by 2030, until…well, it turns out he was.
Everyone thought Elon was overpromising when he said they could sell a Cybertruck with 500 miles of range for less than $70K, until…well, it turns out he was.
It’s highly likely you’ll see the same on FSD. That the promises outstrip the capabilities of the tech. Heck, we’ve already seen that - the promises about the self-driving capabilities of the HW2 and HW3 configurations were false, and it’s fairly probable that most of the HW4 cars also lack the hardware to actually self-drive (no front bumper camera, inadequate self-cleaning capabilities in the camera).
As for Waymo, their sensor array doesn’t cost $75K anymore. As LIDAR has been increasingly adopted in so many ADAS systems and gets high-volume commoditization, the price has plummeted. By the time Waymo gets into having built-for-purpose vehicles (which should be in a few years), the difference in vehicle cost should be pretty modest. Tesla was betting big on camera-only allowing the existing, entire Tesla fleet to be used for Robotaxi service - so the fact that they flubbed that eliminates a great deal of their competitive advantage vs. their competitors.
Third option; China provides very cheap LiDAR.
One can only simulate what one knows about … one of the virtues of real miles driven is that one encounters situations one didn’t previously know about.
Correct, but the vast majority of those miles where likely recorded under HW3 and without the benefit of a front bumper camera - so there is a significant amount of missing data that Tesla does not have that they need for FSD.
Footnote: I have read online (I have not watched the videos) that as recently as three months ago, FSD was not even using the new front cameras and that influencers on youtube had covered the camera and FSD continued to work.
If that is valid, then we are still a long way away from having miles that fully count.
While one clearly wants footage including the front camera, I suggest that it is far too harsh to exclude any footage which doesn’t include one. For starters, lots of stuff happens at the side and back. Also, the prior configuration was hardly blind facing forward.
It would not be excluded, it just would not be sufficient. Or perhaps more accurately, deficient.
In other words, we might need ANOTHER 10 billion millions with front cameras before FSD is fully functional.
My question is when does Tesla start taking advantage of all these bazillions of FSD miles?
Thus far, they have driven 10 bazillion FSD miles and this allows them operate driverless vehicles in three tiny geomapped areas, only during daylight hours, and only in good weather.
How many do they need to operate at night or on the freeway? 20 bazillion? 50?
Sure, adding Lidar would be expensive, but geomapping every city prior to enabling FSD is expensive too.
In other words, you are excluding the existing data from having any value.
The problem is the long tail. The first few miles are very rich in data but as miles accumulate it gets harder and harder to get new edge cases to train the AI.
Google AI:
Yes, the “long tail” problem is the primary bottleneck for achieving fully autonomous, Level 5 systems. As miles accumulate, the return on investment for data collection drops, as most new data is mundane and redundant.
To overcome this by 2026, the industry is shifting from pure data collection to data curation and synthetic generation.
Strategies to Solve the Long-Tail Problem (2026)
Generative AI & Synthetic Data Generation: Instead of driving millions of miles to find a rare event, platforms like NVIDIA Omniverse generate it. Tools are used to create photorealistic, simulated scenarios such as a pedestrian in a costume, extreme weather, or rare sensor failures.
AI-Guided Data Augmentation: New frameworks like LTDA-Drive use LLMs to intelligently replace “head-class” objects (like cars) with “tail-class” objects (like specialized bicycles) in real driving data to create high-quality training samples, resulting in over \(30\%\) improvement in rare class detection.
Active Learning (Active Curation): Instead of training on all data, systems actively identify and flag “uncertain” scenarios. Tools like FiftyOne allow teams to find and prioritize data that represents model blind spots.
Simulation Replay & Variation: Real-world failures or near-misses are captured, then brought into simulation and replayed thousands of times with slight variations (e.g., different lighting, speed, vehicle type) to test robustness.
Scenario-Based Testing: Focusing on building a parametric representation of a scene, reinforcement learning can be used to tweak parameters until the system fails, generating new, high-risk scenarios.
The Shift in Focus
Phase Focus Data Source Past Volume Massive real-world driving data Current Quality Targeted edge case acquisition Future Diversity Synthetic simulation & generative augmentation
By 2026, synthetic data is no longer just a testing convenience; it is a primary training mechanism to bridge the gap between “90% accurate” and “completely safe.”
The Captain
No, it has value, clearly it has helped establish level 2. But without XXXXX miles from the front camera, that data will never allow them to achieve level 4. It will always be deficient.
Edit: As Captain illustrated, it is a March of Nines. Tesla has billions of miles of data but virtually zero miles of data from the front camera. They are JUST STARTING their March of Nines for the front camera. Now, it certainly should not take them as long but it may indeed take just as many miles because at the beginning, everything is new.
The video mentions that. Tesla FSD now infers a “bird’s eye view” from the camera information it has.
If it sees a pedestrian approaching from the right, it waits until it reappears on the left side of the vehicle. And you can follow that in the birds eye view on the flat panel screen.
That’s the AI innovation here. The software is predicting the behavior of the pedestrian even if it’s not seeing it in a front camera.
And that’s the reason Elon said he needed 10 billion miles of user data “to catch all the edge cases”
intercst
As it turns out, as per the Tesla Safety Page, FSD has now exceeded 10 billion miles.
I sure don’t remember it that way. Do I remember correctly that Blue Origin first did controlled landing before SpaceX? By the way, what difference is tandem supposed to make other than just looking nice?