Tesla FSD already works everywhere, even on unpaved roads. It just doesn’t work as well as Waymo.
Waymo works very well, just not in as many places as Tesla.
You might bet on Waymo because they have a fully working product already, but I’m betting on Tesla because of the vast amount of training data they are collecting. There’s a bitter lesson here.
> the vast amount of training data they are collecting
They keep pushing this point. And they do appear to be collecting an absolute firehose of data from the millions of vehicles they have on the road. By comparison, Waymo collects a lot less data from many fewer vehicles.
Which leads to some tough questions about Tesla's tech. If they have (conservatively) 10x the training data that Waymo has, why can't their product perform as well as Waymo? Do they need 100x? 1,000x? 10,000x?
Assuming they were at parity with Waymo today, this would suggest that their AI is only at best 10% as effective as Waymo's, and possibly more like 1% or 0.1% or whatever. But since they can't achieve parity, it's not even possible to bound it.
It's entirely possible that their current stack cannot solve the problem of autonomous driving any more than the expert systems of the 60s could do speech translation.
I haven't heard a compelling argument as to why a system that is at best 10% as effective would ever be expected to be the leader.
Data isn’t as useful here as in other domains, since when you change the car’s behavior even a tiny bit, a lot of the timeseries is invalidated. It’s not evergreen, and it can be quite subjective what it means to “pass” a scenario that one previously failed.
Also, Tesla collects data from its fleet, but that data’s fidelity is likely quite limited compared to other companies, because of bandwidth if nothing else. Waymo can easily store every lidar point cloud of every frame of driving.
FSD and Waymo are completely different products. FSD isn't even autonomous, as the user manual reminds you:
Always remember that Full Self-Driving (Supervised) (also known as Autosteer on City Streets) does not make Model Y autonomous and requires a fully attentive driver who is ready to take immediate action at all times.
I hope for the best for Tesla, but they are many years behind Waymo. The world definitely needs a second working self-driving system! Right now comparing Tesla and Waymo is nonsensical. Once you can sit in the backseat of a Tesla while it drives there might be some worthy comparisons to be made.
My definition of "works" includes the fact that a self-driving car will never drive into a parked fire truck, or many other things i've seen tesla FSD do.
Well and because they actually have real self driving cars without a safety driver. Tesla doesn't have that and only has demoed it in very specific scenarios.
I own a Tesla, though I don't own FSD, but this year, Tesla has given all cars a trial of FSD on two occasions. It works remarkably well. I backed out of my driveway, then enabled FSD and it drove all the way across Portland to a friend's place with zero intervention. It was about a 15 mile, 30 minute drive.
It navigated neighborhood roads without markings and tons of cars parked on the curb. It got onto the freeway and navigated, including changing lanes to overtake slow traffic. Once I got to their place, I was able to tell it to automatically parallel park on the curb.
As far as I'm concerned, Tesla has fulfilled their promise of full self driving. The "supervised" requirement is basically just being used as a legal loophole to avoid liability if it fails.
> The "supervised" requirement is basically just being used as a legal loophole to avoid liability if it fails.
"If it fails" - so it is supervised for a reason then. It makes sense because FSD has an intervention rate in the low double digits according to community trackers like https://teslafsdtracker.com.
I don't understand the "bitter lesson" reference here. The bitter lesson is that general methods of computation are more effective. How is one of the two not using general methods?
My understanding is that Waymo is applying specialized centimeter-scale mapping and lidar to achieve superior results.
In contrast, Tesla is using dumb cameras and just dumping boatloads of data into their model. It’s a more general solution. Maybe the reference doesn’t fit perfectly - the model architecture is likely similar under the hood - but there’s some analogy there.
Just saying they have better results because of mapping and lidar is incredibly reductive. They have an extremely sophisticated AI/ML stack and simulators.
One challenge for Tesla and Waymo has been the piecemeal permitting process. Even though California gave Waymo a statewide permit they have still needed to work through various cities/counties for permits. I imagine one goal of Musk's is to make that all go away sometime next year. I'm not making a comment on whether I agree that is a good idea. Just speculating.
If training data is such an edge for Tesla, how is it that Waymo works so much better than FSD with only 1/1000th the data?
I also don't see any evidence that Waymo can't work anywhere. They recently expanded to Austin, and it seems that it immediately drives better than FSD.
Tesla FSD will never catch up to Waymo until they switch to LIDAR and have human assistance when the vehicle gets into complicated scenarios such as emergency vehicles blocking the road and redirecting traffic.
I'm betting on Tesla not for the technology, but because President Quid Pro Bro is probably going to issue an Executive Order that turns every Tesla company into a federally blessed monopoly.
Indeed, it was not very long ago that FSD would happily take a straight line through a roundabout, lanes or "skirts" be damned, essentially treating it as a standard intersection.
Training data is trivial to collect. Betting on Tesla because they have more training data than Waymo is like betting on Roscosmos because they have more employees than SpaceX.
It looks similar to a cable/fiber rollout where they have to onboard each region individually. I know they are currently doing this in the Atlanta metro.
Like cable/fiber, once they have good models of the business and what it costs to roll out, they have the freedom to accelerate and do regions in parallel. If the business works, I would expect them to scale the pace of rollout.
This is the root of the misunderstanding, I think. You’re begging the question.
More data does not necessarily mean better data. You can collect many more individual driver experiences, but if they do not have sufficient resolution in the necessary dimensions, they may never provide “better data.” Similarly, even if the magic data is hidden somewhere in there, if the model cannot practically extract the insight because of their sizes/disorganization vs the computational/storage capacity, this too would mean they are not better data.
Of course you can make the argument that some of the sensors are unnecessary, but when one fleet has had millions of vehicles for years and isn’t working, and one started with dozens, has recently grown to one thousand vehicles, and is working, the evidence is not in support of the argument.
Waymo was able to do this with less miles. How much data does Tesla really need at this point? Assume you have all the data in the world that you could ever possibly want. How much of that can you really compress into a car for real FSD?
Tesla was supposed to have what they needed when they released the Model 3. Then they had to upgrade the cameras and CPU which meant they had to re train. Then they re-wrote, so again retrain. Now it's new cameras and compute again. Cycle repeats.
How over-fitted are their models to the cameras? I'd expect a layered architecture where a sensor layer does object-recognition and classification and then hands over this representation of the world to a higher-level planning model. You should have to retrain the whole stack for camera revisions - hell that's how it would work across car models with their different camera angles.
Uhm, we already have FSDs, both the USA and China, just not the Tesla FSD. China is running auto taxis in a few limited areas with full setups that rely on LIDAR, and I hear they are pretty good.
There are some "where" issues with Tesla too. I have an intersection where it consistently can't tell that its view is obstructed. It'll just yolo into the intersection then pause (after pulling out into the lane) when it realizes that it wasn't actually able to see. Its consistent behavior, and seems to be a flaw with obstruction detection.
I might argue that every traffic light is sort of a where too. Mystery meat yellow light handling is scarily bad.
Waymo vs Tesla definitely smells like bitter lesson to me yes, 100%. With Waymo being on the bitter side, to be clear. Future will tell if the intuition is right on this one
FWIW, Waymo has more cameras than a Tesla. Both companies are removing sensors over time. In some ways removing sensors is easier to prove out with real-life data than adding them. I think it is going to be fascinating to see how it plays out.
Tesla added back the radar and improved the cameras in HW4. My guess is that ultimately they'll converge to a similarly capable sensor/compute suite with Tesla improving theirs and Waymo paring down.
Well, one thing is that many of us rarely take taxis. (Aside from reserved private cars to the airport now and then.) I'm unconvinced that self-driving changes the equation enough for most of us. I do have a trip coming up that 50% cheaper Uber might lead me to not rent a car but that's rare.
Car ownership is pretty high in a lot of places. It's still pretty much central urban--lower ownership, harder to re-find a parking spot--vs. everywhere else. Taxis may be more common in some places but there are still a lot of privately owned cars and taking taxis or having drivers is really not the norm in most places.
Tesla FSD already works everywhere, even on unpaved roads. It just doesn’t work as well as Waymo.
Waymo works very well, just not in as many places as Tesla.
You might bet on Waymo because they have a fully working product already, but I’m betting on Tesla because of the vast amount of training data they are collecting. There’s a bitter lesson here.