@diplomat33 , thanks for the thorough reply. Now I can contextualize your opinions better.
We're on the same page on many things. Waymo is indeed trying to be a robotaxi service in the most dense of places, where the total adressable market is greatest compared to the area they have to map. And they are making excellent strides in doing so.
In my book that however is not "solving L5", which is Tesla's goal. That goal is further away, and a lot more 'sci-fi'/utopian.
Where you state : "
If you think Tesla is close to removing driver supervision in the entire US, you are deceiving yourself." I'd like to point out I in fact
do not think Tesla is close to removing driver supervision, not in the US, not anywhere.
Where we still disagree is that L5 worldwide is possible or not. You say nobody is able/will be able to scale worldwide. I think that is a gross underestimation of the advances in AI. We are only at the very beginning of the AI revolution.
Will it take longer than most expect? Of course. But not insanely long.
Personal feeling:
maybe within the next 5 years (but low likelihood is my hunch), very likely within 10 years, and I'd be shocked if L5 isn't achieved in 20 years. (Not necessarily by Tesla).
I am in the camp of "L5 will only be unique to one company for a very limited amount of time". Once one company shows they've done it, the door is open for others to follow suit. The second-movers will have a lot less legal hurdles than the first mover, and can start developing based on what is known about the tech of the first-mover (with the state-of-the-art hardware and software that the first-mover didn't have at the time of its development process).
That's why I think ARK's insane valuation of Tesla due to FSD ride-hailing is kind of unrealistic. Their argument however is that, even though L5 is achieved by multiple companies, the first 5-10 years there won't be enough robotaxis to fulfill demand for cheap transport, thus any player with L5 can profit.
Either way, we are living in exciting times.
Thanks for the equally respectful and thoughtful reply.
The SAE essentially defines L5 as equivalent to a typical human driver. They may not always be good drivers but they can drive anywhere, anytime, as long as the roads are drivable of course. That is why the SAE specifically says that L5 only needs to drive when a typical human driver could drive. So for example, L5 is not expected to drive in a white out blizzard. L5 is certainly a laudable goal. I definitely see the appeal of L5, of autonomous driving that is like a human driver in that it could drive anywhere, anytime as long as the roads are drivable. And I agree that eventually AI will be good enough for L5. But here is why I don't think L5 is really needed. I see L5 as more of a long term end goal.
But note that the SAE does not say anything about how good the L5 has to be. The big challenge is not just developing autonomous driving that is capable of L5 but making it safe and reliable. And of course, to be useful, it needs to be safer than human drivers. Most carmakers want their autonomous driving to be at least 10x safer than humans. So there is a lot of testing and validation before you can deploy autonomous driving. Testing and validating L5 is exponentially harder than L4 because you have to test and validate everywhere and anytime. It is much easier to test and validate L4 since you only need to make sure it is safer than humans in a specific ODD. That is another reason why I see L4 as a better path to L5. You can start with L4 and expand the ODD over time until it is essentially equivalent to L5. I think that will be easier than trying to test and validate L5 directly.
This is also why I say "L5 worldwide" is not going to happen. Even if we get generalized AI that is capable of driving anywhere in the world, there needs to be a lot of testing and validation before companies would deploy it unsupervised. And testing and validating that autonomous driving is safer than humans on literally every road, every conditions, in the entire world would be insane. And carmakers won't deploy their cars in literally every single country. Furthermore, the SAE defines as L5 as per country or per region. So I think it makes sense to talk about L5 for a country. Carmakers will eventually deploy L5 in specific countries. Over time, more and more countries will get L5. But I don't know if literally every single country will get L5 to be truly "L5 worldwide". But perhaps you would consider it close enough to "L5 worldwide" once enough countries have it, even if it not every single country does?
Another important point:
L4 and L5 are actually the same autonomous driving. They are both full autonomy, meaning no human driver needed. The only difference is the ODD (when and where the system can be used). L4 is just L5 with a smaller ODD. But L4 does not have to be a tiny geofence. Technically, the ODD of L4 can be anything from a single street to 99.999% of L5. Also, the ODD limit does not have to be geographical. So for example, you could have L4 that works in the entire US, except for one city block, and it would technically be L4. or you could have L4 that works everywhere but cannot work when it rains or cannot work at night and it would be L4 even though it works in the same geolocations as L5. The only real advantage of L5 is that you don't have to specify the ODD since it works "everywhere" and "anytime". With L4, you do need to specify the ODD so you need to know what limits it has. But I hope you would agree that if L4 had a ODD that is say 99% of L5, for all practical purposes, that would be close enough to L5 for most people. And if the L4 can meet all the personal driving needs of an individual, it is essentially "L5" to that person since they never need to take over. So this is why I don't think we really need L5. we only really need L4 in a big enough ODD to be useful and practical and work towards L5 ODD.
With this in mind, I don't think it is quite accurate to say that Tesla is solving L5 while Waymo is solving L4. IMO, it would be more accurate to say that Tesla is trying to deploy L5 while Waymo is focused on deploying L4. But both Tesla and Waymo are trying to solve full autonomy. The deployment ODD is just different. I say Waymo is solving full autonomy because the Waymo Driver has complete perception, prediction and planning designed to handle all driving tasks. In fact, Waymo has demonstrated that their autonomous driving can handle a pretty wide ODD, from rural, suburbs to dense urban and highways and also inclement weather. Waymo is just limiting the full autonomy to a geofenced ODD, so we call it L4. But it is only L4 because the ODD is restricted, not because it is not full autonomy. Waymo is not interested in deploying L5 but they are interested in full autonomy and expanding the ODD over time.
Tesla and Waymo just have different approaches to solving full autonomy. Tesla wants to deploy L5 and their approach is to start with L2 but in a "L5-like" ODD, ie everywhere/anytime, and then work towards full autonomy. This makes perfect sense for consumers who want to go anywhere, anytime and will be in the driver seat anyway to supervise. The Waymo approach is to start with full autonomy but restrict the ODD and expand the ODD over time. This makes sense for a robotaxi model where you need to do driverless and only need to offer ride-hailing in a metro area where your customers are.
There has been a lot of debate about the SAE levels. Some want to get rid of L5 entirely since they see it as a uneccesary. Once "L4" has a ODD that is say 99% of L5, then doesn't it become indistinguishable from L5 for all practical purposes? In fact, wouldn't we probably just call it L5 at that point anyway since it is so close to L5?
So personally, I think it might be better just to think in terms of full autonomy, ie no human driver, and then define the ODD rather than talk about L4 or L5. The goal should be full autonomy with a big enough ODD to be practical and useful. That is one reason I really like Mobileye's taxonomy of "eyes on" or "eyes off" with well defined ODD. I think that is clearer to the consumer. Consumers basically want to know if they need to supervise the system or not and where and when can they use the system. And if the ODD is big enough to meet all their personal driving needs, they probably won't care if the system is technically only L4.
Sorry for the long post but this topic is a passion of mine.
![Smile :) :)](data:image/gif;base64,R0lGODlhAQABAIAAAAAAAP///yH5BAEAAAAALAAAAAABAAEAAAIBRAA7)