odysseus2000 wrote:According to Wiki:
https://en.wikipedia.org/wiki/Motor_veh ... S._by_yearthere were in 2019 36,120 deaths on US roads = 36120/365 = 98 deaths per day. Yet a Tesla accident where apparently the driver was not in control of the vehicle, in violation of Tesla recommendations, becomes from page news, the other on average 97 deaths do not.
The measurements from Tesla show an accident rate with human combined with autopilot of approximately 10x less accidents based on distance between accidents. If we equate accidents to deaths then if the entire US fleet was driving with autopilot the death rate could be reduced to about 10 per day. This would be a very significant reduction in human misery and economic costs.
Whether complete auto pilot (no human driver) happens we still have a technology that has been demonstrated to be better than human alone driving by close to one order of magnitude.
The neural net technology used by Tesla came into existence about 10 years ago and there is no data/theory that I am aware of which limits this technology. Unless some currently unknown limit to the technology emerges, we can expect the software to continually improve, especially as the labelling of images is now being done automatically.
It is easy to argue that the technology will never reach a mass market and be allowed by regulators based on previous experiences and prejudice, but that is not a scientific way to think about new technologies.
Regards,
But 'self driving' and 'full self driving' aren't promising 'driver assistance'.
They're the promise of driver replacement - have a nap or read a book while the car gets you home through the rush hour traffic jams.
While I don't strictly agree with the exact claim that 'there are no limits to the technology', I do firmly (as someone qualified in the field of AI) believe the current state of the art is perfectly capable of producing a self driving car (without the need for human in the driver seat) that can drive better than a human would.
I just don't believe that Tesla are anywhere as near perfecting it (or even just making it acceptable for a first generation) as quite a few people believe, and I wouldn't be surprised if Tesla find themselves wrong-footed on their choice of hardware.
i.e. I agree...
"The neural net technology used by Tesla came into existence about 10 years ago and there is no data/theory"
So on what basis can Tesla know that the hardware computing capabilities they are putting into their cars for which they seem to be promising a 'software only' upgrade to FSD, will actually be up to the job? I don't believe they can know it until they actually get a proven, working FSD using it. And they don't have that yet. So how can they know that their hardware will be adequate?
I think a lot of the Tesla fans misjudge the comparison with Waymo. A lot of Tesla fans seem to think that 'driver-assisted-but-not-geofenced' is somehow better than 'no-driver-but-geofenced'.
I think those people misunderstand the effort involved in those final few 'tweaks'. It's not unknown for development attempts to completely hit a brick wall (metaphorically speaking, no pun intended) when trying to go that final distance. It can be like when you're a kid trying to reach under a fence to retrieve something... you stretch your arm as much as you can and you can just about touch what you're reaching for ... but you can't quite get the extra distance you need to grab it. You have no option but to withdraw your arm and either go and ask your neighbour, or at least find an additional tool to get you the final distance.
It's all about the foundation. Is it robust enough. Is it expressive enough. Is it powerful enough. There is no data / theory, so how can Tesla know for sure?
Waymo, having a geofenced-but-no-driver service running, have proven that they can go all the way (to FSD), albeit in limited circumstances.
Telsa on the other hand have gone for a broad-based, lots of foundation, but haven't gone the whole way (to FSD) in anything - nowhere have they proven that the foundation is fit to support the ultimate aims.
Think of it like building a city of skyscrapers.
Waymo have built one skyscraper all the way to completion. Waymo have proven that skyscraper's foundation can support it. They done full stack beginning to end for at least one stack. It's true that the other stacks won't be exact replicas and may even be 'taller'. But don't underestimate the significance of proving that one stack all the way to real world implementation, developed far enough to completely remove the driver from the driving seat. That last leap from safety driver to no safety driver was no small step for mankind!
Telsa on the other hand have put multiple half built skyscrapers out. None of them have reached their final height. It's not yet proven whether Telsa's foundation will be able to support the final skyscrapers. There's only two things to driving - speed and direction, and just watch the Tesla videos, and you regularly see the driver dialing up or down the speed on autopilot.
Unlike real skyscrapers where the engineering and theory is well established, like you yourself say, with the neural net technology... "there is no data / theory". There's no data or theory that you can look to when you pour the cement of the foundations (or sell lots of cars with the supposed 'final' hardware already in place)
One final thought.
A lot of Tesla fans proclaim about how Tesla are getting huge amounts of real world data from their cars which Waymo with its limited fleet aren't able to match.
I saw a video a few months ago where a Telsa representative (giving the speech) put up a slide, proclaiming about how many millions Tesla had vs Waymo. If forget the exact numbers, but obviously being a Tesla presentation the Tesla miles driven was a couple of orders of magnitude higher than Waymo.
I couldn't help think to myself... yes, but ...
Waymo = "Keep hands
off the steering wheel"
Tesla = "Keep hands
on the steering wheel"
Waymo might have less data, but they seem to be getting far more out of the data that they do have.
If you're interested, there's a guy with a channel largely dedicated to riding in Waymo cars - he isn't paid by Waymo, and is paying to ride just like any other public riders.
https://www.youtube.com/c/JJRicksStudios/videosBut be warned.
Compared to the Tesla videos they can be quite boring, because, well, the Waymo cars tend to just, how can I put this, "work" without any feeling of will it or won't it. They don't quite have the feeling of jeopardy that you get with the Tesla videos.
But they can do some pretty smart stuff. Identifying and reacting appropriately to emergency vehicles, road works, etc.
I mean, put it this way... it's a huge leap from safety driver to no safety driver. To remove the safety driver
and put public passengers in the car,
and allow it to drive in supermarket car parks right up at the front where people are walking in and out of the store all the time, you really do have to believe it will acceptably handle all situations it's ever likely to encounter. And that's no mean feat! Don't underestimate the significance of that.