If you can eventually guarantee you’ll get a robotaxi instead of a car with a driver when you call one (unlike Waymo in Austin, now, thanks to their dumb partnership with Uber), they’ll likely do ok in the longer term.
I'm rooting for Tesla being able to provide some driverless taxi competition to Waymo. Hopefully both can increase supply enough to allow for a guaranteed robotaxi booking, although anecdotally I haven't found it that hard to get a Waymo through the Uber app depending on the time.
I’ve only tried twice (after using it ~30 times during the preview directly through the Waymo app) - neither time resulted in a robotaxi though, and one resulted in a driver that smelled so bad I’ve sworn off ride sharing until the end of summer.
Hopefully the Tesla app forces Waymo to reconsider their deal with Uber altogether and just run it themselves like they do in San Francisco.
Let’s give it a little bit of time to see, but gating the release of self driving taxis to sycophants is not the choice you make when you are confident with a product…
> gating the release of self driving taxis to sycophants
Are they sycophants or are the other sites haters? Take a gander at every tech website's headlines regarding Tesla for the last few years and tell me they aren't following the "if it bleeds, it leads" mantra.
The people invited to preview Robotaxi were not only unedited, they were also allowed to live stream.
Are you looking at the feeds from the cars to reach 32? Each screen is from the cameras around a single car, so you're actually only seeing 5 cars from that.
"Tesla has self-driven 4B miles in every corner of the USA, Waymo has only self-driven 0.1B miles in select cities, how pathetic! They are so far behind Tesla it isn't even funny!"
- Me, if I wanted to be equally ham-handed in throwing the comparison for Tesla.
In reality, we are seeing two bets on two different approaches. Do you scale up supervised driving to maximize data collection/diversity and then go unsupervised? Or do you go unsupervised and then scale up problem solving as you go? The cool thing is that both of these approaches are being tried so we will find out. If you want to place a bet of your own, you can find the casino in your favorite brokerage app!
In any case, the videos of what is possible with supervised FSD are quite amazing, certainly not "pathetic," and what remains to be seen is if they can successfully navigate the supervised->unsupervised jump, which is certainly not trivial.
If I said I've never had a driver's licence, can we get back to the topic at hand?
What's the thesis here, humans are imperfect drivers and therefore we should accept self driving cars driving into incoming traffic, and if anyone objects, focus on their faults instead?
Humans on the aggregate can definitely get better, that's why Norway has 2.14 / 100k traffic-related deaths per year and the United States has 12.84 / 100k.
Norwegians drive half as much. Even if we ignore the other components of the gap, closing that 50% will require rebuilding 300000 square km of urban sprawl into dense urban centers with good public transit. It's a nice thought, but my money is on the robots.
The "human drivers are shit" problem kills a million people every year. Solving it is good. If we let the great be the enemy of the good we will never fix anything.
It's not a binary. That particular case is pretty egregious since the car signals to turn, gets into the turn lane, actually starts turning, and then swerves back (across the oncoming traffic lanes in the middle of an intersection!) and tries to get back into the correct lane to drive straight, but only manages to achieve that after driving for a bit in the lane for oncoming traffic. That is a major fuckup. And this all happens on a road with fresh clear marking, on a sunny day with perfect visibility.
I believe that driving is a privilege, not a right.
However, I'm not saying that their licence should be taken away (we've all made mistakes - I should have made that clear), but if that is representative, it is dangerous. I rode motorbikes for many years, and you can spot people who are dangerous very quickly from their 'car body language'. And an error like that in another circumstance could kill someone.
I knew people whose driving had deteriorated like this (my late mother and her friends spring to mind). They refused to accept they were not capable any more, and she used to do things that terrified me. I had discussions (kindly!) about her driving standards and errors and she refused to accept that she was making the errors she was. And they were not as bad as what I saw in the video.
License should be just revoked instantly for the sake of Tesla. At least with not fully autonomous I can override it. They need to get a feedback that they are not ready.
If I was in a car with a human driving that badly, I'd leave ASAP! No-one beyond a learner would traverse in such an uncertain, jerky way. And this is on a day with perfect visibility and no adverse conditions that are visible in the video.
Anyone who owns a Tesla with “full self driving” knows how this is going to go. Teslas using only camera vision just don’t have the sensor package and programming to actually perform self driving. I just don’t see this going well for Tesla as it’s more likely to reveal the weaknesses in their technology than be a showcase.
But if you're correct, then won't it be a massive harm to the valuation once they are at fault in many accidents over the coming weeks? They have the Q2 earnings call coming up next month.
It seems like it would be a really bad idea to put this out there if they know from the data that it will cause lots of accidents.
It depends on how much autonomous driving they actually do. My guess is that they will only drive a miniscule amount in favorable conditions for the foreseeable future. That would maximize the hype:risk ratio.
It's also possible they quietly rely on remote operators to a large extent. The fake robot PR stunt wasn't even that long ago, the company might try the same thing here.
Yeah, well, companies like Enron show that covering up mistakes with other mistakes ends badly. Fake it til you make it works if you make it, blows up otherwise.
So you think they chose to launch it yesterday despite the data telling them that they will be at fault in lots of accidents over the coming weeks, and will presumably be forced to roll it back? You think they decided that's somehow better than delaying?
I mean, if that were to happen, they wouldn't exactly be the first. It's not particularly rational, but "launch very late thing in an unacceptable state and then pull it a few weeks after" is not an uncommon industry pattern. Admittedly, it's unusual for _hardware_ (though not totally unknown; see the Humane AI thing, say).
Pulling something a few weeks later isn't uncommon when it's unknown how a service will go, but in this case they actually have data that pretty much tells them how this is likely to go.
So you think Elon is not interested in numbers? Or you think Tesla employees are presenting fabricated data to him?
The phenomenon in that video is only possible when management is heavily disconnected from ground truths & data, but that seems the opposite to how Musk's companies are run.
Honestly, I have been hearing this since years now. I was suspicious about it for a long time. But they have been allowing vision based FSD for a long time now and the data shows it's getting better, fast. I am ready to park my skepticism (sometimes cynicism) for some time and give them a benefit of doubt out of optimism. Tesla knows that if things go wrong with Robotaxi launch, they are screwed. They wont' take that risk unnecessarily.
Waymo has had many such incidents like this, but they're still (correctly) considered much safer than human drivers.
Whilst a wheel wobble & veering somewhere it's not supposed to go looks bad, it's very difficult to do worse than the average human driver in terms of safety.
It's actually the opposite. Humans, for all their faults, are amazingly safe drivers. Depending on exactly how you choose to measure things, we achieve anywhere from 5-8 9s of reliability or more.
It's statistically unlikely that we'd see an issue like this on the first day of a limited deployment if FSD was hitting those numbers.
That's true. As part of this deployment they haven't driven enough miles to prove safety, but from their FSD data overall they should have a pretty good understanding of where they stand.
> from their FSD data overall they should have a pretty good understanding of where they stand
They know where they stand when there is a safety driver behind the wheel. I'd expect if that data were really good, they'd be less secretive about it. But still, it says very little about where they stand without that driver.
Seems like a mistake not to be a sensor maximalist. Why not try and get as many data streams onboard as possible. Seems like the generalized robot OS would be better if it was a data fusion platform.
That’s a good point. I coming from the perspective that waymo is more of a “dead reckoning” approach compared to Tesla long(and maybe never ending) road.
Their decision to rely on vision got them billions of miles of training data from every corner of the USA. It did not get them point clouds, but monocular depth estimation works extremely well these days, so would it really have been that valuable? It looks to me like they made a shrewd bet and won big.
You might not have heard of the biggest one. Apollo Go in China is actually slightly ahead of Waymo in terms of vehicles and miles.
Assuming the Austin robotaxi service continues to go smoothly, I expect Tesla will leapfrog both of them in the next 12 months. Tesla's cost-effective approach (cameras) combined with the fact that they have already scaled vehicle production positions them really well.
Media (and therefore people who trust it too much) will point out places where each of the services goes wrong, but the reality is that likely all of them are already much safer than human drivers if you define "safety" in terms of severe accidents per million miles.
More than a million people die each year from automobile accidents with humans behind the wheel.
> the reality is that likely all of them are already much safer than human drivers if you define "safety" in terms of severe accidents per million miles.
There's no chance this includes Tesla with their disengagements (equivalent to the driver passing out) and even then it's only true under the restricted set of conditions these systems actually operate at compared to human drivers.
Yes but what I was getting at is if the autonomous system is allowed to disengage every time it encounters a "difficult" situation (as Tesla "FSD" does) then its safety record can't be compared to human drivers even in otherwise comparable road conditions.
Human drivers don't get the luxury of disengaging and having a more skilled driver take over when they're struggling. If Tesla FSD drives for 100km before overwhelming glare causes the system to disengage, that'll go on record as "100km driven without accident", but when the human driver is blinded by the same glare and ends up in an accident 5km later, that'll go on record as 1 accident per 5km driven for the human driver.
> they just stop if they ever encounter a situation where they don’t know what to do.
Oh my god that’s terrifying if true. I can think of many situations when driving when slamming on the breaks is the absolute wrong choice. Tesla is pushing this out way before it’s safe enough to operate in public.
Not justifying it, but there is a reason the person behind is almost always responsible in accidents. You are responsible for maintaining a safe distance in case the person in front of you (for whatever reason) stops or acts erratically.
Apollo Go doesn’t seem to be very safe yet, see https://www.reddit.com/r/SelfDrivingCars/comments/1hy72y9/ch..., it seems like they still need safety drivers? Watching the video, I don’t know how much if the ride is just being China (chinese taxi drivers on Chinese roads can make you car sick also), but the tech is obviously not where Waymo is regardless.
The video seems biased. What's with calling the car to a location where they weren't happy to access it from, and then acting as though it was a fault of the service? Pretty sure I also spotted a gap in the hedges where a fire hydrant was, so it really felt like an act.
Besides whether the video is biased or not, are we conflating comfort & safety? When a robotaxi jiggles the wheel and makes the car shake, that's clearly terrible for comfort and perhaps brings to mind an inexperienced driver, but that is a personification. Programmers who have ever dealt with debouncing might recognize that it's likely something that could use smoothing to aid with comfort, but not necessarily unsafe.
In any case, Apollo Go is ahead in number of vehicles, cities, rides, and probably also miles. Not that it really matters. Demand is going to lead supply in autonomy for many years to come. These services are not actually going to be in real competition with each other for quite some time.
The video seems pretty honest about their attempts to even get a car in the deep suburbs of Shenzhen, let alone Shenzhen central. And the ride making them sick seems legit, but it does contrast with the many Chinese social media posts saying the service didn't need safety drivers and was comfortable (although, if you watch those videos, they are obviously in the deep suburbs also, so I guess they agree on service area limitations at least).
I was in China (Beijing) a couple of months ago yet didn't have a chance to ride a driverless car there because they were only running them in the deep suburbs (Changping near Lenovo I think?). This is quite different from my experience taking Waymo, where it served the central city just fine (although I'm not sure how Waymo would handle a narrow Hutong alley way like one of the taxi's I took did).
The video seems honest in what way? What part of it makes you think that? The only evidence of ride discomfort that I saw was some steering wheel shuddering.
Once I saw the part where the car was being blamed for the pickup location chosen by the user, I really struggle to believe anything that's being said. Felt like they wanted to say it's not good right from the start.
They constantly mentioned that their phone camera stabilization was pretty good, and that they really wanted to like this from the start (otherwise they could have just taken a normal taxi without waiting around in an outer suburb for a few hours to finally get one). I don't know what to tell you, you seem to believe that self driving cars are running around the core of these cities without safety drivers, while all I've seen so far, even from the most enthusiastic Chinese v-bloggers, are cars running in the deep suburbs and aren't really very deployed yet like Waymo is in downtown SF.
This started with me pointing out that the scale of Apollo Go is greater than Waymo, but then you made it about safety, which is quite a different (but still interesting) topic. Then there's this video that has been brought up in the context of safety, but seems more related to ride comfort. Ride comfort is also interesting, but different than safety.
I'm not sure we can actually have a reasonable conversation without objectively separating such topics.
I haven't looked at what the service area is, but I wouldn't assume that a downtown service area is somehow better or worse than a suburban service area, especially not without defining "better". A suburban-focused service seems like it might actually be a lot more useful than a downtown-focused service.
Back to the original point, you easily see the data yourself. The scale of Apollo Go is slightly larger than Waymo on most metrics that would seem to matter; vehicles, rides, miles, etc.
> Waymo seems to be the furthest ahead in my complete outsider opinion.
There's a high floor to Waymo's costs though, and therefore a lower ceiling to their profits. Tesla's aim was always to make it (way) more scalable. If they can match Waymo's performance at Tesla's costs, Waymo goes the way of the dodo pretty quick.
I really feel we are headed towards a future where humans driving cars will be illegal outside of few exceptions. There will be no traffic as all the robocars will coordinate unselfishly. Car ownership won't be a thing, rides will be a service. Parking spots won't be needed. It's just getting over a safety threshold we are comfortable with.
I have predicted this for years. I think they will make city cores self-driving cars only, with no other cars, or even bicycles because they are too chaotic. China should have done this with those ghost cities, so that they could have dominated with full self driving before other countries.
Someone will own the platform that coordinate car movement, and then all cars will need to pay to get onto this coordination platform that will tell each car how to drive, which route to take, etc. Each car know what all the other cars in its area will be doing, so that mass coordination is possible. This is how you can get a completely full highway but all traveling at 65 mph 1 feet away from each other.
Imagine all the land being used for parking that wouldn't be needed anymore. Car accidents will be like plane accidents where a full investigation will be launched to improve the car and coordination software.
I surely hope not. I'd rather have a city centre without cars than one filled with constantly self-driving cars. Pedestrians and animals are still "chaotic elements". At 105 km/h (still slow by European standards), what happens when a deer suddenly runs onto the motorway with a distance of ~30 cm between the vehicles? Talk about carnage.
I suppose the solution is to seal in all roads then with high walls; what dystopian future.
Ghost cities are really just ghost districts, and they simply aren’t economically viable which is why they are deserted, traffic won’t make something economically viable, and it’s quite the opposite, economic viability brings traffic. Self driving isn’t going to make Kangbashi or Tianjin’s new financial district popular.
China is demoing a lot of self driving cars in their suburbs however, so not core Beijing, but out in Daxing or Changping, for example. China could and will mandate self driving cars in dense city cores when they become viable to optimize traffic flow in cities that can’t really fit many more new roads. And Chinese companies are working hard to make mass produced Lidar economical, while America will probably just put high tariffs on that.
The shoddy past 10-15 years of self driving improvements has led me to believe that this "inevitable future" is at least 20-30 years away, if not longer given that a huge portion of existing cars would need to be effectively outlawed for that that to happen, people and businesses will basically always have legitimate reasons for wanting a human in the loop (and not remotely), and the tech has been worked on actively for decades and still doesn't work well enough to avoid loss of life (which, similar to a terrorist attack, might not be as devastating in numbers as the current state of affairs with humans driving - but will absolutely cause a deeper psychological impact on the general public as there is something seemingly more cruel and dystopian about a company killing people via "error" or cost-savings than a person killing someone by accident).
The psychological distinction is definitely something that was perhaps overlooked. we need to put the blame on some one, for closer. Loosing to some rounding error definitely is haunting
Yeah, will probably be the same as with horses. Quiet a few of them standing around at the outskirts of the my town. Maybe we'll get an equivalent to horse girls
Car ownership definitely isn’t going away. Sharing a vehicle with the general public gets old fast if you have the means to avoid it.
There are also many reasons people use cars for more than going from point A to point B on a purely transactional basis. Many professions need to leave things in the car or truck like tools or even your laptop. Having to take everything you own into every building in case the self-driving car gets called back home for service or whatever isn’t going to work.
That's probably a thought experiment in business models. Like a franchise. Waymo is experimenting with Uber as it's retailer, essentially. Personally, I'd prefer to do business with Waymo, directly.
If the cost of the Waymo Driver hardware falls to the point where it's not prohibitive for the low duty cycle of a private vehicle, I could see that eventually happening.
I think you're partially right... it'll start with insurance companies, they'll charge people that insist on driving it themselves more for the right (due to safer standards (not that I'm saying this is the case, just that THEY will say it is)).
Then local Gov will charge more to allow non robot cars on the road ( less wear and tear on the roads, fewer roads needed due to more reliably predictable drivers and fewer accidents).
Then lastly manufactures will get to a point that they need to simplify their production range and will pretty much only produce self driving.
lastly, culturally the demand will change. the ipad generation dont want to have learn to drive or have to spend their screen time driving. the damand from them alone will push for self driving cars.
It's already a reality in many places around the world. Most of people don't own a car, just participate in the ride sharing economy. In many places they still have a safety driver like Tesla but there are plenty of places with %100 autonomous driving.
Also authorities are getting giddy when a human tries to drive on railways, so it's effectively illegal to drive in certain places where the ride sharing is the default mode of transportation. It's also very privacy focused, even though there are cameras everywhere you can just buy an anonymous travel pass that you top-up every once in a while. It also allows you to hop between rides for free or at discount.
In the larger cities they often use hyperloop, so you never get stuck in the traffic.
We're so many years from that point that it's impossible to make reliable predictions. I was hearing these claims over a decade ago and we're only marginally closer. The vast majority of miles driven still happen in locations and/or conditions that self driving isn't reliable at.
There is a massive long tail of unusual cases that happen outside of major cities that nobody is even really trying to solve at the moment because there isn't much of a profit motive compared to dense cities.
In US, I find it doubtful regardless of how well the tech works, for purely political reasons.
On the right, car ownership is deeply culturally ingrained as a signifier of personal freedom, and the notion of the government requiring that a robot drive you instead would be unpalatable, never mind the notion of no car ownership at all.
On the left, you have a generally hostile attitude towards cars in general as undesirable and to be replaced with public transport. Then there's a growing current of FSD being associated with "big tech" in general and Musk in particular that makes it an even stronger knee-jerk reaction.
Difference between this and Cruise or Uber is that Elon is in charge of regulation. A driverless Tesla could plow through a farmer's market and service wouldn't even stop for the afternoon.
That's not usually the argument. It's that LiDar has historically been working more reliably than tech that doesn't use it - and reliability is pretty important in that industry.
In any case I'm curious what gives you such blind faith that Tesla's vision-only tech will work out. There's a lot of really practical and frankly obvious reasons why it hasn't been working for a very long time and similarly lot's of videos (which are aggressively taken down by Tesla or can't even be posted due to shady customer NDA's) of their self-driving equipment getting dangerously close to murdering people. And of course, there are the _actual_ cases of Tesla's self-driving tech just you know, straight up murdering people.
You can shift blame to the humans not being attentive enough to take control when self-driving mistakes come up, but then you've lost the broader argument that self-driving Tesla's are "ready", even if you're right about the former.
Edit because fuck-you-HN-comment-limit-soft-ban:
(response to child comment below):
> Obviously, only vision and hearing are required for driving
Human vision is drastically better than even the most advanced computer vision systems we have or are likely to have any time soon - especially when there are realtime and hardware constraints. That's before you introduce the relatively (!) shoddy approximations that even SOTA vision models learn via SGD.
> I'm curious what gives you such blind faith that Tesla's vision-only tech will work out.
Obviously, only vision and hearing are required for driving. The question isn't "can it work," the question is "which approach will be economically viable on what time scale, and is there a winner-takes-all dynamic that excludes the approach that comes second."
I've never understood the argument "humans only use vision and hearing so self driving cars should too". We don't constrain what we do with machines to what humans do.
Humans don't have wheels either. Is Tesla planning to replace wheels in their cars with legs, Flintstones style?
That's not the argument and never was. Humans aren't proof that you can't do better than vision, we are proof that the problem can be solved with vision.
This isn't a question of max-sensors vs min-sensors, though. It's a question of max-sensors vs max-data under economics constraints that forbade "why not both?". Tesla chose max-data, and they now have a 40x data advantage thanks to their choice, and probably more once you consider the improved diversity of data collected from private cars in every corner of the country and world.
If you were training a self-driving car, would you rather have 40x the data or would you rather have point clouds that are only a smidgen better than what you can get from modern monocular depth reconstruction on the vision data?
> Humans aren't proof that you can't do better than vision, we are proof that the problem can be solved with vision.
It's proof that the problem can be solved with vision when your driver is a human. That doesn't necessarily prove it for non-human drivers implemented on currently known AI architectures.
Humans can learn from a lot less data then current non-human AIs require. We also learn in the field, often recognizing when we've made a mistake and consciously updating ourselves to handle similar situations better in the future. It is possible that this is necessary to make vision only driving work well.
I didn’t say nobody should try to solve the problem using lidar, I just said humans are an existence proof that it can be done. Lidar gives you more data, but maybe that data will turn out to be overkill. We’ll find out.
Calling this 'Robotaxi' rides seems somewhat disingenuous given that they are Model Ys with a full seating provision. People will be conflating this with the (IMO a very bad choice) 2-seat design seen on the demo months ago with 2 doors and no provision for more than two people (which wouldn't work too well if the drives are being chaperoned, reducing seating to one).
This conflation and mixing of words is something which Tesla seems to do intentionally (Full Self Driving, Autonomy, etc), or am I being cynical about that?
I agree that 'cybercab' and 'robotaxi' are similar names, but what ulterior motive would it serve to intentionally confuse people about the difference? Seems like it's just bad naming. Cybercab isn't late or anything, it was announced for 2026 at its unveiling. And Robotaxi is an accurate and descriptive name for a self-driving Model Y that you can hail.
What good is the safety driver doing in the passenger seat? Even the die hard Tesla FSD fans will admit that it isn't uncommon to have a serious disengagement. I give this less than a month before they pull the plug.
In the first instance they're asking to see the passenger's app to verify identity. Seems like a necessary precaution considering what happened to the Waymo vehicles recently.
Presumably if there were an incident they would be trying to remedy any situation.
135 comments
[ 72.3 ms ] story [ 504 ms ] thread> 10-or-so cars
> human driver behind the wheel (except in this beta)
> invitation based (apparently a very limited audience)
> geofenced not only to a city, but to a small handful of neighborhoods
> early reports suggest disengagement requires remote re-engagement
I hope they get there, more competition in this space is good. But, this is pathetic. They're so far behind Waymo it isn't even funny.
Hopefully the Tesla app forces Waymo to reconsider their deal with Uber altogether and just run it themselves like they do in San Francisco.
Are they sycophants or are the other sites haters? Take a gander at every tech website's headlines regarding Tesla for the last few years and tell me they aren't following the "if it bleeds, it leads" mantra.
The people invited to preview Robotaxi were not only unedited, they were also allowed to live stream.
Are you looking at the feeds from the cars to reach 32? Each screen is from the cameras around a single car, so you're actually only seeing 5 cars from that.
If they are still doing this in 3 months, it'd be a bad sign, of course. Their plan is for rapid growth next year.
We'll see if they are able to do that.
- Me, if I wanted to be equally ham-handed in throwing the comparison for Tesla.
In reality, we are seeing two bets on two different approaches. Do you scale up supervised driving to maximize data collection/diversity and then go unsupervised? Or do you go unsupervised and then scale up problem solving as you go? The cool thing is that both of these approaches are being tried so we will find out. If you want to place a bet of your own, you can find the casino in your favorite brokerage app!
In any case, the videos of what is possible with supervised FSD are quite amazing, certainly not "pathetic," and what remains to be seen is if they can successfully navigate the supervised->unsupervised jump, which is certainly not trivial.
Arc de Triomphe: https://youtu.be/o2xKpbKZLVA?t=7
Busy China: https://youtu.be/ybBpRN4Hqbc?t=13
Manhattan: https://www.youtube.com/watch?v=qafr3RrJRfU
https://youtu.be/_s-h0YXtF0c?t=0h7m15s
What's the thesis here, humans are imperfect drivers and therefore we should accept self driving cars driving into incoming traffic, and if anyone objects, focus on their faults instead?
However, I'm not saying that their licence should be taken away (we've all made mistakes - I should have made that clear), but if that is representative, it is dangerous. I rode motorbikes for many years, and you can spot people who are dangerous very quickly from their 'car body language'. And an error like that in another circumstance could kill someone.
I knew people whose driving had deteriorated like this (my late mother and her friends spring to mind). They refused to accept they were not capable any more, and she used to do things that terrified me. I had discussions (kindly!) about her driving standards and errors and she refused to accept that she was making the errors she was. And they were not as bad as what I saw in the video.
Still seems nuts to me, they're pretending Waymo doesn't already regularly drive without supervision.
https://en.wikipedia.org/wiki/List_of_predictions_for_autono...
It seems like it would be a really bad idea to put this out there if they know from the data that it will cause lots of accidents.
It's also possible they quietly rely on remote operators to a large extent. The fake robot PR stunt wasn't even that long ago, the company might try the same thing here.
I just don't see how this train of thought makes sense.
https://www.youtube.com/watch?v=1-5s4JlBesc
Nobody's gonna be the one to tell Elon that the stuff isn't ready.
The phenomenon in that video is only possible when management is heavily disconnected from ground truths & data, but that seems the opposite to how Musk's companies are run.
Tesla: a company that doesn't take risks unnecessarily.
Whilst a wheel wobble & veering somewhere it's not supposed to go looks bad, it's very difficult to do worse than the average human driver in terms of safety.
It's statistically unlikely that we'd see an issue like this on the first day of a limited deployment if FSD was hitting those numbers.
They know where they stand when there is a safety driver behind the wheel. I'd expect if that data were really good, they'd be less secretive about it. But still, it says very little about where they stand without that driver.
How is it relevant though?
People have some sort of beef with Elon and thus FSD, but Waymo has made WAY WAY worse mistakes
https://news.ycombinator.com/item?id=44302724
Waymo seems to be the furthest ahead in my complete outsider opinion. I feel like everyone else is a way distant second.
- Elon, probably
Assuming the Austin robotaxi service continues to go smoothly, I expect Tesla will leapfrog both of them in the next 12 months. Tesla's cost-effective approach (cameras) combined with the fact that they have already scaled vehicle production positions them really well.
Media (and therefore people who trust it too much) will point out places where each of the services goes wrong, but the reality is that likely all of them are already much safer than human drivers if you define "safety" in terms of severe accidents per million miles.
More than a million people die each year from automobile accidents with humans behind the wheel.
Will sound engineering prevail over brain rot in the C-suite? I am skeptical.
There's no chance this includes Tesla with their disengagements (equivalent to the driver passing out) and even then it's only true under the restricted set of conditions these systems actually operate at compared to human drivers.
I'm guessing that they just stop if they ever encounter a situation where they don't know what to do.
Human drivers don't get the luxury of disengaging and having a more skilled driver take over when they're struggling. If Tesla FSD drives for 100km before overwhelming glare causes the system to disengage, that'll go on record as "100km driven without accident", but when the human driver is blinded by the same glare and ends up in an accident 5km later, that'll go on record as 1 accident per 5km driven for the human driver.
Oh my god that’s terrifying if true. I can think of many situations when driving when slamming on the breaks is the absolute wrong choice. Tesla is pushing this out way before it’s safe enough to operate in public.
1. Keep driving in some way
2. Stop
Which one do you think Waymo (or any other system) does when it doesn't understand a situation?
Besides whether the video is biased or not, are we conflating comfort & safety? When a robotaxi jiggles the wheel and makes the car shake, that's clearly terrible for comfort and perhaps brings to mind an inexperienced driver, but that is a personification. Programmers who have ever dealt with debouncing might recognize that it's likely something that could use smoothing to aid with comfort, but not necessarily unsafe.
In any case, Apollo Go is ahead in number of vehicles, cities, rides, and probably also miles. Not that it really matters. Demand is going to lead supply in autonomy for many years to come. These services are not actually going to be in real competition with each other for quite some time.
I was in China (Beijing) a couple of months ago yet didn't have a chance to ride a driverless car there because they were only running them in the deep suburbs (Changping near Lenovo I think?). This is quite different from my experience taking Waymo, where it served the central city just fine (although I'm not sure how Waymo would handle a narrow Hutong alley way like one of the taxi's I took did).
Once I saw the part where the car was being blamed for the pickup location chosen by the user, I really struggle to believe anything that's being said. Felt like they wanted to say it's not good right from the start.
This started with me pointing out that the scale of Apollo Go is greater than Waymo, but then you made it about safety, which is quite a different (but still interesting) topic. Then there's this video that has been brought up in the context of safety, but seems more related to ride comfort. Ride comfort is also interesting, but different than safety.
I'm not sure we can actually have a reasonable conversation without objectively separating such topics.
I haven't looked at what the service area is, but I wouldn't assume that a downtown service area is somehow better or worse than a suburban service area, especially not without defining "better". A suburban-focused service seems like it might actually be a lot more useful than a downtown-focused service.
Back to the original point, you easily see the data yourself. The scale of Apollo Go is slightly larger than Waymo on most metrics that would seem to matter; vehicles, rides, miles, etc.
There's a high floor to Waymo's costs though, and therefore a lower ceiling to their profits. Tesla's aim was always to make it (way) more scalable. If they can match Waymo's performance at Tesla's costs, Waymo goes the way of the dodo pretty quick.
Someone will own the platform that coordinate car movement, and then all cars will need to pay to get onto this coordination platform that will tell each car how to drive, which route to take, etc. Each car know what all the other cars in its area will be doing, so that mass coordination is possible. This is how you can get a completely full highway but all traveling at 65 mph 1 feet away from each other.
I suppose the solution is to seal in all roads then with high walls; what dystopian future.
China is demoing a lot of self driving cars in their suburbs however, so not core Beijing, but out in Daxing or Changping, for example. China could and will mandate self driving cars in dense city cores when they become viable to optimize traffic flow in cities that can’t really fit many more new roads. And Chinese companies are working hard to make mass produced Lidar economical, while America will probably just put high tariffs on that.
Car ownership definitely isn’t going away. Sharing a vehicle with the general public gets old fast if you have the means to avoid it.
There are also many reasons people use cars for more than going from point A to point B on a purely transactional basis. Many professions need to leave things in the car or truck like tools or even your laptop. Having to take everything you own into every building in case the self-driving car gets called back home for service or whatever isn’t going to work.
If the cost of the Waymo Driver hardware falls to the point where it's not prohibitive for the low duty cycle of a private vehicle, I could see that eventually happening.
Then local Gov will charge more to allow non robot cars on the road ( less wear and tear on the roads, fewer roads needed due to more reliably predictable drivers and fewer accidents).
Then lastly manufactures will get to a point that they need to simplify their production range and will pretty much only produce self driving.
lastly, culturally the demand will change. the ipad generation dont want to have learn to drive or have to spend their screen time driving. the damand from them alone will push for self driving cars.
Also authorities are getting giddy when a human tries to drive on railways, so it's effectively illegal to drive in certain places where the ride sharing is the default mode of transportation. It's also very privacy focused, even though there are cameras everywhere you can just buy an anonymous travel pass that you top-up every once in a while. It also allows you to hop between rides for free or at discount.
In the larger cities they often use hyperloop, so you never get stuck in the traffic.
There is a massive long tail of unusual cases that happen outside of major cities that nobody is even really trying to solve at the moment because there isn't much of a profit motive compared to dense cities.
On the right, car ownership is deeply culturally ingrained as a signifier of personal freedom, and the notion of the government requiring that a robot drive you instead would be unpalatable, never mind the notion of no car ownership at all.
On the left, you have a generally hostile attitude towards cars in general as undesirable and to be replaced with public transport. Then there's a growing current of FSD being associated with "big tech" in general and Musk in particular that makes it an even stronger knee-jerk reaction.
In any case I'm curious what gives you such blind faith that Tesla's vision-only tech will work out. There's a lot of really practical and frankly obvious reasons why it hasn't been working for a very long time and similarly lot's of videos (which are aggressively taken down by Tesla or can't even be posted due to shady customer NDA's) of their self-driving equipment getting dangerously close to murdering people. And of course, there are the _actual_ cases of Tesla's self-driving tech just you know, straight up murdering people.
You can shift blame to the humans not being attentive enough to take control when self-driving mistakes come up, but then you've lost the broader argument that self-driving Tesla's are "ready", even if you're right about the former.
Edit because fuck-you-HN-comment-limit-soft-ban:
(response to child comment below):
> Obviously, only vision and hearing are required for driving
Human vision is drastically better than even the most advanced computer vision systems we have or are likely to have any time soon - especially when there are realtime and hardware constraints. That's before you introduce the relatively (!) shoddy approximations that even SOTA vision models learn via SGD.
Obviously, only vision and hearing are required for driving. The question isn't "can it work," the question is "which approach will be economically viable on what time scale, and is there a winner-takes-all dynamic that excludes the approach that comes second."
Humans don't have wheels either. Is Tesla planning to replace wheels in their cars with legs, Flintstones style?
This isn't a question of max-sensors vs min-sensors, though. It's a question of max-sensors vs max-data under economics constraints that forbade "why not both?". Tesla chose max-data, and they now have a 40x data advantage thanks to their choice, and probably more once you consider the improved diversity of data collected from private cars in every corner of the country and world.
If you were training a self-driving car, would you rather have 40x the data or would you rather have point clouds that are only a smidgen better than what you can get from modern monocular depth reconstruction on the vision data?
It's proof that the problem can be solved with vision when your driver is a human. That doesn't necessarily prove it for non-human drivers implemented on currently known AI architectures.
Humans can learn from a lot less data then current non-human AIs require. We also learn in the field, often recognizing when we've made a mistake and consciously updating ourselves to handle similar situations better in the future. It is possible that this is necessary to make vision only driving work well.
I think you dodged the question because you know that 40x data is the clear winner.
I responded to the part of your comment that interested me, which concerned what we can infer from the ability of humans to do vision based driving.
This conflation and mixing of words is something which Tesla seems to do intentionally (Full Self Driving, Autonomy, etc), or am I being cynical about that?
Presumably if there were an incident they would be trying to remedy any situation.
The same as a driving instructor. He has a brake pedal, and he can pick up the steering wheel if necessary.