Hacker News Reader

Comment by lolinder | Hacker News Reader

What it comes down to for me is the marketing. The other manufacturers are very careful in how they market the software, Tesla is not.

If you look at Mercedes, for example, their marketing page describing their driver assistance technology[0] (with a very similar feature set to Tesla's) uses the word "assist" more than 30 times and in practically every header. Few people would come away from that marketing thinking that their car is going to drive itself without them paying attention.

Tesla, in contrast, advertises their "autopilot" and "full self driving" capabilities. The word "assist" is used exactly once on the Autopilot landing page[1]. The rest of the words and names are carefully chosen to convey a sense of total autonomy.

[0] https://www.mercedes-benz.com/en/innovation/autonomous/the-n...

[1] https://www.tesla.com/autopilot

View on HN

Tesla has also used this wording to advertise their purported self-driving features since 2016[1]:

> The person in the driver’s seat is only there for legal reasons. He is not doing anything. The car is driving itself.

[1] https://www.tesla.com/videos/autopilot-self-driving-hardware...

I find it really interesting that for Mercedes L3 self driving they are even willing to put their money where their mouth is and take liability for the car in self driving mode.

Smart. This will be the move that triggers widespread adoption. Cheaper insurance.

It makes sense as long as it is failproof. But is their tech better than Tesla's? I assume it is not and they will have to severely limit the capabilities of the system. Maybe it will work in traffic jams, simple routes.

Why are you assuming it isn't better, just Musk's relentless promotion and exaggerated claims?

Read the fine print though.

They do disable the system in complex situations.

And they give ten seconds warning in advance of this happening.

Which sounds great until you think it through…

It means they schedule the disabling of the system way earlier in the progression of any potential traffic complication. Ten full seconds is long before incidents even start unfolding. The car has to be psychic, but it's not. So in order to accomplish the ten-second warning, what that means is those warnings need to be on a hair trigger with many false positives.

The driver, if they have it enabled, will be constantly getting warnings that the system may disable itself in ten seconds when the car sees even the very slightest possibility that things may potentially get complicated way down the road. I don't see this being endurable for most people.

So my conclusion is this is just marketing hype, until they can get rid of the auto disabling thing.

I haven’t read the fine print but you can have ten seconds of warning without predicting complications ten seconds in advance.

That 10 seconds simply means that the car must handle any situation safely for at least that amount of time. In practice, this means decisions such as emergency braking and steering must be autonomous but possible more complex scenarios (e.g. moving for emergency vehicle on a tight street) can be delegated to a human (or the system can simply pull over and stop).

They disable the system in cases of rain but at least from the marketing and videos demonstrating it, there’s no prediction of complex scenarios like you described.

No, you are not understanding what I explained. There is a moving horizon of time during which the ten seconds is continually extended, and the system needs to draw the line somewhere. When does the final ten seconds before cutoff begin? Your examples don't explain this. This is not easy stuff to solve without some compromises.

If it’s limited on what roads it can run on the 10 second window makes perfect sense. Just enable when you’re entering an unsupported roadway.

> And they give ten seconds warning in advance of this happening.

which is infinitely better than the <1 second that tesla gives you.

> those warnings need to be on a hair trigger with many false positives.

unless it is actually reasonably good at tracking threats and traffic. look, if you have multiple sensors of different types, then you are able to take firm action when things turn south quicker than a human.

A lot of tesla's problems are because they have shit sensors, and a stupid approach to designing the system (no lidar, no radar, no multiview cameras, no high res GPS or maps) They also have an overwhelming pressure to just yolo shit, rather than test, and redesign.

>which is infinitely better than the <1 second that tesla gives you.

Not at all. With Tesla you are constantly paying attention, or should be to the same level that you are with any other car.

Meantime it is actively intervening to keep things safer with better follow distance, collision avoidance, lane departure detection, and automatic braking. Most scenarios leading toward accidents are entirely avoided due to all this.

So the <1 second scenarios (which are not talking about warnings, but about when the system was in a complete outlier situation it does not know how to deal with and disabled itself) are very unusual things.

Like "a Prius driver drove off a bridge and is now landing on our hood." Of course the system will disable itself in that situation; what else would you expect? What sensors do you suggest for that?

> With Tesla you are constantly paying attention, or should be to the same level that you are with any other car.

you should be paying attention, but you are not. Human attention is a difficult thing. from what I recall, in a driving it takes about 7-14 seconds to re-gain situational awareness. This means that 1 second isn't enough.

> Like "a Prius driver drove off a bridge and is now landing on our hood." Of course the system will disable itself in that situation;

I'd expect that the system would slam on the brakes, not disengage to avoid liability. Thats the point here, its not about tech, its about legality. Thats the worse part, the entire system appears to be designed to stop tesla being taken to court.

That’s a paranoid and particularly uncharitable view which ignores the perfectly valid reasons that it’s best the system behaves as it does.

The human has responsibility for the safe operation of the vehicle, whether they step up and fulfill it or not. If they don’t fulfill their responsibility, all bets are off.

There is no other (equally good or better) way this could work in practice. You can imagine other ways, and I’m guessing you will, but they are imaginary, not practical.

> I'd expect that the system would slam on the brakes

It can and it does, even while they autopilot system is disabled. Who said it wouldn’t?

You should learn more about the cars before hardening your opinions so much.

> paranoid

I doubt its paranoid, I am not worried about it, just rather annoyed that driver's aids are being marketed as something they are not, cheapening an entire industry.

> There is no other (equally good or better) way this could work in practice.

We are literally discussing a company that has another way.

> It can and it does,

https://www.youtube.com/watch?v=45XMhMzMDZY

suggests otherwise.

>You should learn more about the cars before hardening your opinions so much.

I work in machine perception, this is my bread and butter. more over I have worked with life critical infra, and I know corner cutting when I see it.

I think the clever thing about this is that it breeds a culture of responsibility at Mercedes - if you are programming the self driving, you will be more considerate of how it works because your company (and perhaps ultimately you) bear that responsibility. The consumer gets confidence in the product, and Mercedes holds itself accountable because it doesn’t want to ship a product that will endanger people.

The page literally talks about "full self-driving capabilities [...] through software updates designed to improve functionality over time". Which to me sounds like FSD is not actually being advertised as currently existing, as opposed to something you'll eventually get.

Yes the actual finished FSD software release does not exist and Tesla does not claim it exists. There is a SKU you can pay up front for (in other words, pre-pay for) called FSD, so the SKU exists, but it does not yet exist as a released, finished software release.

A lot of people can't get their head around this level of subtlety. Those people probably should not have too much confidence in their take on this, but they do. Dunning-Kruger effect. It seems like you do get it.

Just for completeness to soothe anyone triggered by details missing here, certain features of the current beta version are enabled for drivers who have purchased the SKU. And there is also a full beta release that is available to some drivers who opt in. All of this does not mean FSD exists yet as a public non-beta software release.

> A lot of people can't get their head around this level of subtlety.

Have you considered that this might be intentional on Tesla's part?

In my original post I never said anything about full-self-driving being a real thing, I simply said that their marketing uses that term a lot and leans heavily on those capabilities. I'm well aware of the distinction, but I'm also aware that Tesla seems to intentionally cultivate the ambiguity as to what exactly FSD means. It's not fair of you to cast blame on those who get confused when the company makes the line very, very blurry in their marketing.

EDIT: I'm also not accusing Tesla of outright lying. They can "not claim it exists" while still making sure plenty of people miss that "subtlety".

Yes, I've considered the possibility of it being intentional on Tesla's part. It's impossible not to consider it with all the conspiracy theorists and short sellers repeatedly bringing it up on Hacker News and elsewhere.

However, I don't buy into the theory, for the reason that Tesla's proactive, frequent, and aggressive informational notices clarifying the point are impossible to miss when you are in the car, and that goes completely counter to the theory.

And their informational messages are also there prior to that during the buying experience.

And prior to that in the marketing.

Why would Tesla do all these reminders about the human needing to maintain oversight and control, if they are trying to trick you into thinking the opposite?

Haters latch on to the marketing as if it's the only thing, and ignore the caveats in the marketing, and ignore the informational messages during the buying process, and ignore the in-car information and the in-car active measures the car takes to make sure you are paying attention.

Haters dismiss all that and pretend it does not exist.

To what end, I'm not sure, I think it is to stay in a comfort zone regarding a delusion they have about their opinion being the right one. Confirmation bias doing its job.

I do appreciate you bringing up the point. It's a fascinating phenomenon.

You're assuming that FSD is a software problem, and that there is any realistic chance that the feature will be available in the lifetime of any of the cars they sold including it.

Both of these beliefs seem unwarranted: FSD as described in the advertising is most likely much more than 5 years away, and will almost certainly require LIDAR to achieve any kind of safety.

People pay for a pre-order based on the promise that the item will be delivered. If I pre-order a game and it is later cancelled, or even delayed for many years, I will get my money back, I won't just be told "well, you knew it wasn't ready at the time".

> and will almost certainly require LIDAR to achieve any kind of safety.

Is there a physical reason for that? We know that humans do just fine with just ~8cm of stereoscopic separation, and for example cars have the potential for a significantly higher amounts of stereoscopic separation.

Not a physical reason, no, but an AI one.

Humans and most other animals don't rely solely on stereoscopic vision to navigate the world, we rely on a model of the world where we recognize objects in the image we perceive, know their real size from experience, and use that as well as stereoscopic hints to approximate distances and speeds. We additionally use our understanding of basic physics to assist - we distinguish between an object and its shadow, we can tell the approximate weight of something by the way it moves in the wind (to know if we need to avoid an obstacle on the road), and there are other hints we take into account.

We also take into account our knowledge of the likely behavior of these objects to judge relative speeds (e.g. thr car is moving away, it's not the tree coming closer).

Without this crucial aspect of object recognition and experience about the world, our vision is actually very bad at navigation. If you put us in an artifical environment with, say, pure geometric shapes at various distances, no/fake shadows, objects with non-realistic proportions and so on, we will have much more trouble navigating and not bumping into things even at walking speeds. And this is the level the AI is currently operating at, more or less.

And if you don't believe me, note that humans with one eye, while having impaired depth perception, are still perfectly able to drive safely, with ~0 physical mechanisms for measuring distance (I beleieve the spherical shape of the iris may still give some very subtle hints about distance as you move your eye around, but that is minimal compared to stereoscopic vision). A LOT of our depth perception is just 2D image + object recognition + knowledge about those objects.

While all of this may be true, this doesn't explain why stereoscopic vision wouldn't work where a LIDAR would. Both provide identical geometrical information and neither has anything to do with AI. Neither tells you approximate weights of things, or judge based on human experience how things might move in the future depending on their type (tree vs car), or anything like that. And if you swap one system providing geometric information for another one that provides identical information, I don't see how this makes the cognition of any AI later in the pipeline magically any better, no matter how good or bad that AI was previously.

However, one benefit that long baseline stereoscopic vision (for example with cameras in corners of the front windscreen) would have compared to a short baseline stereoscopic vision (a human) or a point measurement (LIDAR) that could be relevant for safety would be the ability to somewhat peek around the vehicle in front of you from either side. Admittedly, this may overall be a small-ish benefit relative to a LIDAR but it does provide strictly more information (slightly) than a LIDAR would.

Well, LIDAR uses very well understood physics to give you precise measurements of distance from the world around you, without any need for object recognition. It is not enough on its own, but it is an excellent safety technology. It's basically impossible to run into an object that's moving slow enough to avoid based on LIDAR input.

Stereoscopic vision first relies on object recognition of the elements of the pictures taken by each camera, then identifying the objects that are the same between the pictures, and only THEN do you get to do the simple physical calculation to compute distance. If your object recognition algorithm fails to recognize an object in one of the images; or if the higher-level AI fails to recognize that something is the same object in the two pictures, then the stereoscopy buys you nothing and you end up running into a bicycle rider crossing the street unsafely.

LIDAR does have limitations of its own (for example, it can't work in snowy conditions, since it will detect the snow flakes; not sure if the same applies to rain), but the regimes under which it is guaranteed to work are well understood, and the safety promises it can make in those regimes don't rely on ML methods.

> Well, LIDAR uses very well understood physics to give you precise measurements of distance from the world around you, without any need for object recognition. It is not enough on its own, but it is an excellent safety technology. It's basically impossible to run into an object that's moving slow enough to avoid based on LIDAR input.

Again, claiming that LIDARs make things magically safer sounds like a lot of snake oil to me. Both LIDARs and stereoscopic systems use well-understood physics. Stereoscopic rangefinders were being used in both World Wars for gun-laying and you wouldn't say that you don't need precise measurements for gun-laying.

> Stereoscopic vision first relies on object recognition of the elements of the pictures taken by each camera, then identifying the objects that are the same between the pictures, and only THEN do you get to do the simple physical calculation to compute distance. If your object recognition algorithm fails to recognize an object in one of the images; or if the higher-level AI fails to recognize that something is the same object in the two pictures, then the stereoscopy buys you nothing

As for whether stereoscopic vision relies on object recognition, that seems like a mild stretch to me. Generally it, like for example SfM (of which it is a special case), seems to rely on local textures and features for individual data points -- and in a simple single-dimensional stereoscopic vision case, your set of possible solutions is extremely limited, so matching features from SIFT or SURF in stereoscopic vision is way simpler than even the general SfM case. Those individual data points do not require in any way for individual objects to be recognized and separated. I have NOT seen in my life an SfM solution that would not give you a point cloud if it failed to separate objects -- in fact, SfM software doesn't even try to identify objects when generating a point cloud because it doesn't even operate at such a high level. Note that this actually provides the exact same information as a LIDAR would, namely a point cloud with no insight how the points are related to each other.

Pretty much the only situation where stereoscopic vision or SfM fails to provide depth information is with a surface of highly uniform color completely devoid of textures. Whether this could or couldn't be solved with structured light is an interesting problem.

Human stereoscopic vision could also be fooled by specifically designed optical illusions in science museums. We just avoid them when designing roads.

I don't think I said anything about hardware or whether FSD will ever be delivered in the lifetime of the current fleet.

Myself, I am skeptical of Elon's timelines. But I also understand they are not his promises, they are just his (foolish, imho) expectations.

He admitted he vastly underestimated the problem… but what he might not admit is continuing to do so. I think he continues to underestimate it.

On the other hand, the power of compounded returns of improvements over time is counterintuitive, and he probably understands that better than most of us. Maybe he used that understanding to get overconfident, or maybe it's still beyond reach. We really don't know. It's still possible that at some point, his team might just crack it. Not just with vision, though. They will need a world model for things like predicting the behavior of a group of children occluded by a bus near a crosswalk.

The money back thing is another question. I hope Tesla offers money back to ease the experience of those who are bitter, but I probably won't take it back myself, because I don't mind supporting the effort even though it seems like the results are far away. I don't think money back was an option previously. As a matter of company survival (which in Elon's mind equates to humanity's survival, take it or leave it, but suffice it to say he doesn't treat it as a normal throwaway company) Tesla just didn't have the money. Now, they probably do.

> I don't think I said anything about hardware or whether FSD will ever be delivered in the lifetime of the current fleet.

You said "Yes the actual finished FSD software release does not exist and Tesla does not claim it exists." (emphasis mine).

Even if you didn't say it, the whole false advertising investigation revolves around the difference between FSD being a software or a hardware problem. Tesla marketing and Musk personally have stated clearly (at least in the past) that all cars sold with the FSD option are FSD ready on the hardware side, and that FSD will be delivered as an over-the-air software upgrade to all of them once it's ready.

If they can indeed enable (working) FSD without a hardware upgrade, then they have not lied (even if the timelines they suggested were wildly optimistic). If they in fact need hardware upgrades to support FSD on the cars sold with this option, then they have lied in their advertising, and people who bought this are entitled either to a refund or to a free upgrade when the feature is available.

Reply on news.ycombinator.com