Hacker News Reader

Comment by dreamcompiler | Hacker News Reader

Should have happened sooner. I love my Model Y. It's the most fun car I've ever owned. It's also the most comfortable on long road trips and I can sleep in it in climate-controlled comfort without generating a gas that could kill me. I leave the house every morning with a full tank and never have to stop at a gas station. And my cost per mile is 1/4 what it was with an ICE car.

I just wish Tesla would focus on this stuff without hyping an AI feature that can never happen.

It's like what if somebody invented the world's greatest backpack? It's bigger on the inside than the outside, and it incorporates negative gravity so anything you put in it becomes 50 lbs lighter. And all that was true; it actually worked? But for an extra $10k, the company will add on the "anti-grizzly bear" feature which makes you invulnerable to grizzly bears while you're wearing the backpack. Except most of the people who buy the anti-grizzly bear feature end up getting eaten by grizzly bears. Whoops! Just quietly stop advertising the anti-grizzly bear feature and you'll still have a great product.

View on HN

I also love my Model Y. It’s by far the best car I’ve ever owned or driven. But the software has stagnated in favor of “FSD” beta releases.

Autopilot on freeways is great. I use it all the time and it works. I wish Tesla would focus on getting that to a level 3-type system - something where I can take my hands off the wheel and just be ready to take over after 10-15 seconds. It feels close enough that it could be done. But despite the complaints about the name it’s exactly what is claimed.

Auto park is bad. It doesn’t detect the spot half the time and it’s not great at actually parking. Especially on tight spaces where it would actually be useful.

Navigate on autopilot is underwhelming. It’s often too timid to actually switch lanes if there’s traffic. I like the signal-to-lane-change.

Full self driving is a boondoggle and absolutely false advertising at this point. Put all the disclaimers on it you want, but I paid for something that (the language at the time) claimed would get me from point A to point B while I relax and don’t pay attention. Every year it’s another few months away. I doubt it’s actually possible with the hardware on the car at this point. Elon’s claim of “humans drive with eyes so a car can drive with cameras” is naive. It’s glossing over the very important bits of “with the compute power in the car” and “within the limits of moser CV/ML techniques”. I have no doubt that autonomous driving is possible - it’s actively happening in San Francisco roads with Cruise/Waymo. But I doubt my Model Y can do what I paid for, and that’s unethical.

This doesn’t need a law, it should be enforced by the FTC or by a lawsuit. It seems like someone scoring some cheap points.

> But the software has stagnated [..]

I've only owned ICE vehicles (lifetime total: 3, 2 of which are parked outside right now), and have only ever rented ICE vehicles (lifetime total: several hundred, the latest of which is in the long-stay at London Heathrow right now...)

I'm unable to get my head around talk of a vehicle's software "stagnating"? AFAIK none of the vehicles I've ever driven had OTA updates, and from the driver's PoV they weren't the worse for it.

If you're hoping to convince current ICE drivers to upgrade to EV, then talking about shipping rapid software updates probably isn't going to work...

I think the implication is that the current software is just not very good.

It's obviously fine for techie early adopters, who like the new shiny and are willing to put up with a lot. But it sounds like from the perspective of a normal person's sustained daily use, the gee-whiz feelings wear off and leave you with annoying issues.

But it could also partly be a case of differing expectations. The web has trained us all that software generally gets better over time. I never expect a car's too-low doorframes to make room for my head. But on an ICE car when I'm confronted with, say, the same clunky dashboard nav system, it seems more frustrating that, year after year, it's still just as bad. In fact, I ended up paying good money to replace the car stereo with something that was compatible with Android Auto mainly so I could have an ever-improving Google Maps in the same spot on my dashboard.

> on an ICE car when I'm confronted with, say, the same clunky dashboard nav system, it seems more frustrating that, year after year, it's still just as bad

I drove my first ICE vehicle for almost 10 years. The driver experience was exactly the same in year 9 as it was in year 1. The replacement is currently in year 7. It's the same as it was in its year 1.

Q: Why would that be frustrating?

It's just not as much fun when you're not renting your UX so next year could bring you better collision avoidance or require you to pay for an upgrade to continue to be able to recline your seats

> better collision avoidance

Mark I eyeballs still in use here, touch wood they're apparently still working ok.

Was driving round greater Innsbruck in the snow 10 days ago, someone was coming downhill on a snowy sidestreet too fast and out of the corner of my eye I clocked they weren't going to be able to stop, yanked my wheel left, I slid(!) out of my lane into the oncoming lane (happily empty) and the unlucky driver slid out over the white line right into the space my vehicle would have been in.

Talk about adrenaline rush :(

Did you have a touch screen? Or a screen that controls any part of the car?

The frustration is that those tend to have awful interfaces, with extra frustration that they could be fixed but don't get fixed.

If you had normal buttons then there wasn't anything to get annoyed with in this way in the first place, so it didn't need updates.

Given that this is a fairly common sentiment, it’s a bit wild to me that SpaceX uses touch screens to control almost everything in their Dragon capsule.

It's a different environment. Astronauts are passengers, not drivers, and they rarely need to make sudden actions while multi-tasking the way car drivers do. If they made all-glass cockpits for fighter jets I would agree with you, but they don't.

This isn’t a very accurate take, anymore than an airline pilot is a passenger. There’s an reason why the first astronauts to fly in it were military pilots. Dragon has a manual mode that requires an astronaut to control the vehicle for safety reasons. I’m not aware of any human-rated NASA software that doesn’t also require a human in the loop to meet the required safety thresholds.

Modern cars have much more features and tech than older cars. UX on a decently designed 30 year old car couldn't be improved much. While my current car is nice, I'd appreciate automatic wipers and Bluetooth that were more reliable. The voice recognition and navigation wasn't cutting edge the day it left the factory. There are a few behaviors of the automatic climate control that need improving too.

The distinction I'm drawing is that we expect software to improve.

My first car had no software, so I didn't expect anything to get better. But the most recent thing I drove frequently had significant amounts of software, especially the touchscreen that did audio and navigation. Even when it first came out it was not very good software, so it was the kind of thing I'd want to get better. And 5 years later, having watched all the rest of the software in my life improve, the flaws were ever more grating.

And I should add that I suspect there's a technology stage issue here. If you look at a model T dashboard, the controls were by modern standards terrible. All of us are used to cars where they spent decades refining physical controls, all of which were pretty uncomplicated compared to what we routinely expect software to do. So it's reasonable to me that if we count the number and severity of driver experience issues, the graph bottoms out sometime in the mid-1980s (1986 being the first time they put touch-screens in cars).

> It's obviously fine for techie early adopters, who like the new shiny and are willing to put up with a lot

"Buttons beat touchscreens in cars, and now there’s data to prove it": https://arstechnica.com/cars/2022/08/yes-touchscreens-really...

I did actually chortle when I read that :)

That isn’t ICE vs EV that’s new vs old

A few months ago, Tesla rolled out a new feature where it will give a gentle ping if you’re standing in front of a stop light and the light turns green.

I love it.

That used to be part of the expensive FSD package, and it was the only part I liked. Glad it's standard now.

> Elon’s claim of “humans drive with eyes so a car can drive with cameras” is naive.

Humans drive with our eyes and we’re pretty bad at driving. We should want systems with redundancy and using multiple modalities so that a car crash isn’t an accident, it’s a choice.

> we’re pretty bad at driving

Are we? I know it is a fashionable thing to say, but I am personally fascinated by the relatively low number of accidents given how omnipresent cars are.

For germany it's ~5 dead / 100000 motorized vehicles, and ~4 dead / 100000 inhabitants. [1]

More anecdotal: Growing up I lived near a pretty busy street and never personally witnessed an accident there.

Makes me think that we are actually pretty good at driving, even on average.

[1]: https://en.wikipedia.org/wiki/List_of_countries_by_traffic-r...

It's a bit of both. We're pretty bad at minimising basic errors driving (at least until you consider we as a species spent a billion or so years of evolution travelling <20mph and generally picked up handling a vehicle in a few hours in our late teens) but so strikingly good at handling the sort of edge cases where people are at particularly at risk of death I'm not sure collision avoidance AI will catch up despite all its speed advantages.

When measured by deaths per passenger mile, passenger vehicles are, by a large factor, the most deadly mode of transportation[1].

1: https://injuryfacts.nsc.org/home-and-community/safety-topics...

7 deaths per billion passenger-miles. To put that into perspective, if walking had a death rate on par with passenger vehicles, a small village of 150 people going on a walking trip for 60 years, at 3 mph for 16 hours per day, should have seen one death in that time.

I don't think we really quite appreciate how incredibly safe the modern world is.

Germany has WAY more difficult driving licensing standards than the US.

My high school English teacher father used to ironically remark about how students who failed his class repeatedly (often for all four years) somehow always managed to pass the drivers license test. Funny that.

Your calculation also neglects the cost of auto accidents in terms of wasted life. There will be an accident on I-5 between San Diego and Los Angeles every single day--generally multiple. When that happens, an enormous number of people wind up sitting in their cars for an extended period of time. That accident may not have actively killed someone, but it caused many thousands of people to waste an hour+ of their life. If a thousand people waste an hour every day, that's one "dead" year of human life every 9 days. And that multiplies across the entire city and then across every city. The cost of automotive fender benders and the time they waste exceeds the "death toll" of driving by quite a lot.

Humans suck at driving. It's a boring, repetitive task with sudden moments of sheer terror. As long as everything goes "normally", it's okay. When something unusual happens when driving, most outcomes are random. The lack of deaths and injuries due to driving is more due to automotive, traffic and safety engineering than it is due to active crisis decision making by human drivers.

> Humans drive with our eyes and we’re pretty bad at driving.

I disagree that humans are bad at driving.

Let me ask you this, how many miles would a fully autonomous self-driving car need to drive without a fatality for you to be able to reasonably claim that it's statistically a better driver than a human?

And feel free just to round to an order of magnitude here. 2.75 x 10^z <-- what is `z` here?

In 2013 in the US there were 32,719 crash related fatalities and 2.3 million reported injuries. So that must mean humans are bad a driving right? Well those numbers are tiny compared to the 3 trillion miles that humans drove that year.

Humans have about 1.09 fatalities per 100 million miles driven. The probability of a human causing a fatality for a mile driven is low: 0.00000109%

To show to a 95% confidence interval that a self-driving car is better at driving than a human it would need to drive 275 million miles flawlessly. This would require a fleet of 100 vehicles driving continuously for almost 13 years.

Not all humans are the same either

Some people can flawlessly do stunts with their car after driving for a few years and have a perfect driving record

Others make you shake in fear every time you sit with them as you see them fail to notice other cars when they change lanes, despite driving for 10 years

It’s a bell curve and I think the median is better than current car driving AI, even with LIDAR, can achieve

But combo AI (or just safety features) with humans and I think it shifts the bell curve up a lot

I think part of the problem is the lax licensing standards in the US. For everyone's benefit, we should have higher standards. There also can and should be regional differences. Someone who has learned to drive in a rural or suburban area has not necessarily had the training and experience necessary to safely drive in a city alongside pedestrians and cyclists.

I think we’re not really willing to up the standards because you need to drive to survive in a lot of places.

You can’t change your genetics, you can’t necessarily change where you live, and you can’t just fix transit without rebuilding an entire city that took 100+ years to build so the standards are just what they have to be

Over the last many decades, we’ve starting requiring things like backup cameras because cars naturally break down and so it’s something we can actually change on a short timescale (but it does add a lot of cost)

Fixing transit is a long timescale thing

A lot of cities were rebuilt in the 1950s to accommodate cars and car oriented design, destroying neighborhoods in the process. Before that they had public transport and better walkability. The real solution is to undo that.

Yes but how would you give a downtown Boston experience driving test in North Dakota?

I still remember the time my mom almost drove into a tram tunnel! She wasn't used to the city.

It wouldn’t have to do that much driving flawlessly - many crashes occur without there being a fatality.

It would actually need to be driven significantly more, see this report: https://www.rand.org/pubs/research_reports/RR1478.html

To quote:

> Autonomous vehicles would have to be driven hundreds of millions of miles and sometimes hundreds of billions of miles to demonstrate their reliability in terms of fatalities and injuries.

> Under even aggressive testing assumptions, existing fleets would take tens and sometimes hundreds of years to drive these miles — an impossible proposition if the aim is to demonstrate their performance prior to releasing them on the roads for consumer use.

There are lots of leading indicators that a car (or human) is bad at driving, so you don’t actually have to wait for an accident to know if you are on the right path.

E.g. keeping track of near misses / interventions can get you data points much more quickly.

To add to one of your points, The Tesla auto park is legitimately awful. https://youtu.be/nsb2XBAIWyA

Every time we try it for a spot to reverse into, the linear regression it must be doing is so apparent. It violently tries to turn left and right rapidly trying to straighten itself.

It works better for parallel spots but often gives up half way because it made a bad angle choice.

We’ve tried it in various models in addition to our own Model Y. They’re just embarrassingly bad at it.

This is kind of scary, it seems they’re trying to use black box Ai/ml for parking where a simple PID controller would do excellent given an input of a target box and current location.

Indeed. Why do Tesla go for ML to poorly solve problems long solved by well understood PID controllers?

The cynical in me feels this is to do with employees at Tesla want to work with new and shiny ML instead of old boring PID tech for the sake of "resume driven engineering".

I suspect these decisions are driven from the top down. We know who loves a shiny tech choice at Tesla. The company has been quite proud of their end to end ML system in presentations. But it sounds like they’re struggling with execution.

My brother in law works as an engineer for Tesla Palo Alto: according to him these features are all driven from the very top. Always at the expense of fixing bugs, or getting even basic things working reliably in ways that would make driving safer and more enjoyable. Make your own conclusions about buying one. I would steer clear, literally, knowing some of the bugs in there.

This tracks with everything we’re seeing at Twitter now. A passion for high level feature ideas and an apparent disinterest in confounding issues or technical complications.

As a robotics engineer for a long time I thought they were ahead of the curve. But I am beginning to see that their big bet isn’t paying off, and I’ve taken more of an interest in going back and listening to those people who said it wasn’t going to work.

Where is the evidence it doesn't use any PID? The wobbling described in the parent is a very PID like behavior (poorly tuned)

I guess everyone is giving Tesla engineers the benefit of the doubt that if they are using a PID they would be competent enough perform a simple proven task for a PID that other car companies have no problem with. I refuse to believe the Tesla engineers are incompetent. So they must be forced into using a "full black box ML" paradigm. There may be a completely different issue that none of us are aware of (faulty sensor, poor mix of both, etc). It is blatant that Tesla vehicles are poor at parking compared to other electric vehicle offerings from BMW, Audi, and Ford.

I think parallel parking works wonderfully (only know autopilot 1)

Remember it uses sonar to park, not the lines.

Strangely, I've never tried backing into a spot - I guess because the backup camera is great at seeing the lines and I can back in quickly.

They’re switching to lines.

That's odd. My 2015 Model S auto-parks very well, faster and neater than I can do it myself.

Is that the older Mobileye stack? It had much better autopilot than Teslas solution.

In the video I linked, you can see it out perform Tesla’s implementation.

Yes it is.

That's embarrassing for you, no?

Not at all. I'm sure that with practice i could outperform the car. But why bother? My view of the automation in my car is that together we both drive better than either of us do on our won.

Thanks for sharing the video. It's kinda telling seeing every single other car's auto-parking feature work flawlessly while Tesla's fails miserably

LOL. They did much better job parking Falcon 9

> Elon’s claim of “humans drive with eyes so a car can drive with cameras” is naive.

Specially because we don't drive with eyes only, we use at least our ears too for hearing unusual noises and the balance sense for detecting vibrations, changes on road surface, and to get feedback about orientation changes.

Teslas have those sensors, but don’t currently use some of them (like microphones) as inputs.

Unlike human drivers, a Tesla autopilot can track the rotation of each wheel independently and in real time.

Yet it still doesn’t work.

> a Tesla autopilot can track the rotation of each wheel independently

But most humans can in most cases feel when there is a loss of traction on the road. Although agreed it comes nothing close to what the car knows.

That said I am not sure that this is a core feature needed to make FSD work.

No idea what resolution those cameras are, can they be as good as our auto-adapting eyes together with our brains? They are pretty neat features after all, trained right. And will we loose our abilities if we don't keep working on them, building our feature set?

The resolution is pretty good, but they got one other massive advantage over human eyes even without sound - a full 360 degree view at any given time that doesn't require turning head, thanks to the array of cameras covering everything all at once.

Our eyes also have a much greater dynamic range than most cameras. Those are bad in low light and often get blinded easily.

With the amount of theorising and handwringing people do wrt Elon’s motives or reasons for having this view of FSD, there seems to be a lot of people that don’t want to acknowledge him for the boring stereotype that he is: a manager / salesperson completely out of his depth making wild unsubstantiated claims. It’s just that the autocratic role he demands within organisations, and his relatively newfound publicity fuelled by his legion of fetishists, means that we all hear about it.

I only vaguely (for a developer) know about AI / computer vision, yet I’m still made deeply anxious by the implication that we will somehow in the foreseeable future be able to pull this off at ALL, let alone with 1.5 hands tied behind our backs as Tesla insists.

> I also love my Model Y. It’s by far the best car I’ve ever owned or driven. But the software has stagnated in favor of “FSD” beta releases.

I'm not sure that I agree with this, tbh. More of the regular AP stack is moving to the FSD underpinnings, and I'd expect the nav on autopilot feature to be a part of that eventually. All of those are much needed fixes, IMO. I bet the recent big improvements to the auto-high-beam functionality is related too.

I do wish they'd focus on something more like a FSD-lite for the highway, though. Let me go hands off for a little longer as long as the driver facing camera is detecting attention.

> I just wish Tesla would focus on this stuff without hyping an AI feature that can never happen.

There's a good argument that Tesla wouldn't exist today without the hype. The enormous positive public image mad Tesla's cost of capital extremely low, and ditto its cost of marketing. Tesla needed a lot of both to get to break-even. Musk himself has talked repeatedly about coming close to bankruptcy. It's reasonable to believe that less hype would have meant less capital and fewer sales, nudging them at some point over into bankruptcy.

There’s a tipping point where the hype becomes excessive or even fraudulent.

The ends justifying the means doesn’t absolve a business of needing to conduct themselves ethically.

Oh, yeah, I think Tesla's well past the point of fraud. Musk promised "1 million robotaxis by end of the year", and that year was 2020. I don't think the end justifies the means at all here.

Elon Musk wants to be Steve Jobs so badly but the reality distortion field isn't working.

"You can fool all of the people some of the time and some of the people all of the time, but you can't fool all of the people all of the time."

And buying Twitter really put him in front of all of the people all of the time.

Fake it to you make it. FSD will never come on the current hardware.

Or fake it until you crash...

Remembering that they also added video games, crazy light effects, ... Tesla is surfing from one hype wave to the next one

It's almost as if someone invented an amazing new technology that really works, then some VC came along, bought it out, (er, I mean invested!) ran the original inventors out of town and bolted on some unasked-for snake oil. Now, you can still buy the still-awesome product (mostly made in America too) but you have to take all the VC crap along with it. Welcome to America!

This is typical engineer fairy tale thinking:

that “inventing” v0.01 defines all other versions and nothing else is needed even if v0.01 didn’t get sold. And all other versions have ton of work in them.

And seasoned with the big bad wolf, oops investor, buying everything at gun point.

Edit: of course this is downvoted lol, hate Musk all you want ( I do but because he is pro CCP) but reality can’t be denied

The idea of FSD is, obviously from the effect of hyping it, enormously attractive. It can't happen now. But in the 1990s, speech recognition that would enable me to dictate this comment in a relatively noisy space was science fiction. Now that we have orders of magnitude more computing power, and much better microphones and associated noise rejection technology, you can blithely dictate messages with reliability that was previously a dream.

Thing is to go from being unable to distinguish pencil from cancel to today's almost magical speech recognition, took 25 years. That might be the same interval it takes to go from mowing down motorcyclists to delivering all the safety and productivity potential of level 5 autonomy.

Someone commented on the Waymo thread that insiders believe they already have FSD, but the cost doesn’t work.

Two things about that: if by FSD you mean level 5 autonomy, waymo is not level 5. It is a robotaxi. Fully autonomous in that role. But the definition of level five includes driving in situations where you don't have any prior data. Outside the map.

Fair but I’d imagine 90%+ of taxi/uber rides are inside cities.

Really hoping that they bring the cost down. I’m a way it seems like Musk was half right - the cost for these just doesn’t work (according to the comment).

> It's also the most comfortable on long road trips and I can sleep in it in climate-controlled comfort without generating a gas that could kill me.

Why are you sleeping in your car for long trips? :)

I know you're kidding but it's a good question. I don't do it while the car is moving obviously. I like to camp in the car at RV spots at lakes, state parks, and other pleasant places while on a long road trip. Plug in the car, sleep in the back, and wake to a full charge.

Because hotels can be expensive, and when you want to bomb from one side of the country to the other driving 16 hours a day, it's just easier to sleep in the car. I've done it many times in a 20-year-old Jetta, very comfortably.

> Because hotels can be expensive

Perhaps buying a $60k+ car isn't the best idea if that means you can't afford a hotel room occasionally.

But my car cost less than $2000.

The context of the discussion is someone sleeping in their Model Y, not someone penny pinching to get by.

Just because someone spent a few tens of thousand dollars on an electric car doesn't mean they should be willing to drop $200 a night on hotels for a week straight.

My uncle bought a Tesla 3 (coincidentally, to replace his 20-year-old totalled Jetta) because he needs to drive several hours to the office a few times a month for 2-3 days at a time, and the silent climate control and flat trunk floor means he can cheaply and comfortably spend the night in his car rather than driving 3hr back home or renting a hotel room. It is also much cheaper to operate.

> Just because someone spent a few tens of thousand dollars on an electric car doesn't mean they should be willing to drop $200 a night on hotels for a week straight.

I disagree. In my opinion if you can't afford to do that a few times per year without financial concern then you have no business spending $60k+ on a car. Having to sleep in your expensive car just so you can afford a vacation is ridiculous. There are much cheaper alternatives that won't force you to sacrifice other aspects of your life, such as your $2000 car.

I'm not even going to comment on your uncle's situation. Let's just say I would be working somewhere else if I had to resort to that.

It's not financial concern. My uncle could definitely afford to rent a hotel room, but he doesn't mind sleeping in his car and he would rather spend that money on something else. Heck, I could afford a hotel too, but I would rather sleep in my $2000 car.

I don't think it's about the price. Just the money saved on fuel may be enough for a hotel.

I think it's more about Tesla owners becoming obsessed with their car, it's a strange group of people.

After driving a few hours, they only thing a normal person would want to do is get out of the seat, walk a bit, take a shower.

It's like when you were a kid and got your first walkman, you would take it anywhere with you, while eating, sleeping, even in the shower if possible.

Some people assess value independently of their means.

As per my uncle's situation, he works for the US army, is due to retire in a year or two with a full pension, and moved to be closer to family and friends. He mostly works remotely. It being the army, sometimes he has to go to the nearby base 20 minutes away, but other times he has to go to the far away base for some reason or another.

Or maybe, the person enjoys their car a lot more than they value a night at a hotel. Perhaps the extra value at a hotel versus sleeping in a Tesla is not worth the extra $200 per night. Someone wanting to save a few thousand dollars on a hotel is NOT the same thing as "financial concern". It's just a preference that is lot different than yours.

Many people with money spend huge amounts on some things but can be penny-pinchers in other areas.

Not sure why you are getting downvoted. I'm guessing the downvoters have never camped? TBH, quite a few people do this in Teslas, sometimes in remote areas.

> I just wish Tesla would focus on this stuff without hyping an AI feature that can never happen.

Why can that never happen?

Arbitrary limitations on the sensor suite because "we figured it out, humans only need cameras" are one reason.

Tesla is not running in the same race as actual AV companies.

All of the "actual AV companies" have only released curated marketing videos and/or restrict operation to curated geofenced areas. Meanwhile we've seen exactly one company's autonomy technology exposed to unsupervised and uncontrolled adversarial conditions across hundreds of wildly diverse cities in the USA. Whether you are impressed with their state of progress or not, there's nobody you can compare them to.

Operators like Waymo do look plausible while operating within their tiny geofence. Driving pre-validated roads is impressive in its own way. Right now if you dropped a Waymo and Tesla FSD on an unsealed road in Michigan, one of them will drive at least as well as a human learner driver and the other will probably refuse to move.

The point is, I agree that there could be many firms which are far ahead of Tesla, but there isn't enough information in the public domain to say this with any confidence. Right now Tesla's is the only major system which can be assessed with critical objectively.

> ...one of them will drive at least as well as a human learner driver and the other will probably refuse to move.

If it was just one single trip, I was incapacitated and just needed to get home, then fine, level of human learner driver is good enough. But for everyday usage? No way, the risk is just too high.

> one of them will drive at least as well as a human learner driver and the other will probably refuse to move.

refusing to move is a feature not a bug. and human learners don't slam into a jersey barrier because there's some glare.

why tesla bros continue to insist that they have a good vision stack while they can't even do contrast adaptation is beyond me. the dril tweet comes to mind

Instances of "slamming into a jersey barrier" are on highways and by definition involve the legacy Autopilot stack and not the FSD beta vision stack. I agree that the tesla bros are extremely cringe, but I'm equally unimpressed with criticisms that rely upon ignorance of basic facts about the thing being criticised. Tesla is deserving of a lot of criticism for a lot of things, but they are clear about the distinction between Autopilot and FSD. Eliding this distinction in one's criticism gives the appearance of being uninformed.

Also your claim about contrast adaptation seems awfully specific, enough that I'm interested in a citation. There's plenty of evidence that their imaging sensors do adapt well in high dynamic range circumstances.

humans don't crash into shaded jersey barriers, as a tesla famously did. contrast adaptation of CCDs pale in comparison to that of the retina. there a ton of literature on contrast adaptation in the retina and in neuroscience in general, i would suggest you look there

> humans don't crash into shaded jersey barriers

They literally do, which is why jersey barriers are literally designed with vehicular impact in mind.

> contrast adaptation of CCDs

You can't just mix terminology like that. Contrast adaptation describes behaviour in biological systems which isn't analogous to anything a CCD does. There's a reason why literally nobody uses the term to describe digital systems, because it's not relevant.

If you were trying to say that the CCD doesn't have sufficient dynamic range, then just say that instead of using a term which doesn't describe dynamic range.

> Right now if you dropped a Waymo and Tesla FSD on an unsealed road in Michigan, one of them will drive at least as well as a human learner driver and the other will probably refuse to move

One will drive and likely get into an accident, and the other will responsibly not do what it wasn't designed for. But if you took away Waymo's responsible disposition, I'm sure it would drive better than the Tesla, because it's just way further along in development, more and better sensors, more rigorous testing and simulations, more corner cases that absolutely had to be dealt with to roll out a real self-driving product.

> One will drive and likely get into an accident

What are you basing this prediction on? There is extensive independent documented video evidence of FSD beta driving and navigating on unsealed roads in Michigan with no human intervention required. This includes driving at night, in the rain, and with a diverse array of obstacles.

> [Waymo is] just way further along in development

This may well be true. However there's no publicly available information to make any objective assessment. For better or worse (and I see valid arguments both ways) Tesla is airing their dirty laundry in public for all to see. It can be assessed by objective third parties in privately owned vehicles without the blessing or oversight of the parent company.

By comparison we have no way of assessing Waymo's stack objectively. For example we don't know to what extent corner cases are embedded into the software or handled by humans remotely. Was the 15th Avenue incident resolved by human intervention or an improvement to the stack? To what extent is their vision and planning stacks capable of universality or over-fitted to regional specifics? No objective analysis is possible, so no comparison is possible.

As someone who works at one of the "actual AV companies", there are just as many examples of FSD making mistakes that show Tesla's approach is broken in unacceptable ways, so assuming it will get in an accident is very fair.

What laypeople don't seem to get is that AV mistakes don't follow the same "scale of alarm" as human mistakes.

Tesla is making mistakes that, for a human, might not be world ending, but for an AV make absolutely no sense.

Things like recognizing it needs to make a turn and making no attempt at all to slow down: https://www.youtube.com/watch?v=6wGAyAN3iRo

Or recognizing a vehicle is in its path and not even trying to slow down: https://youtu.be/6wGAyAN3iRo?t=21

Or recognizing a stop sign and ignoring it to follow the car in front: https://youtu.be/21ShKWLAnWA?t=548

Classification issues in AVs are supposed to be the achilles heel, but Tesla is still running into situations where perception works correctly and the AV just completely ignores ground rules that any serious AV must be built around.

> Was the 15th Avenue incident resolved by human intervention or an improvement to the stack? To what extent is their vision and planning stacks capable of universality or over-fitted to regional specifics? No objective analysis is possible, so no comparison is possible.

It doesn't matter when the product you're comparing it to ignores stop signs. Tesla's FSD approach is the definition of the local maxima problem for AVs. It makes progress on the axis that impresses lay people by sacrificing progress on the axes that matter to the long term success of an AV. Geofencing and arbitrary sensor limitations are just the tip of that iceberg...

> All of the "actual AV companies" .. restrict operation to curated geofenced areas

That's because they aren't being reckless and releasing a dangerous product onto the road that they know has issues. Unlike Tesla.

And those geofenced areas are at least with Cruise almost entire cities.

No lidar or sonar, not enough sweating of the thousands of dangerous edge cases, and an insistence on a software architecture whose failure modes cannot be adequately characterized.

I'll disagree with the others and just say "not enough cameras". Cameras alone can work, but you really need stereoscopic cameras in every direction without huge blindspots less than 3 feet from the car.

Until they fix that, there will be lots of limitations. It could get to be good enough to be hands off a lot more than it is now, though.

I suspect Moravec's Paradox strikes again.

I'd like to see what happens if someone did the following to test the assertion that "cameras are all we need". Cover all the windows of a Model S with LED screens, and route all the cameras to display onto them. Then see how well human drivers do with that visual input which roughly simulates the fraction of a second delay to perform the FSD processing, and removes our stereoscopic perception when we turn / move our heads around.

Many software and hardware engineers do not appreciate the amount of work done by evolution upon our kinesthetic / proprioceptive packages. Sure there is wide variability in its implementation (that friend you won't under any circumstances let behind the wheel with you in the vehicle). But by and large, it performs under a set of energy and materials constraints that continues to school our current engineering state of the art.

So I sometimes also wonder if the ultrasonic/sound sensors in Teslas are sufficient to reproduce the control system our bodies' proprioception engages when we drive.

Tesla's evaluation is closely related to its FSD promise. Not to mention it already sold future "fsd" options. If FSD is out then you have just a dumb car. How do you justify your evaluation(i.e worth the entire auto industry EV or ICE)?

Maybe you don't? Maybe you learn to live with an accurate valuation based on the company's record of innovation in batteries, motors, and castings? Not to mention heat pumps, aerodynamics, nearly flawless worldwide charging infrastructure, etc?

Is it really better to have a hyperinflated valuation based on vaporware that risks falling off a cliff when people get tired of the lies?

It's because he knows that those things are relative commodities, and in 5 or 10 years the rest of the industry will eventually catch up in every single way. Hyundai and ford are pretty much %80 of the way there already.

Could be, but Apple could serve as a good example of staying ahead even as others are catching up.

> n 5 or 10 years the rest of the industry will eventually catch up in every single way.

Seems to have already happened? Mercedes EQS/EQB, BMW i4/i7/iX, Polestar, Ioniq

All seem better than the comparable Tesla vehicles.

Though Tesla does have better features for sitting at chargers, and obviously the super charger network which is a huge thing for anyone taking longer trips.

I did a test drive of a bunch of EVs, and Hyundai makes the Ioniq which I was referencing.

I found that the ford mustang EV and Ioniq had about a %20 worse driving experience than a several year old tesla model 3 LR, and tesla did have a bunch of things that were better. But that also came from being years ahead in making EVs and having a more refined product and driving experience as a result. Driving all the v1 EVs made me realize that they were close, and in a few years they'll probably catch up to tesla on the driving experience. Will they be able to make nice software like tesla will? I'm not so sure.

Mercedes and BMW sell their EVs even more expensive than tesla in many ways too, so their price category is different.

Also cars after a certain point are way more commodity, they wont be able to differentiate like apple does with their products nearly as much.

And that supercharger network advantage will go way in 5 to 10 years too.

   > Should have happened sooner.

Which tells you everything you need to know about California's government.

What, that they are #1 in understanding the self-driving technological landscape? They remain more responsive than the other 49 states.

Their timing tells you what you need to know. This isn't new information, their decision is motivated by something other than a new set of facts on self driving vehicles.

So they were wrong, and they stayed wrong (for money) long after they knew they were wrong. And now that it's convenient they've reversed course. Sounds like the other 49 states was right to be cautious.

This is illogical. You can’t simultaneously criticize California for taking so long to do this, while praising other states for being more cautious by not doing it.

You’re also ascribing some laterite motive that you fail to provide a description of.

Reply on news.ycombinator.com