Just vision is a very reductive way of describing how we drive. We use sound to hear if something else is coming and where from, or if we're losing traction on the road, or if the road surface has changed. We can judge acceleration and deceleration in all directions using our inner ears. We can feel if the car's performing differently to usual indicating an issue with the car or different conditions. In addition when we are using our vision, if it gets obscured (i.e. snow covers the windscreen) we know how to get it off with wipers, and most importantly we're very adaptable to new conditions in a way that computers aren't, if we experience something completely new to us, chances are we'll make a reasonable in the moment decision.

> Lots of people manage to drive in poor visibility without radar, lidar, etc. If that's safe and normal for people, the principle that self driving should work with just cameras isn't such a strange one.

No camera system comes close to the capabilities of human eyes, combined with general intelligence.

I am one of those who drive in bad weather from time to time. I'm 'good at it' - but I cannot honestly call myself safe. I've been in the ditch. I've spun a 360 and only didn't hit someone else in the process because it happened nobody was there.

i grew up where bad weather was common enough that we cosidered it not worth shutting down for bad weather so we risked driving in it - but it was always a risk and many do die from taking that risk.

The vision-only approach isn't even able to trigger the windshield wipers correctly.

they are when you can open a line of credit against them

Sure. But when the option is priced at strike price way below the market price, it can be pretty much considered as salary.

It seems quite likely that once self driving cars are well perfected, we will demand more than just human level driving which is currently horrendously dangerous. If lidar systems can exceed vision only, we are going to demand it as a baseline standard.

> He is not wrong, but we demand superhuman performance from our machines

I have a model 3 with v3 FSD hardware. FSD is an objectively terrible driver compared to the average human.

"He is not wrong, but anyway, here's why he's wrong."

He's wrong. Cameras are not enough... but they're certainly cheap enough.

He is very obviously wrong since Waymo cars drive millions of trips with 0 drivers while every single robotaxi still has a safety driver in it at all times.

Correct.

You have to pass theory exam - when I was doing it it was 30 questions with four possible answers out of ~500 pool [1] now the pool is 3700. If I'm not mistaken you could make two mistakes and still pass the test. The questions are either road diagrams, ie. intersection with three cars, road signs and/or lights and you have to tell in which order cars will go or a picture from drivers POV and a question what you should do in such situation - like three lanes, car on the middle, the right is a bus lane and the question is if the driver is allowed to take the left lane.

Once you get that done you can take practical driving test.

It starts on a training yard where you have a series of tasks, like said driving forth and back on a curve in one sweep motion, starting a car on a incline without going back or losing your engine - mind you we're driving manual - parallel parking between cones [2], etc.

If you ace them instructor will take you for a 30-45 minutes ride around town. Apart from normal driving he will ask you to few random tasks like parking in normal conditions.

Any mistake will end the trial with a fail and you have to start all over again. Three failures and you need to redo your theory exam.

[1] I have really good memory and when I taking my exam I went through all questions few times - they are publicly available - and when I had my exam I only looked at the picture and double checked if the question ended up with question mark or period as some images were reused and I knew my answer. I don't think I would be able to do that again with that 3700 question database lol.

[2] before taking exams you go to a private driving school and my instructor gave me a cheat code for parallel parking - which is extremely tight, but also, as pointed by my instructor, government regulated, so all the cones have to be in very precise spots. Not only that, but you take the exam in government selected car (whoever won current bids). So he told me to back up until I saw a cone lining with door post, then full stop, rotate the steering wheel by a exact amount of degrees, etc.

PS. gun permits are given on similar grounds, plus you are required to have a regulated gun safe at your premise and it has to be permanently attached, so given the fact that many of our apartments are smaller than your garages and most people are renting significantly reduces access to guns :)

I know, I know, murka bad, but my driving test decades ago had a left turn out of a driveway, a 3-pt on a road with real traffic, parallel parking, and an intersection with right on red to return to the testing center, located in a terrible parking lot (i.e. designed in accordance with modern standards) that had a ton of stupid lines that "normal people" cut but you have to follow as part of the test.

I (1997, upstate NY) didn't even do that much. I got in the car with the instructor, we drove around the block once. I ran the stop sign because a truck was parked in front of it. The instructor didn't notice (I only noticed later).

Stop signs became universal. No reason why machine readable signals/devices to communicate don’t become the norm with law enforcement and emergency response workers.

Quite frankly, many drivers don't do well here either since hand signs can be very ambiguous. And many times there are contradictory signals that require interpretation.

Look at Scottie Scheffler's arrest for an extreme case of how very hard this is to get right.

Sounds preferable to a dystopia where AI driven cars are getting into wrecks because they’re overconfident in their abilities.

I see no issue with that as long as you can override and drive manually.

I picture one where it locks the doors and drives you right to the ICE center as soon as the facial recognition cameras realize who you are

even better if this is the only way to get around. no transport for whoever the Trump admin decides is insufficiently loyal!

y'all need to get more creative with your dystopias

It's really interesting because that's something they definitely don't teach you when you first learn to drive. Growing up in Florida, I learned to pull over and turn on emergency blinkers if the rain gets bad enough. The reason I know to do this is because I saw other drivers do this on the highway and realized that's pretty wise. It's tempting to imagine that a younger version of me would have been smart enough to realize this on my own but I think most of us learn a lot by observing the behavior of others. Or maybe I would have learned eventually after a few close calls with skidding. Or maybe I would have never learned until it's too late. I wonder if the different responses to averse conditions you've observed is a function of the different experiences we've had as drivers. You might be a more experienced driver than some of those around you.

The limit is much higher than human performance given enough low latency compute. [1] is probably the limit, the actual issue is being able to do that while also avoiding colliding with other road users. The challenges of state estimation and control should be the same.

[1]https://www.youtube.com/watch?v=2MWib6bnnT8

Humans have one advantage over autonomous cars in ice: they can pull over and put on chains. Cars can’t do that (yet).

(I’d love to see a serious winter vehicle that can deploy traction devices by itself, perhaps while rolling at very low speed. Off the top of my head, it seems like it might be easier to put them on then to take them off.)

I hate this sort of take. Humans aren't horrible, on average, they're about average. Your opinion is just a statement of "my judgement is decently far from the fat part of the bell curve" but dressed up in "and I know better" type snobbery.

Eh, it's a pretty big distinction weather-wise. Extreme Western New York and the Tug Hill plateau are all susceptible to somewhat frequent lake effect snow. Given the right time of year and wind fetch, you can see narrow convective / lake-effect snow bands from the Finger Lakes. But broadly speaking the actual annual expected snow and the phenomenology of the storm systems that produce that snow are very different over the rest of the state.

It's easier in bad weather because the ~10% of stupid people who traffic normally just kinda flows around slow it to a traffic jam that's pretty impossible to screw up too badly.