Here's a Rorschach test for you:
Stuck on the Streets of San Francisco in a Driverless Car
A reporter and a photographer went for a ride in an experimental autonomous vehicle operated by the General Motors subsidiary Cruise. There were bumps in the road.
You can legitimately read this story two ways:
- Driverless cars are pretty far from being ready for prime time. They're still slow, limited, and can strand you for no good reason.
- Wow! Driverless cars still have glitches, but they do damn well even in a famously crowded city with tough traffic. They're closer than you think.
Read the story and decide for yourself.
If you want an idea about how San Franciscans feel about the fleet of autonomous murderbots, please google "san francisco cruise self-driving". Skip past the normal stuff.
So that's how that's going. The weird flocking behavior, where bunches of them will converge on an intersection and stop, is super annoying to people who need to drive around them. I think that's only happened twice, but it is... odd.
One upside I can point to is cops keep getting confused, trying to give them tickets. So we get America's Funniest Cop Gags in our bleak, horrible future, at least until the cops are replaced with robots, too.
flocking together....hmmm....collusion. They're going to take over!!!
I haven't heard much out of Pittsburgh recently. Searching didn't help...
Tomorrow is TESLA Robot Day!
Take your human to work day?
Even in the year 2061, George Jetson wasn't going to be cleaning toilets, as a white man, under Android Rosie's direction.
Just another robot flash mob. Move along. Nothing to see here.
Our Terminator Future.
I doubt many people come back to these, but, Update: It has happened a lot more than twice.
https://www.sfexaminer.com/news/the-city/s-f-to-feds-cruise-driverless-cars-keep-blocking-our-roads/article_7c881bba-4023-11ed-9732-6baa6e037597.html
"The City's letter to NHTSA provides specific data on these incidents. Between May 29 and Sept. 5 of this year, 28 incidents of stopped Cruise cars blocking traffic were reported to 911. The City identified an additional 20 such incidents reported on social media over that time period, which does not include the events of the past week. The City estimates that these figures represent "a fraction of actual travel lane road failures."
Rather than read about San Francisco in a New York paper, there is always a local source: https://www.sfgate.com/local/article/driverless-cruise-cars-block-SF-traffic-17467985.php
Which may not have paywall issues either...
That link works - no paywall for me.
I can't read it at all as I used all my free articles for Krugman, but both the Cruze and Waymo cars seem to work pretty well under perfect conditions. This surprises nobody. The issue with driver less cars has always been what happens during exceptional conditions.
Dancing bears. It's not how well the bear dances, it is that it can dance at all.
My view will always be that if we had started work on this kind of stuff decades ago, ignoring the complaints and bribes of the auto industry, we would have flying cars by now. We were promised flying cars!
This year is George Jetson's birth year so they may still come faster than you think! He was 40 in the series so the first flying cars must have started flying around Orbit City several years before that.
As an aside, while factchecking this rumor I'd heard last summer on Snopes, I learned that The Jetson's originally just ran a single season (1962 - 1963). Can you believe that? Just 24 episodes repeated over and over since then -- and still so formative during my childhood and countless others. Wild!
The very young me fell in love with Judy Jetson. And also Penny from Lost in Space. The original show.
https://youtu.be/QYpoCzkHh0s
But the Apollo 440 recorded theme for the 1998 film update is an absolute banger!
If George Jetson was 40 the year I was 2 why does he look so much older than I do now after having spent all that time in the future and he still isn't even that age yet?
He even had the world's first treadmill for exercise, which Wikipedia says we're only patented in 1968, and still looks like he smokes three packs a day and needs a liver transplant.
We have had flying cars since the 1940s. We just call them helicopters.
Do you want Katy Tur to show up? Cause this is how you get Katy Tur to show up.
They've been working on these things for years as proponents keep saying they will be ready soon. They could eventually work under realistic conditions, but it will be more years. They certainly have not yet developed an AI "robot" which could actually learn the way humans do. As far as I can see, the human programmers have to code for every possible situation. The process is accumulative, as computers don't forget and don't keep making the same stupid mistakes that humans do, so they keep getting closer.
Yes, computers can learn. It’s called an “artificial neural network.” It’s not human learning, but it’s a far cry from human programmers coding every possible situation.
Now, I can’t speak to San Francisco’s experience, but having been part of Telsla’s FSD beta program for about ten months, I have to say it’s getting pretty close. I’ve used it in Los Angeles, Manhattan and across the country.
Cities definitely give it more trouble than highways or towns, because edge cases happen more often there. But it’s constantly improving. Maybe we’re years away, but I think 2-5, not 30.
Computer 'learning' is just lots of bog-simple linear algebra. To call it 'learning' is really a misleading attempt at anthropomorphism.
I’m the first to admit I don’t understand the math. I was just pointing out it’s not people literally coding in every scenario they can think of.
A neural network with linear activation functions would be incapable of learning.
Yet another innumerate who thinks his opinions deserve equal weight. Watch this series. And please don't come back until you do.
Back when driverless cars first hit the discourse, the discussion had the air of "We have solved 95% of the problem. How hard could the last 5% be? This is just around the corner!" That was about eight years ago. Now the discourse is more that they have solved 98% of the problem. But of course this just means that the really knotty parts are still hanging.
Driverless cars and AI will not save mankind from extinction.
Rather than trying to make machines super-intelligent, it would be much a much more worthy goal to help reduce the ignorance and stupidity of actual living human beings.
One of these things is much difficult than the other.
Genetic engineering isn't that advanced yet.
Sounds to me like it will have to be an "all or nothing" scenario
The ONLY way to benefit is to have all cars be antonymous because one human failure could cause an accident for a self driving car
I like the technology - I also like the pursuit of this. Imagine shipping w/o drivers? Or use it on ships?
Thank goodness I won't live to see the day this all happens.............
Ships already have autopilot systems that navigate in open ocean areas with radar to identify other ships. I think they have or, are in late stage development of automated docking systems as well. Like driving the more complex situations will be the last to be automated. For ships these include canals and docking that requires tugs.
Sort of like driving in Naples. If you follow the rules you create all sorts of confusion and dangerous situations because the locals don't follow any rules and assume you won't either.
"In the back, above the two tablet computers, there’s a button that lets you call tech support for help and a speaker through which that disembodied voice can talk to you."
Uh oh! As the usage expands, the inevitable next step is that when we call tech support for help we will reach someone halfway around the world for whom English is not their first language...nor, apparently, their second or third language, either.
This was obvious from the first time we heard about a remote help desk for when the car couldn't figure it out. These desks were manned by highly paid engineers. What are the chances of this remaining the case?
I think the real challenging frontier for their application are the edge cases where it's challenging to gather data to inform the ML, e.g., driving on icy roads, through snowstorms where the normal stopping distance is radically altered, where the behavior of others on the road is not as predictable, where car sensors are likely to fail. My guess is that these technologies are going to be challenged to make it to operational use in areas with frequent bad weather any time soon.
I am team, "Wow! Driverless cars still have glitches, but they do damn well even in a famously crowded city with tough traffic." The technology deciding that it does not have complete information and deciding it cannot continue is a good thing. For the simple getting to point a to point b use case I expect autonomous vehicles will be significantly better than a human driver within 5 years. If they did not have to predict the irrational behavior of humans they already would be. The hard problem is not identifying obstacles and street signs. It is being able to predict that based upon what someone is wearing, what time of day it is, the neighborhood and other details whether they will drunkenly stumble into the crosswalk.
Based on the sfgate story linked by rick_jones, I can't tell whether "they perform damn well" or not. There's no indication about number of trips or percent of trips that require intervention. It's a lot of people feeding their confirmation bias by noting when the cars have problems.
Personally, I'm in the "They're a long way from driving on unmarked roads in the winter at night" camp.
If by "in the winter" you mean "roads covered with snow, and maybe fresh snow falling" you're describing a situation that many human drivers can't handle. It's very difficult to stay on the road when you can't tell where the edge of the road is.
I moved to snow country when I was about thirty and had to learn to drive in it. What I wonder about with driverless cars is the throw-in-the-towel point. If I am driving through the mountains through a driving snow storm, I might decide to pull off at the next exit with a motel and check in. What will the driverless car do when it concludes that this is a really bad idea? Pull over to the side and shut down? I honestly don't know. I have never seen this discussed, what with the excitement over driving the mean streets of Mountain View.
I was born in snow country and have been driving in adverse winter conditions for over fifty years. There are still times when I just say "not going anywhere today" or, if I'm already out when conditions get bad, I get off the road at first opportunity. Usually, by the time I say "not going" the radio is telling me that the local police are asking everyone to stay off the road.
One would hope that an autonomous vehicle would do exactly as you say, pull over and shut it down, or at least return control to a human, when conditions get really bad.
I have a friend who is well steeped in the industry. He told me self-driving technology works in ‘about 90% of circumstances.’ There are two categories of challenges (he called them use cases). First, are information challenges: situations, such as weather, where the technology can’t see or where markers (lane lines, speed signs) are absent. Second, are context challenges. For example, due to road construction, a car needs to drive on the wrong side of a double yellow line or a policeman is using hand signals.
The challenge, per my friend, the difficult use cases are very difficult to solve.
The cops will have to be wearing some kind of indicators on their gloves the car can understand, which will immediately figure in a movie about hijacking someone which will immediately inspire street gangs or lunatics to try it in real life.
Not only the cops need th3 magic gloves, but also road maintenance workers, tree trimmers, tow truck crews, the guys who help semi drivers back into tricky docks, all of whom direct traffic as part of their work. Now, what about the motorists who arrive on an accident scene before the cops can get there, and try to stop a chain-reaction?
It is worse than that. One thing I quickly learned when I moved to snow country was that the lanes shift in snow. Where before you had two lanes going in one direction, in snow they rapidly consolidate to one lane, which may or may not straddle the line between the two. This new lane is defined not by markers, which you can't see anyway, but by the tire ruts of previous cars and the path of the plough on its most recent pass through, How would a driverless car handle this? I have no idea. This is another topic that is discreetly elided by the enthusiasts.
As for context challenges, an utterly routine part of my commute is the trash truck making its rounds. My drive is mostly on rural roads with one lane each way and a double yellow line down the middle. In theory, if you get behind that truck you are in for stopping and starting indefinitely. In practice, you wait for an opportune moment and zip around it, double yellow line or no. Often the trash guys will helpfully signal when it is safe to do this. Everyone does it and it and a cop would have to be really hurting to make his quota before he would pull you over for it. So how would a driverless car handle this? Again, we aren't told
At least where I live, it's legal to pass stopped vehicles even on double yellow lines.
Resident of northern-tier states for over half a century. One of my winter habits in snowy/icy conditions is waiting until I have clearance around my vehicle, then braking gently to gauge the degree of grip I have. That goes well beyond just sensing and reacting.
Once in Connecticut when I hadn’t had the opportunity to check traction, I was descending a moderate grade toward several cars waiting at a stop sign. As soon as I started braking I realized I would not be able to stop fully in time. I steered toward the edge of the road and scrubbed the front tire against the embankment*, and brought the car to a safe stop.
*Outside of town centers, lots of roads in New England lack curbs. Or markings. Occasionally, pavement.
You just described why this is very unlikely to ever work in most environments. And if they don't work in the vast majority of environments in the vast majority of conditions then they aren't really fully AV and can't be relied upon.
I live in SF and there is just constant road construction and each construction area has a different means of getting around the construction area. There will never be enough training data for every scenario at every intersection.
One of my big beefs with the pursuit of AV technology is that we have spent God knows how many billions of dollars fruitlessly trying to, effectively, increase the use of cars instead of investing in alternatives to get cars off the road for a wide variety of reasons. And I say "fruitlessly" because I think the huge number of use cases will mean that the last 10% to full AV in all circumstances and if they aren't pretty close to 100% they'll just not be practical except in very proscribed areas and conditions.
And just for the record and for those who are new to Kevin and think this is going to happen, Kevin is on the record of predicting that mass produced AVs would be available around 2020 (I don't remember the exact year, but his prediction was definitely before the current year of our Lord 2022). We are no closer to general availability of that than we were when he made those predictions around 2016 or so. This is genuinely not intended to be a dig at him - he was far from alone in that sort of prediction - but just to show that it is far better to be skeptical than not of the widespread availability of true AVs.
I have a Tesla running the latest self driving beta software. People who have it understand it, but for those who don't.
1. The Waymo/Cruise plan is based on accurate GPS and mapping, as well as sensors. So the theory is that in a geofenced (i.e. selected) area, the car not only uses sensors, but in theory knows where it is on a map at all times. These systems are the closest to actually fully self driving, but you don't really have to be a critic to point out that we already have technology where you can get places but don't have to drive, but the technology only travels in a limited area (i) trains, (ii) buses. Plus, my understanding is that Waymo/Cruise does not have a plan to actually sell cars to people its basically going to be robotaxis.
2. Tesla is actually working on a mostly sensor non-geofenced system that can drive anywhere. Its not really comparable to Waymo. And, of course, you can at the moment still buy it. The best way of thinking about the developement is to imagine a map. On highways Tesla's system is very close, of course, you have to merge highways and leave them for roads and stuff. It can do the merging, but not 100% well. But on the highways its very, very close. I drove from LA to Tahoe and the Tesla drove for like 400 miles of it. On major divided roads, well marked, Tesla is, I would say, more than halfway there. On my street, which is unmarked and so narrow if cars are parked and an opposing car comes one car has to pull over, Tesla's system disengages. I have seen how the development works and there is no question Tesla will achieve it, there isn't any part of driving which is not solvable by identifying an object and controlling the car to not hit that object or otherwise react to it. The difficult part is getting the software that can convert camera images into objects and into decisions based on that object. With each iteration of the software more and more objects are correctly identified.
Incidentally, long before the cars actually drive themselves, under any system they will be safer than humans. That's because as the system develops at a minimum there are warnings and at a maximum the car actually takes over from the human to avoid an accident.
Also, per rick jones' sf gate cite above about the cars just stopping, there are going to be two hurdles to driverless cars, not just safety.
Most of the time I disengage the system on my car is not because it is behaving unsafely or about to be unsafe, its that I don't want to be rude.
Not being rude driver is not the same as not being a safe driver. The cars blocking intersections in the sf gate article are safe, but rude. And therefore not ready.
Its quite a hurdle to be not only safer then human drivers, but also as considerate, because being a considerate driver has a whole set of decisions independent from safety.
How much you wanna bet Tesla does not succeed?
Ugh on those comments declaiming flying cars will ever be a thing. The fact that these wights even bother to discuss why rather than point out why not is a whopping big clue stick as to their technical literacy, which is to say, 'below grade level.'
When a rider-monitor can tell the car "Kevin*, you don't have to worry about the squirrels running across the road - they'll get out of the way (or won't crash the car if they do get run over", you will have some real artificial intelligence, that is ability to actually learn in real time. There are reports of supposed AI devices, for example ones that can devise complex images to order. But if the device is doing something wrong, can you tell it how to do it right? Hopefully the programmers are learning when the Tesla crashes into a parked vehicle, but the car itself isn't, even if its brain is not incinerated.
*That is the default name for smart cars. Why should the names of these things be all female anyway?
There is a reason why consumers aren't allowed to rename Alexa,Siri and so on. And it's a damn good one, IMNSHO.
As a Tesla driver who has been using Tesla's FSD Beta software for the past year, I think Kevin's Rorschach test is completely off base. This is not surprising from somebody who knows nothing about self-driving cars. He's properly skeptical of information from reporters when it's stuff that he knows about, but....
It's not a question of how advanced the cars are, it's more a question of why a reporter who knows nothing about vehicular autonomy is writing an article and spreading his ignorance.
I don't have, unfortunately, much insight into when Tesla's FSD will be truly autonomous everywhere. It is, as Yikes mentioned, getting better, but it's a long way from handling city driving and exceptional weather and road conditions properly. The approach is completely different from what Waymo is doing, so it's hard to compare. Tesla is building a software system that is supposed to be a much better driver than a human, under all conditions. It has much better senses and reaction time and precision than a human, but (currently) less contextual understanding of what's going on around it.
So the progress for Tesla towards true autonomy is pretty much the progress towards making an AI generally intelligent about everything to do with driving. I don't think these other companies' approaches have any path to true autonomy. They can get close, in limited areas, by pretty much brute forcing a solution. But, if Tesla solves the general case then Tesla will have a much, much cheaper solution. So it becomes very difficult for anybody else to build a business around it.
Full selfdriving beta for a year!?
...hey hey, Jon Zeigler, how many kids did you kill today?
Oh what the H.
On the off chance anyone reads this far, the first iteration of the FSD software I saw in July of 2019 recognized cars and lanes, basically two objects. Enough for pretty good fwy driving.
Around 2020, I remember the software update where the first object was recognized, and it was an orange traffic cone. Not sure why that was first, but I remember it.
Then came people, motorcycles, trash cans, then stop lights and stopsigns, then reacting to the color of stop lights, then bicyclists. Each of these things requires a different reaction but it hasn't been the reactions, the improvement has been identifying of objects.
So now, last weekend, on a very, very busy LA city street, one lane each direction with a bike lane, the car successfully navigated the street, several stop lights, gave a cyclist a nice six foot pass, which required going over the center line a bit, and then nicely reacted to a car pulling out of a parallel space.
What it could not do was finish a left turn on to a narrow street with no lane markings. Streets with no lane markings are not part of the software yet, basically.
Its ironic because as humans we tended to learn how to drive by speed and amount of traffic. But that's not how the computer is learning. Its learning on an object by object basis from what I see.
And I should give up, but.
Its not, by the way, "AI." So far its all reaction. And a computer with eight cameras, from what I see, can react just as well as a human if not better.
Most of driving is reacting.
Most of being a consciencious driver is anticipating. Nothing in Tesla's software anticipates as far as I can see. So its not really AI in any sense.
But I don't think it has to be.
I really should buy a Tesla. Then by the Grace of Elon I will hold superior insights and while the rest of you will know nothing. Except I hate going two deep in the touch screen menu to open the glove compartment door.
For non-urgent stuff like opening the glove compartment, many people just use voice commands. Tesla mentions them here (https://www.tesla.com/support/voice-commands).
But, of course, various diligent Tesla owners have experimented and compiled extensive lists. See, for example, here: (https://www.notateslaapp.com/tesla-voice-commands), a list which includes "Open the glovebox". Surprising to me is that it also includes "Gas".
But yes, if you haven't driven a Tesla for a significant amount of time, you're pretty much unqualified to comment on the experience. Driving an ICE vehicle around doesn't magically generalize to Teslas.