Re: Guest Post - Checking back in on self-driving cars

1

The last time this came up was no fun... Partial disclosure: I work for one of these companies. Skepticism is warranted but the speculation I see from outside doesn't entirely connect with reality as I know it.

I think there's a gap that causes some confusion: Does the tech work well enough to imagine safely and profitably providing a taxi, trucking or delivery service within a few years? Versus: Does this look like we imagined "true" level 4-5 autonomy? We are very, very far from you buying a personal vehicle for any reasonable price that can take you anywhere you want to go while you nap in the backseat. We're also pretty far, when you think about the scale, from getting a robotaxi to the airport in every one of the world's major cities. But those aren't the short term goals. The goal is turning the pilot services you see out in the world right now turning into real businesses.


Posted by: Opinionated Lincoln Towncar | Link to this comment | 10-10-22 7:42 AM
horizontal rule
2

Selfishly, I am very glad to see that someone with vastly more technical expertise than I have shares all of my prejudices and beliefs about self-driving cars, and has evidence to back them up. Thanks, Sifu!

(Also, the social element of driving is a good example of why I loathe tinted windows with a passion. At least two or three times a week I have to make a more dangerous decision as a pedestrian or a driver because I can't see the driver. And there is no good solution except for people to voluntarily stop doing it, because police enforcement of overly-tinted windows is highly racialized and biased. Argh.)


Posted by: Witt | Link to this comment | 10-10-22 7:42 AM
horizontal rule
3

Does the tech work well enough to imagine safely and profitably providing a taxi, trucking or delivery service within a few years?

The problem from my perspective is that we have ample evidence that "safely" for many of these companies is going to mean "only kills people who lack social power," not "is actually as safe as human-operated equipment."

Which, to be clear -- human operators kill people all the time too. But self-driving vehicles kill them at a higher rate and are vulnerable to design flaws that kill people in newly horrible ways (cf the data on how they are worse at recognizing dark-skinned pedestrians).

I am not at all confident that self-driving vehicles will be held to the same criminal law standards as human operators; I suspect it will instead be more like other occupational safety disasters where capitalism and insufficient regulation combine to create a workplace where "expendable" people die with some frequency, but invisibly enough that policymakers don't feel the need to act.


Posted by: Witt | Link to this comment | 10-10-22 7:52 AM
horizontal rule
4

Cory Doctorow put out an essay that mostly riffed off the Bloomberg piece (which I can get with incognito BTW) but also noted that the promise of self-driving cars helped boost the value of the companies working on them, at least for a while (a period when some early investors cashed out)

1: Maybe, although Levandovski's one-route industrial haulers technically fit that definition.


Posted by: Minivet | Link to this comment | 10-10-22 7:54 AM
horizontal rule
5

3: Yeah, some VC guy in the wake of this suggested self-driving cars would work if only all our roads were designed for them.


Posted by: Minivet | Link to this comment | 10-10-22 7:55 AM
horizontal rule
6

Great Tweety reads, thanks! And worth thinking about - how many times have you been behind another car or truck and said something like "watch out - that guy has no idea what he's doing" from observing the car's body language?

(Also the bit about unprotected lefts suggests to me that the autonomous car Turing test will be figuring out the Pittsburgh left.)


Posted by: Cala | Link to this comment | 10-10-22 7:59 AM
horizontal rule
7

3: moreover, laws are already pretty friendly to homicide by car!


Posted by: Cala | Link to this comment | 10-10-22 8:00 AM
horizontal rule
8

(Also the bit about unprotected lefts suggests to me that the autonomous car Turing test will be figuring out the Pittsburgh left.)

Ha! I had a similar thought.


Posted by: Witt | Link to this comment | 10-10-22 8:03 AM
horizontal rule
9

Which, to be clear -- human operators kill people all the time too. But self-driving vehicles kill them at a higher rate

Citation?

I really can't jump into a debate on this topic for obvious reasons but if anybody has any questions I can answer without getting myself fired I'll do my best.


Posted by: Opinionated Lincoln Towncar | Link to this comment | 10-10-22 8:04 AM
horizontal rule
10

The thing about 5 is that you could say the same thing for ordinary cars, they weren't that useful until we redesigned all our roads for them. It's not totally crazy to redesign a lot of roads for self-driving vehicles, in particular, it seems like you could redesign interstates for them and then still require vehicles to be driven by people in town.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 8:40 AM
horizontal rule
11

When I bought my first car almost 10 years ago, I really thought it was going to be the last non-self driving car I bought. No matter whether there's still some promise left in self-driving cars, they pretty clearly aren't going to pan out the way it looked 10 years ago.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 8:42 AM
horizontal rule
12

There's a lot of really really bad drivers out there. I'd expect that technology developed by reputable companies that care about safety (i.e. not Uber or Tesla) are putting cars on the road that are much much safer than solely human-operated cars. I'd expect human drivers are at least as racist as the automated drivers. The current situation with cars in the US is so so so bad that a lot of things that are still bad are going to be an improvement.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 8:47 AM
horizontal rule
13

Really enjoying Sifu's substack!


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 9:18 AM
horizontal rule
14

12: It's not a matter of whether or not human drivers are consciously racist. It's about whether the software can accurately detect darker-skinned pedestrians. There is good evidence suggesting it can't.

(And we know from other studies on Fitbits and pulse oximeters that companies cheaping out on the type of sensors used mean that technology works more poorly on darker skin, sometimes with life-or-death consequences. I have no reason to trust that self-driving car companies are going to be more ethical than their peer companies in other industries.)


Posted by: Witt | Link to this comment | 10-10-22 9:23 AM
horizontal rule
15

I particularly liked this post of Sifu's: https://apperceptive.substack.com/p/what-if-we-were-actually-trying-to

There are some places in the world that do genuinely care about safety for pedestrians and bicyclists (The Netherlands is an obvious one), and it'll be interesting to see what they look like in 20 or 30 years. Are there Dutch autonomous car companies? What are they trying to do?


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 9:26 AM
horizontal rule
16

. I'd expect that technology developed by reputable companies that care about safety (i.e. not Uber or Tesla) are putting cars on the road that are much much safer than solely human-operated cars.

I agree you might expect that. But what we're seeing on the road suggests it might not be the case because of all the things natural for humans and unnatural (so far) for computers we discounted; in any case they should have to prove it before widespread adoption.

It's not totally crazy to redesign a lot of roads for self-driving vehicles, in particular, it seems like you could redesign interstates for them and then still require vehicles to be driven by people in town.

If it were just the interstates I wouldn't mind, but they'd still need to get to and from the interstates, so I can't imagine the push would stop there. I don't care to rip up / reconfigure our cities even more than we already have for regular cars - or require pedestrians to wear special vests or some shit.


Posted by: Minivet | Link to this comment | 10-10-22 9:27 AM
horizontal rule
17

Oh right, the other thing I meant to say in 15, it's sort of astounding to contemplate the divergence in death rates between Western Europe and most of America that's starting to open up and seems likely to only accelerate. It's not just guns and opioids, cars are going to play a role too, and of course the increasing Republican hostility towards medicine.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 9:35 AM
horizontal rule
18

Ugh, we're going to end up with cars that are remotely operated by underpaid drivers in some warehouse somewhere, where each person is remotely operating many vehicles at a time, aren't we? Like the autonomous cars decide when they don't know what to do and some driver just gets dropped into a video game dealing with whatever the problem is? And then if there's an accident you just fire the driver.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 9:45 AM
horizontal rule
19

15: I have no idea if there are autonomous car companies in NL, but I know they have redesigned their roads over time into these three categories:

1. Stroomweg / through roads: closest to limited access highways
2. Erftoegangsweg / access roads: right by residences, no more than 30 km/h in urban areas, sometimes only 15, lots of consideration for cyclists and pedestrians
3. Gebiedsontluitingsweg / distributor roads, connecting the above two, also designed solely for cars (with ped/cycle access provided by other routes, or at least thoroughly separated)

I could imagine AVs navigating 1 and 3 well, but 2 is where you need the social driving sense and if they can't do 2 they can't really serve people.


Posted by: Minivet | Link to this comment | 10-10-22 9:52 AM
horizontal rule
20

Sifu's blog is great, even if it's called a "substack."


Posted by: fake accent | Link to this comment | 10-10-22 9:59 AM
horizontal rule
21

If I took Sifu's point correctly, it's not that autonomous autos are worse drivers as that they are worse/differebt in ways that we as humans are not. We get distracted; they don't. But we can read that the bus is pulling out or that the car across from us is going to gun it when the light turns green. - they can't. We don't confuse clouds for trucks.

But we've designed traffic safety systems around the assumption that human minds are driving the cars and the AIs aren't thinking like we do.


Posted by: Cala | Link to this comment | 10-10-22 10:00 AM
horizontal rule
22

Sifu's Substack is indeed very good! I've long been very skeptical about self-driving cars and it's gratifying to see the conventional wisdom start to come around on them.


Posted by: teofilo | Link to this comment | 10-10-22 10:01 AM
horizontal rule
23

I feel like I missed my chance to write a skeptical blog post about the hyped up future presented for AVs circa 2014 when I got a brief ride in one. I can't believe that was 8 years ago.


Posted by: fake accent | Link to this comment | 10-10-22 10:01 AM
horizontal rule
24

Though who knows, maybe the silver bullet for AVs is limit them to 20mph everywhere that's not an interstate. And by extension everyone else.


Posted by: Minivet | Link to this comment | 10-10-22 10:01 AM
horizontal rule
25

19: I could imagine using different vehicles for 2 (some kind of smaller version of container shipping, where you pull into a warehouse and unload onto smaller vehicles that are designed to be very light and go really slowly in town), or just use human operated vehicles for 2 (which again would only need to be designed for 2). There's problems there for sure, but they seem eminently solvable if your goal was actually to solve them rather than try to ignore them away.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 10:02 AM
horizontal rule
26

For passenger cars you could either just have them switch out of autonomous mode for roads of type 2, or you could have big lots of autonomous cars that you rent at the outskirts of towns and you just bike to the lot, rent a car, it drives to wherever else you're going and then you bike on the other end (either renting a bike, or having cars that are designed to carry bikes easily).


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 10:05 AM
horizontal rule
27

12: this is actually a good example. There isn't, as far as I know, a general problem of racist drivers failing to see people of color crossing the street that leads to different accident rates.* There is a pretty well known problem of training algorithms on light skin and then discovering that it doesn't work on darker skin. Pulse oximeters, hand sanitizer, modulo Facelock, etc. So crosswalks are designed with the assumption that drivers who aren't distracted or drunk can tell when there are people in the crosswalk. If that's not true, the crosswalk isn't safe, even if you tell the AI to check if there are people in the crosswalk before proceeding.

*Wouldn't be surprised tho.


Posted by: Cala | Link to this comment | 10-10-22 10:06 AM
horizontal rule
28

The asterisk is exactly my point, we probably haven't looked for it, but it seems likely to me that similar effects exist with people.

(I should read the article, but it seems really weird to me that self-driving cars would have this problem at anything like the scale that pulse oximeters do. The ratio of visible clothing to visible skin means surely they're judging mostly off clothing? Especially if someone has their back to you then you're often not going to be seeing any skin at all.)


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 10:12 AM
horizontal rule
29

Remote takeover a la 18 is a cool idea for e.g. Amazon warehouse work but iffy for driving because the lag is a safety issue.

Witt, I think you're missing the forest for the trees with the skin perception stuff. The tech may work slightly less well for darker-skinned people than lighter-skinned people while still being much safer for all groups than human drivers.

I'm unsure how/why Sifu is CTOing a self-driving car company (excuse me, Autonomous Driving company) while also continuously blogging about how self-driving is so fundamentally unworkable. Wondering if it has something to do with the fact that his company seems to use a targeted module-level supervised learning approach for one or two specific subtasks (predicting pedestrian behavior) when it's perfectly clear that the way forward is giant-ass SOTA vision transformers pretrained on large-scale corpora hooked up to giant-ass SOTA RL models GATO-style.


Posted by: torque | Link to this comment | 10-10-22 10:15 AM
horizontal rule
30

29.last: Former company, hence starting the blog. (This is more obvious if you read in forward chronological order rather than blog order.)


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 10:20 AM
horizontal rule
31

28: so, ex recto, but - if we did find a difference (e.g. pedestrian fatalities by race) given humans, we'd probably find that it was tied to structural racism (more PoC walking in congested areas, badly designed unwalkable core, bad interface with public transit, e.g.) rather than vision and our fixes would be like: make the crosswalk have caution lights and blink if someone pushes the button before crossing, or reroute the bus line. Those fixes wouldn't work if the problem is literally "car/driver didn't recognize pedestrian" because it is OK to proceed through crosswalks if there is no one there.


Posted by: Cala | Link to this comment | 10-10-22 10:24 AM
horizontal rule
32

He's still front and center on their Our Team page ¯\_(ツ)_/¯. (Incidentally, I wonder if J3ff Z@r3mb@ is related to Wojci3ch of OpenAI fame.)


Posted by: torque | Link to this comment | 10-10-22 10:24 AM
horizontal rule
33

29.last: the issue isn't one of finding the right model architectures. But best of luck!


Posted by: Sifu Tweety | Link to this comment | 10-10-22 10:27 AM
horizontal rule
34

The tech may work slightly less well for darker-skinned people than lighter-skinned people while still being much safer for all groups than human drivers.

See, it feels like every time another flaw is identified the tech instinct is to say "But this must be outweighed by the manifest other across-the-board safety improvements, right?" And this still seems to be nothing more than an assumption.

The Doctorow post I didn't link above pointed out:

The average school-bus driver clocks up 500 million miles without a fatal crash... US traffic deaths are a mere one per 100 million miles driven, and most of those deaths are due to recklessness, not inability.
Waymo claims to have driven 20m miles - that is, 4% of the distance we'd expect a human school-bus driver to go before having a fatal wreck. Tesla, meanwhile, has stopped even reporting how many miles its autopilot has managed on public roads. The last time it disclosed, in 2019, the total was zero.

Solutions like 26 are interesting but still feel like they're solving for the problem of not having AVs, as opposed to the problem of street safety. I'm happy to focus on narrowing and slowing traffic and increasing public transit and cycling, the solutions proven to work at scale (and that will still have a major effect if AVs do make it through the current gauntlet).


Posted by: Minivet | Link to this comment | 10-10-22 10:30 AM
horizontal rule
35

33: Agreed! It's mostly about harnessing larger datasets and more compute. (Guessing you've read this already but posting for other curious people re: AI progress: http://www.incompleteideas.net/IncIdeas/BitterLesson.html)


Posted by: torque | Link to this comment | 10-10-22 10:34 AM
horizontal rule
36

I... sort of forswore arguing about things on here years ago but, no, it's not those things either. Nobody even knows how to ask the right questions to train models around. The companies that picked up my former team might have a leg up, we'll see! But even then there are social and structural constraints that make me doubt robotaxis per se will ever be a viable product. Certainly they won't ever be the product (most) people (including in the industry!) expect them to be. ANYHOW hi all bye now


Posted by: Sifu Tweety | Link to this comment | 10-10-22 10:43 AM
horizontal rule
37

36 IIRC Sifu's excellent blog substack covers those issues in a number of posts.


Posted by: Barry Freed | Link to this comment | 10-10-22 10:50 AM
horizontal rule
38

Framing these comments as questions to maybe get past your arguing filter. What's wrong with this question: "How do we maximize speed of travel while minimizing (accident risk * E[harm caused by accident])?"?

I feel like the biggest issue in the way of just throwing giant transformers at the problem is that the existence of the vehicles themselves will cause people to behave differently and lead to out-of-distribution data. But could this not be addressed by gradual rollouts + frequent retraining?


Posted by: torque | Link to this comment | 10-10-22 10:56 AM
horizontal rule
39

Does Standpipe have a substack now?


Posted by: Minivet | Link to this comment | 10-10-22 11:00 AM
horizontal rule
40

I don't even understand what 38.1 is asking. You want to train them by just having them drive around killing pedestrians and after they've killed a few billion pedestrians then they'll be trained ok? Or you want to train them driving around a virtual world where you've somehow realistically captured pedestrian and driver behavior? The latter sounds much harder than making a self-driving car.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 11:02 AM
horizontal rule
41

39 is, of course, already answered at Standpipe's substack.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 11:02 AM
horizontal rule
42

40: train mostly on historical data from cameras installed in human-driven cars + throw in some simulation data too for the RL piece.


Posted by: torque | Link to this comment | 10-10-22 11:07 AM
horizontal rule
43

How are you deciding on the safety of decisions made on the basis of recorded camera data?


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 11:14 AM
horizontal rule
44

Not sure if this is exactly what's done (I don't work in that industry) but a reasonable approach would be to try to take as your X the last few thousand frames of visual input, and your Y as the next action by the human driver. This does not give you real-world safety guarantees as you point out but is a good place to start.


Posted by: torque | Link to this comment | 10-10-22 11:19 AM
horizontal rule
45

Let's hash out all the solutions to a problem that people have been working on for years via arguing from first principles in comment threads. Seems promising!


Posted by: fake accent | Link to this comment | 10-10-22 11:45 AM
horizontal rule
46

45: It worked when we reinvented the wheel!


Posted by: peep | Link to this comment | 10-10-22 11:47 AM
horizontal rule
47

44: How do you know if what you did was close enough to what the driver did to be safe?


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 11:51 AM
horizontal rule
48

First, someone explain "substack" to me. Then I'll explain why I personally drive to deter Pittsburgh lefts.


Posted by: Moby Hick | Link to this comment | 10-10-22 1:05 PM
horizontal rule
49

This article made it look like there was some progress towards self-driving robot vehicles (starting on a smaller scale; traveling shorter distances, with less at risk if there was an error). I don't know if that offer a route towards bootstrapping more complex self-driving vehicles: https://www.fullstackeconomics.com/p/the-robot-takeout-revolution-is-closer


Posted by: NickS | Link to this comment | 10-10-22 1:09 PM
horizontal rule
50

Also, thanks for the link to Sifu's blog; that does look good, and I'll plan to read the archives.


Posted by: NickS | Link to this comment | 10-10-22 1:11 PM
horizontal rule
51

On second thought, no one explain substack to me. The reason is that I saw a guy drive into a kid while doing a Pittsburgh left.


Posted by: Moby Hick | Link to this comment | 10-10-22 1:12 PM
horizontal rule
52

44: the problem is that stacking up Lots More Data gives you lots more common events. what you need is a much, much greater density of a much, much wider diversity of rare events. you can generate as many synthetic events as you want, although the synthesis is not going to give you any new unpredicted events you didn't ask it for.


Posted by: Alex | Link to this comment | 10-10-22 1:38 PM
horizontal rule
53

Great Vox paragraph from further up:

The report, "Predictive Inequity in Object Detection," should be taken with a grain of salt. It hasn't yet been peer-reviewed. It didn't test any object-detection models actually being used by self-driving cars, nor did it leverage any training datasets actually being used by autonomous vehicle manufacturers. Instead, it tested several models used by academic researchers, trained on publicly available datasets. The researchers had to do it this way because companies don't make their data available for scrutiny -- a serious issue given that this a matter of public interest.

yes, the last sentence is bad, but didn't sentences 2 and 3 worry anyone?


Posted by: Alex | Link to this comment | 10-10-22 1:44 PM
horizontal rule
54

Ugh, we're going to end up with cars that are remotely operated by underpaid drivers in some warehouse somewhere, where each person is remotely operating many vehicles at a time, aren't we?

This made me laugh, and I'm pretty sure Elizabeth Holmes is now stealing this and planning her new autonomous cars VC campaign.


Posted by: heebie-geebie | Link to this comment | 10-10-22 2:55 PM
horizontal rule
55

I'd missed the most recent Waze post. There's this one spot in town here where Google Maps tells you to avoid a traffic light by instead turning left from a small road across a major divided highway that's two-lanes in each direction. It's completely insane, and all to save like 10 seconds of waiting at a light. Does anyone actually want that behavior?

I'd always rather spend an extra minute or two than turn left without a light.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 3:02 PM
horizontal rule
56

55: This person?


Posted by: Minivet | Link to this comment | 10-10-22 3:20 PM
horizontal rule
57

49: Just the other day saw an article that autonomous robot experiments in the wild have all completely failed as businesses. Like, those cute little coolers-on-wheels with the light on a mast? They don't end up working as delivery vehicles, even if they're able to perform the basic navigation tasks.

To me it's all always seemed like a not-yet-existent solution in search of a problem. Like, of all the problems cars pose to society, how many are unavoidably caused by human drivers? We know that US death/accident rates are a choice, not inherent. So the bar for "safer" is quite high. Fantasists liked to imagine AVs that would be tremendously efficient (eg those computer animations of zero-clearance intersections), but that's clearly nowhere on the horizon. Meanwhile, AVs don't ameliorate any of the negative environmental impacts of regular cars, and both empirical research and the vaguest familiarity with human nature show that dream AVs (cheap ones that come when summoned and drive you anywhere while you read your phone) only increase VMT.

I get why academic researchers pursued it--it's an obvious target--but the commercial vision never made any sense. Not denying that there's any possible use case, but the sentiment in 11 was never justified. No offense to UPTEGI, but a lot of the enthusiasm has ISTM always been driven by people who are very unconfident drivers who project that onto others. People who think it's self-evident that almost no humans should be allowed to drive bc it's so inherently dangerous. But, again, the literal majority of the danger of driving in the US is social/political choice, nothing inherent to the task.


Posted by: JRoth | Link to this comment | 10-10-22 3:35 PM
horizontal rule
58

Like, those cute little coolers-on-wheels with the light on a mast?

The main example of that business I know, the robots were guided remotely by Colombians. The automation worked well enough that one person could manage several at a time and intervene as needed when one encountered anything complicated.


Posted by: Minivet | Link to this comment | 10-10-22 3:40 PM
horizontal rule
59

Christ that's a long comment. I'll try to be more succinct:

Every intervention we could/would make to make human driving safer would also improve the rest of our world/lives*. If AVs arrived fully-formed, meeting all the lofty claims, the only real improvement would be for long distance commuters, wealthy people, and edge cases (eg it might be better for handicapped people who don't need assistance getting themselves or their stuff into vehicles).

*street design that protects peds/cyclists, lowering VMT by densifying cities, speed cameras, etc


Posted by: JRoth | Link to this comment | 10-10-22 3:40 PM
horizontal rule
60

Its a weird irony that flying autonomous cars would work a lot better than regular autonomous cars. There is a lot that can be done with controlled airspace that just isn't going to work on city streets. We just have to figure out the gravity problem.


Posted by: Spike | Link to this comment | 10-10-22 4:21 PM
horizontal rule
61

It seems to me that the biggest most accessible payoff of autonomous vehicles would be not requiring drivers for long-haul trucking.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-10-22 4:32 PM
horizontal rule
62

47 is a great/tough question and finding that threshold probably requires on-road testing and hundreds or thousands of deaths.


Posted by: torque | Link to this comment | 10-10-22 4:52 PM
horizontal rule
63

It's true that we could reduce many maybe even most deaths with increased enforcement but to say this means safety benefits of AVs don't matter misses the point which is to give us better tradeoffs on the safety/speed and convenience frontier


Posted by: torque | Link to this comment | 10-10-22 4:54 PM
horizontal rule
64

i fully & enthusiastically support the one already existing proven techno vehicle innovation - speed controllers. 20 mph max on all urban streets, done & dusted.

if you want people to be able to get places faster, build fast reliable dense transit (including regional & interregional trains,) & bike (scooter, skatey, etc.) infrastructure.

for long haul freight ... 🚂🚂🚂

the only "problem" avs solve is creating a giant hole to pour capital into without regard to social outcomes.


Posted by: dairy queen | Link to this comment | 10-10-22 5:39 PM
horizontal rule
65

AVs don't ameliorate any of the negative environmental impacts of regular cars

AVs are fresh-off-the-line EVs and hybrids, albeit with a giant bolted-on battery-suck. Their carbon impact depends a lot on what they're replacing: ICE vehicle miles, public transit, walking or biking. Nobody has the data to say for sure how it nets out.


Posted by: Opinionated Lincoln Towncar | Link to this comment | 10-10-22 7:34 PM
horizontal rule
66

Being an electric vehicle has nothing to do with being autonomous. The point of the comment you're responding to is that they don't have any environmental benefit over the same miles driven in an otherwise equivalent non-autonomous car.


Posted by: LizardBreath | Link to this comment | 10-10-22 7:52 PM
horizontal rule
67

Yeah, and the IRA added $7,500 subsidies for EVs. those are coming regardless.


Posted by: Minivet | Link to this comment | 10-10-22 8:02 PM
horizontal rule
68

I stipulated that the effect is muddled but the assumption is that shifting vehicle miles traveled to EVs will happen faster if we also shift human-driven private-car miles to taxi-fleet AVs. It is much cheaper to shift trips to an AV than it is to buy an EV.


Posted by: Opinionated Lincoln Towncar | Link to this comment | 10-10-22 8:12 PM
horizontal rule
69

The post I find myself mulling over is, It's not clear anybody wants autonomous cars. He points out that many of the fantasies of self-driving cars are fantasies of not having to deal with a world of cars (not having to drive, but also not having to deal with traffic, etc . . . and that's not something that AVs can actually deliver.

The autonomous vehicle industry--particularly the companies developing and testing robotaxis--has gotten away for too long with selling a vision of the future that they should know perfectly well is never going to come to pass.

I have a slightly more optimistic reaction -- but with no reasonable basis at all. I'm somewhat curious about AVs, but am so very unlikely to be an early adopter that my views are completely meaningless in terms of imagining a market.

I am also, like Sifu, someone who dislikes driving and has been primarily a bike commuter for my adult life. In fact I will somewhat often not do something that I would otherwise do because I don't want to drive*. But, unlike Sife, I live in a much smaller and less urban city, so this doesn't quite describe my experience:

One way to explain the value proposition of autonomous cars--particularly robotaxis--is that they will make taking trips by automobile in urban centers more pleasant, more seamless, and less aggravating. This is a goal that I have never supported; some of the implications of aiming for it could lead to a wave of urban desecration as painfully misguided as that which accompanied the advent of the automobile. Attempting to avoid that outcome was my primary personal justification for the work I did in autonomous cars.

From my perspective (as someone who is generally healthy and likes walking) one of the better uses for an AV would be if it could drive me to a light rail stop on the outskirts of [neighboring urban center], _park itself_, let me spend the day heading into the city on transit, and pick me up somewhere when I wanted to head home. It would be nice to have a car take me directly to my destination, but it isn't necessary, and that's not how I currently use my car**.

Expanding on that, the way that I could imagine AVs being a win for urbanism is separating parking from popular destinations. Rather than having a giant parking lot at the mall, could you have a small parking lot (for dropping off and picking up), and then a larger (free) parking lot 5-7 minutes of drive time away that the AV could be routed towards automatically (and a shuttle bus for people who aren't in AVs) or, similarly, could we separate the parking from the downtown core (with restaurants and bars), even by 6-8 blocks, and make half the downtown streets pedestrian only?

That idea seems unlikely, and a synergy between AVs and urbanism isn't, directly, an environmental benefit, but in the long run it could offer a much more environmentally friendly lifestyle (if you can put the parking lots a mile away from the light rail stop, it means you can put more housing within walking distance of light rail). But I appreciate his point that this isn't the future that existing AV companies are trying to create.

* for example, a local pizza company sent me a promotional offer of a free pizza for my birthday; I put off getting it until days before the offer expired largely because I didn't want to make the 12 minute drive (one-way) to pick up the free pizza.

** When taking a vacation out of town my preference is, when possible, to drive to the place I'm staying and then park the car for as much of the vacation as possible and walk around the local neighborhood. This only works for certain sorts of destinations, but I'm not good at taking vacations in general.


Posted by: NickS | Link to this comment | 10-10-22 8:17 PM
horizontal rule
70

That is, you are speculating that it would be cheaper if there were AVs that could function as taxis. Other than in very limited places and times, there aren't. I don't think you can give them credit for carbon reduction when they don't functionally exist yet.


Posted by: LizardBreath | Link to this comment | 10-10-22 8:18 PM
horizontal rule
71

We're working entirely in the world of hypotheticals so banning speculation doesn't seem like fair play? You are comparing an AV to an alternative vehicle, 1:1. I'm comparing a functioning future world with AVs to an alternative one without, and I think it likely (but not certainly) represents a net carbon win. The positive marginal case would be: folks who skip ICE car trips in favor of an AV taxi. (Also, to some extent: folks who don't buy a private EV and rely on AV taxis instead.) The negative case would be: folks who skip a bus or bike trip in favor of an AV. It's totally unclear which group would be larger.

It's also totally unclear if we can get to that hypothetical future world. But that doesn't mean we shouldn't try?


Posted by: Opinionated Lincoln Towncar | Link to this comment | 10-10-22 8:55 PM
horizontal rule
72

the avs infesting sf are loaded up with a large amount of equipment adding to the weight of the vehicles, with additional weight already being an not inconsiderable environmental disadvantage of evs. shittier aq & increased toxic runoff, to not solve problems via a method that even if it ever worked as advertised would only ever *solve" an issue (the unpleasantness of traffic) for a tiny number of v v wealthy people in a community that didn't ask for & doesn't want to be the laboratory - it's not great!


Posted by: dairy queen | Link to this comment | 10-10-22 9:00 PM
horizontal rule
73

Counterpoint to 72 they have cute names like "Muffin" and "Piccolo"


Posted by: torque | Link to this comment | 10-10-22 9:06 PM
horizontal rule
74

I didn't want to make the 12 minute drive (one-way) to pick up the free pizza.

Zume Pizza, where have you gone? They were going to have robots make pizza in self-driving vans.


Posted by: fake accent | Link to this comment | 10-10-22 10:46 PM
horizontal rule
75

71: But you're talking about giving autonomous cars credit for being clean new EV taxis that could be a faster replacement on a vehicle-mile basis for the dirty old ICE cars currently on the road than waiting for the ordinary rate of turnover in the fleet of cars on the road. That works, even as speculation, if the replacement happens this year, when the fleet of cars on the road is almost entirely dirty old ICEs.

But it's not going to happen this year, because functioning AVs don't exist yet. I wouldn't count on next year or the year after either. And every year that goes by, the non-autonomous ICE fleet turns over, and more and more of the cars on the road are clean new EVs. The benefit you're speculating about getting from AVs on this front is time-sensitive: it's completely dependent on their being a large-scale replacement for vehicle miles before EVs are otherwise most of the cars on the road, and that timing doesn't look terribly plausible to me.


Posted by: LizardBreath | Link to this comment | 10-11-22 1:02 AM
horizontal rule
76

When I think about AV's I imagine a vehicle that would take my 90-year old mother to concerts or films and, more critically, be there to pick her up at the end of concerts. Being safer and more reliable than Tel Aviv cabbies is a pretty low bar.


Posted by: peep | Link to this comment | 10-11-22 3:55 AM
horizontal rule
77

I think it probably really isn't a low bar at all. You may find the way they drive unsettling, and less safe than human drivers you approve of more, but that doesn't mean that AV's are anywhere near being able to drive as safely and usefully as a Tel Aviv cabbie in Tel Aviv.

As noted earlier in the thread, there's a lot of casual eye rolling about how badly humans drive in general, so improving safety has to be in reach. But the only standard you're comparing bad human drivers to is excellent human drivers -- there's nothing yet that's not human that drives better than even an ordinarily lousy human driver.


Posted by: LizardBreath | Link to this comment | 10-11-22 4:15 AM
horizontal rule
78

77.1: You're right, of course. I was kidding about the safety part, although not about the reliability part. But the one is useless without the other.


Posted by: peep | Link to this comment | 10-11-22 4:43 AM
horizontal rule
79

The post I find myself mulling over is, It's not clear anybody wants autonomous cars.

That post contained a frankly extraordinary argument - that AVs (assuming they are built and fulfil their builders' promises completely) will be unacceptable to the public because they will always drive safely and follow traffic laws, and passengers will be infuriated by this. It's so strange and repellent, in fact, that it makes me doubt the soundness of everything else in the post.


Posted by: ajay | Link to this comment | 10-11-22 5:16 AM
horizontal rule
80

There's also a huge amount that could be done quite easily to make human drivers better. As already pointed out by multiple people above in the thread, death rates from automobiles are not a fixed constant.

Lots of countries have _vastly_ fewer deaths per capita even controlling for miles driven, population and the other variables. That might involve changes to road furniture/signage, reduced speed limits, better enforcement, making it really undesirable to drive insanely huge SUVs, etc. But it also might just be better driver education and certification. Scandinavians and British people, for example, have far fewer fatalities (between a approx 1/4 and approx 1/2 depending how you are making the comparison).


Posted by: nattarGcM ttaM | Link to this comment | 10-11-22 5:21 AM
horizontal rule
81

The Doctorow piece too seems to be a bit off the rails with things like this:

US traffic deaths are a mere one per 100 million miles driven, and most of those deaths are due to recklessness, not inability.

Oh, well, that's OK then. As long as it's only recklessness, that's sort of fun, in a way.

He's wrong about the figure by about 46%, mind you, and you note that he's exploiting our natural inability to understand big numbers. A hundred million miles sounds like a terribly big number, so one in a hundred million miles must be a trivially small risk! But you only think that because you don't know that Americans drive around three trillion miles per year. Hence why traffic deaths are at 46,000 a year (and another 4.4 million injured).

It's a lot scarier when you put it that way, isn't it? Doctorow knows that.


Posted by: ajay | Link to this comment | 10-11-22 5:23 AM
horizontal rule
82

Scandinavians and British people, for example, have far fewer fatalities (between a approx 1/4 and approx 1/2 depending how you are making the comparison).

I've seen a convincing argument that driving on the left accounts for some of this, at least in the UK - you're using your (generally) dominant right hand to do the tricky task of steering, and leaving the non-dominant left hand to do less safety-critical things like changing gear.


Posted by: ajay | Link to this comment | 10-11-22 5:28 AM
horizontal rule
83

Americans vastly predominately drive automatic, so their right hands are free to be on the wheel despite being in the center of the car. I'd believe it was possible that driving on the left was safer, but it'd have to be something more complicated about which is your dominant eye or something, not that they're steering with their left hands.


Posted by: LizardBreath | Link to this comment | 10-11-22 5:44 AM
horizontal rule
84

I use my right hand to hold my coffee.


Posted by: Moby Hick | Link to this comment | 10-11-22 5:57 AM
horizontal rule
85

Come to think, that is a point. It's not just shifting, anything else you're doing with a hand at all is going to be in the center of the car. Okay, maybe there is a handedness effect.


Posted by: LizardBreath | Link to this comment | 10-11-22 6:02 AM
horizontal rule
86

I was trying to be pointless.


Posted by: Moby Hick | Link to this comment | 10-11-22 6:04 AM
horizontal rule
87

What autonomous cars do now is pretty amazing but there is no proof that they will get to anything actually useful and given the money already invested I really doubt we will get anything any time soon.


Posted by: lemmy caution | Link to this comment | 10-11-22 6:36 AM
horizontal rule
88

79 last is either fabulously hilarious in which case kudos! or premised on a belief about the desires of drivers completely unsupported in the us


Posted by: dairy queen | Link to this comment | 10-11-22 7:09 AM
horizontal rule
89

Driving in the UK is so much more objectively difficult than driving in the US, but the drivers are so much better that you get fewer accidents anyway. God help you if you dropped all American drivers onto UK roads (putting aside the left/right differences). Roads and lanes here are so wide and traffic patterns so simple (Boston aside).

The point in 79 strikes me as so obviously true once you think about it, that I'm really surprised that ajay had that reaction. It is certainly true that Americans speed a lot more than British people, but I think that's not the only thing that's going on here. A good UK example is that you can't drive through a town in the UK while staying in your lane. I'm not actually sure how this is dealt with in UK law (presumably it's actually legal to go into the oncoming traffic lane when it's safe and you're in town?) but there's a lot of social behavior here where people will let someone go through under certain circumstances (maybe signaled by blinking lights, maybe signaled by noticing that people are taking turns in some natural way). But if you have an autonomous vehicle it really could just get stuck for a long time trying to get around parked cars going through town on a street that's marked like it's 2-track but is 1-track in practice.


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-11-22 7:19 AM
horizontal rule
90

The front grill of new American SUVs are targeted right at the heads of pedestrians and people in regular cars.


Posted by: Moby Hick | Link to this comment | 10-11-22 7:21 AM
horizontal rule
91

I think there's a good chance that the greater difficulty of driving in the UK is a cause of the greater safety -- if it's impossible to zone out and not pay attention while you're driving, people will be systematically more attentively engaged.


Posted by: LizardBreath | Link to this comment | 10-11-22 7:22 AM
horizontal rule
92

Americans also really really like to go through red lights. I'm always shocked by how predictably when I have a close call where I decide to go through a light that just turned orange a car, another car behind me goes through too (sometimes two!). And then the person waiting to make a left turn after the light turns red has to wait for all those people going straight to go through the red before they can turn! (Pittsburgh aside.)


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-11-22 7:27 AM
horizontal rule
93

Driving in the UK is so much more objectively difficult than driving in the US, but the drivers are so much better that you get fewer accidents anyway.

It's slower, certainly, and routefinding is probably harder, but is the driving more difficult?

The worst driving experience I ever had was in the US - LA traffic at rush hour. I felt very strongly that I was in conditions that were only safe for someone with considerably more driving skill than me. I've never felt like that driving in the UK (at least not on the road).

But that may well have been familiarity rather than objective difficulty. Drop the average US driver in London traffic and he'd probably feel much the same.

Americans speed more than British people, I would think, for two reasons: first, there are more roads on which it is obviously safe and physically possible to speed. You just aren't going to be able to drive through most British towns at 45mph, even if you want to. And the other is that the UK has a lot of speed cameras and, AFAIK, the US doesn't.


Posted by: ajay | Link to this comment | 10-11-22 7:32 AM
horizontal rule
94

Lots of Americans are working out frustration by driving.


Posted by: Moby Hick | Link to this comment | 10-11-22 7:32 AM
horizontal rule
95

There's an irony that the physical traffic-calming measures contemplated in, e.g., 80 would make it easier for AVs to AV and would constitute, in some respects, the very same new infrastructure considered so outrageous in the OP.


Posted by: Opinionated Lincoln Towncar | Link to this comment | 10-11-22 7:33 AM
horizontal rule
96

I was nearly in an accident Saturday with a big grilled, lifted truck that couldn't see me as it began backing out of the parking space. The back of the truck would nearly have cleared the hood of the Accord. Guy had no idea I was there. Nationally, pedestrian fatalities are up and it's because the cars are too damned big.

And adding to the theory: Utah is full of wide roads and astonishingly horrible drivers, in a clueless way - like don't make left hand turns from the far right lane. My thinking is that idiot behavior like that would have been drummed out of you by close calls in places with narrower, busier streets before you earned your license. Here most of the time nothing bad will happen.


Posted by: Cala | Link to this comment | 10-11-22 7:38 AM
horizontal rule
97

I feel like I'm constantly a second away from an accident in the UK just from accidentally straying out of my lane or scraping a wall or hitting parked car. Fortunately I've so far avoided it, but it requires so much more focus than the US. (I guess part of this is that I drive a UK-sized car on roads designed for enormous SUVs.)


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-11-22 7:39 AM
horizontal rule
98

All of 93.last is true. I'd add to that in many cases the UK has speed limits that are higher than the US, and where safe driving requires driving below the speed limit. (This leads google maps to think you should take a "shortcut" on some 1-track logging road because the legal speed limit is 60mph.) The UK generally has clear rules of what the speed limit will be in a given situation, while the US will often have weird bespoke speed limits that change a lot (and which everyone ignores).


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-11-22 7:44 AM
horizontal rule
99

It's weird how angry I still am about various google map errors. But the one mentioned in 98 was one of the all-time "go home Google Maps, you're drunk" moments:
https://goo.gl/maps/aHER5jyjZWdGqWa16


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-11-22 7:55 AM
horizontal rule
100

Speaking of which, how should AV handle "advisory" speed limits?


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-11-22 8:42 AM
horizontal rule
101

Probabilisticly.


Posted by: Moby Hick | Link to this comment | 10-11-22 8:56 AM
horizontal rule
102

One of the horrid things about speed limits in most of the US, due to laws originally against local speed traps but adapted to the car-speed-above-all ethos, cities essentially don't have the power to lower speed limits for safety reasons. they can only set speed limits by doing a traffic study and setting traffic to the 85th percentile of speed their study finds, e.g., bound to increase over time. (Of course engineering has as much/more power to slow cars down, but we also engineer streets with speed first in mind.)


Posted by: Minivet | Link to this comment | 10-11-22 8:59 AM
horizontal rule
103

102 is bizarre - the idea of saying that you can only make it illegal if fewer than 15% of the population are doing it!


Posted by: ajay | Link to this comment | 10-11-22 9:09 AM
horizontal rule
104

That's freedom baby!


Posted by: Unfoggetarian: "Pause endlessly, then go in" (9) | Link to this comment | 10-11-22 9:12 AM
horizontal rule
105

I meant to conclude that given that prospect, cities often opt to not do a traffic study so as to just leave the speed limits frozen in amber for decades. So that might explain some of the "weird bespoke limits" in 98.

103: I'm told it may sort of make sense for very rural roads, where you want to set the limits to what is natural. But even there, I bet there are better methods.


Posted by: Minivet | Link to this comment | 10-11-22 9:29 AM
horizontal rule
106

if it's impossible to zone out and not pay attention while you're driving, people will be systematically more attentively engaged.

In the middle of last century, there was a lot of energy around reducing the impact of advertising signage--not billboards per se, but neon, animated, highly engaging signs. Opponents claimed that they were safety risks, distracting drivers. But it turned out that they got drivers' attention in a useful way--their eyes might briefly leave the road, but they'd also raise their overall alertness level.

In general, US road design is dominated by the idea that the smoothest possible path for drivers will be the safest, and the fact that it's not at all true is treated more as a fun fact than as a reason to change engineering practice.


Posted by: JRoth | Link to this comment | 10-11-22 12:56 PM
horizontal rule
107

98 is interesting, but it's 100% the case that US road design policy essentially says that road design has to be safe for speeds in excess of the limit*. Basically the only exceptions are local roads that are 25 mph by default but are actually too narrow/twisty for that speed. But that's a municipal condition, not state or federal DOT.

*this is partly embodied by the "85% rule" that says the limit should be set at whatever the 85th percentile driver will go. I used to understand this as a rule of thumb, but a lot of activists insist it's literally SOP**. I now believe the truth is in between.

**in line with how Calvin's dad says they determine weight limits on bridges: keep driving heavier trucks over until they collapse, then rebuild the bridge and mark the weight of the last truck.


Posted by: JRoth | Link to this comment | 10-11-22 1:02 PM
horizontal rule
108

Oops, 107 crossed with 102. And explains the actual mechanism usefully. Thanks.


Posted by: JRoth | Link to this comment | 10-11-22 1:04 PM
horizontal rule
109

107: Whenever this comes up, Caltrans and other car-throughput-focused bodies love to insist the 85th percentile rule isn't the straitjacket activists make it out to be, but when city staff are asked to reduce posted and/or engineered speeds, it's suddenly extraordinarily hard to get around.

For a qualitative idea, look at all the hoops you have to jump through under the latest law letting CA cities disregard it, even just in a few specified circumstances. And cop associations still opposed it! (I guess I can't rule out the possibility that California law enshrines it more fixedly than in other states.)


Posted by: Minivet | Link to this comment | 10-11-22 1:20 PM
horizontal rule
110

109: I assume the rule exists here, but dollars to donuts it only restricts cities on state-owned roads. I've never heard of it having an impact on any street that isn't a state route, and I've been paying attention to this stuff for 25+ years.

It would certainly be in character for CA to have more imposition from state bureaucracy on local decisionmaking.

Come to think of it, Pittsburgh has a couple state routes that are effectively limited access highways with posted speed limits that famously don't match IRL speeds. Wouldn't surprise me in the least if those limits are leftover from the past when IRL speeds were much closer to the posted*, and that the city keeps them in place rather than find themselves with a 60 mph arterial smack in the center of town.

*one of them has been gradually rebuilt to full interstate standards, without a single grade crossing from end to end, but retains a 45 limit, possible 35 closer to town.


Posted by: JRoth | Link to this comment | 10-11-22 1:30 PM
horizontal rule
111

The folk wisdom around here was that the traffic-study requirement (which contains the 85% rule, likely among other things) was a possible route to getting speeding tickets dismissed. You're caught doing 45 in a 35 zone, you find out who put up the 35 sign, and see if they can cite the study that justified it - if not, the sign wasn't put up legitimately and the actual limit is the state default for that type of road, which is probably 45 or 55, so you're not actually in violation.

I don't know how often this works in practice (and clearly doesn't help if you're speeding over 65), but it's at least sometimes.


Posted by: Nathan Williams | Link to this comment | 10-11-22 6:34 PM
horizontal rule
112

111: Yes, one activist I know has suggested reducing speed limits without studies, in the full knowledge that it can't be used to ticket people, but for the sake of the benefits of lower speeds from the signage - and potentially getting the traffic engineers to redo the road to the new signed-if-unenforceable speed limit.


Posted by: Minivet | Link to this comment | 10-12-22 10:40 AM
horizontal rule
113

What about making robot dogs that dart into the street and slow down traffic?


Posted by: Moby Hick | Link to this comment | 10-12-22 10:48 AM
horizontal rule