Skip to main content


Techbros: self driving cars are inevitable!

Also techbros: prove you are human by performing a task that computers can’t do, like identifying traffic lights.

Why do you think the tech bros want us to identify traffic lights, bicycles etc. Its so they can better train their "AI" models to spot them so the self driving cars become a reality. We are all beta testers

@RobCornelius I’m fully aware. I’m also aware that they are STILL doing both, simultaneously. Only one can be true. Which is it?

(Hint, it’s “computers can’t identify traffic lights”. Self driving cars are, in fact, randomised murderbots)

reshared this

@RobCornelius notice, BTW, it’s always American street furniture, a country well known for being a huge international outlier in terms of not using internationally standardised vernacular for street furniture.

reshared this

@Alexandra Lanes @RobCornelius Their road signs are utterly unlike anyone else's, because fuck us, that's why.
@ajlanes Man, these aren't even that hard to follow, why isn't Canada a signatory for this either :I Probably just playing follow the leader with the US on that one, bah

Its down to certain peoples egos (specifically Muskrats) versus reality. Its not going to end well, that's for sure.

Have you ever noticed you are never asked to identify street objects in upscale neighbourhoods? OK posh neighbourhoods are less cluttered but....

@RobCornelius Or anywhere that’s not North America, with its own “not invented here” take on how streetscapes should look.
Perhaps it is because the self-driving AI's already know what objects are unambiguosly traffic lights but still require human input to answer the philosophical questions of whether the iconic depiction of a traffic light painted on a sign counts as a traffic light
reCAPTCHA asks user to "Select all squares with traffic lights" but the photo is of a yellow diamond shaped warning sign with a stylized icon depicting a traffic light

@bornach

🎶
I like traffic lights
No matter where they've been
I like traffic lights
But only when they're green
🎶

@goatsarah @robcornelius

@axnxcamr @bornach @robcornelius I'm more worried about the baggage retrieval system they've got at Heathrow
@David S @Sylvain @Bornach @RobCornelius Also the fashions today. They don’t seem good for your feet.

I'm so worried about whether you ought to have stopped...

@bornach @Pionir @robcornelius

@Bornach @RobCornelius Except that's nothing like the widely internationally ratified sign for a traffic light, as used in most of the world (completely the wrong shape for a start!)

https://en.wikipedia.org/wiki/Vienna_Convention_on_Road_Signs_and_Signals

@robcornelius

There's also the issue of context. Those are indeed physical traffic lights but the AI should realise why it shouldn't obey them
https://futurism.com/the-byte/tesla-autopilot-bamboozled-truck-traffic-lights

@robcornelius

Self-driving cars are for sure not ready yet, and some companies and people are pushing their deployment too hard.

I'm convinced they will get good enough though.

On a related note, I saw in your profile you moved to Portugal. I hope you like it there, and there is indeed much to like, but you must have noticed how *people* drive over there. It shows up in the accident and death stats too.
I'm Portuguese and grew up there. I'm not at all convinced that people can drive safely.

@robcornelius
In the context of the trolley problem, they literally must be.

(As are BTW humans, generally speaking, when an accident happens, humans do not have the time to take a well reasoned decision, it generally is more or less a random decision if at all.)

Taking any stand on the trolley problem, and related ones, immediately raises questions of liability. So random() is the "safe" out for moral cowards.

@robcornelius
But coming back to the more general problem, and taking the point of view of a budding data scientist.

It's not a question if algorithm-driven cars cause accidents. (the AI buzzword gives me migraine)

Humans DO cause accidents too.

So the question is: Do computer-driven cars cause more or less accidents than humans.

It's hard to assess this, at the moment, as there are literally only a tiny number of really self-driving cars.

@robcornelius
And the involved entities are commercial, so they tend to throw the veil of trade secrets over most of their data.
@yacc143 @robcornelius There are roughly 70 companies researching driving automation here in the Bay Area that provide data to the state. Safety of these vehicles peaked around 3 years ago with the best performing cars having 5 times as many accidents per mile as the average driver. They appear to have reached the asymptote of the improvement curve.

Pippin reshared this.

@Marty Fouts @Andreas K @RobCornelius why, it’s almost like there’s a whole pile of bullshit surrounding the whole endeavour. So unlike the tech industry, that.

[object Object] reshared this.

@yacc143 @robcornelius @MartyFouts In really advice to be watchful for companies trying to solve the problem by making cities more self-driving-car friendly. By making streets even more the insta-death zone in front of your door.
@MartyFouts
That begs the question how come they got the permission to test on public roads and endanger the public of that's their best case?
@goatsarah @robcornelius
@yacc143 @robcornelius The politics at the state level are fascinating or would be if lives weren’t at stake. Google started using public roads without permission as Tesla still does. Cal DMV stepped in and designed a program requiring safety drivers in the test vehicles but Google’s Waymo spin-off got another state agency involved and so they and Cruise have licenses to run driverless taxi service in San Francisco. Or as someone else pointed out: money

@MartyFouts
Don't take it wrongly, but the German car makers literally spent years in R&D and ended up offering way less (co pilot system for limited use cases eg high ways/autobahns), literally citing this as the safe state of the art. They could offer more of they were willing to associate their brands with unsafe cars.

@goatsarah @robcornelius

@Andreas K @RobCornelius @Marty Fouts My car has a similar autopilot system (lane following and distance maintenance). It works very well, but you HAVE to be aware of its limitations. You CANNOT remove the human from the system
@MartyFouts @yacc143 @robcornelius This article jibes with my outsider (out of the car, not out of the area) impression. Human drivers have gotten a lot worse since covid cleared the streets in 2020 from what I see.
https://arstechnica.com/cars/2023/09/are-self-driving-cars-already-safer-than-human-drivers/
@elithebearded @yacc143 @robcornelius I have seen several similar articles but the statistics don’t support the claim
@MartyFouts @yacc143 @robcornelius
What statistics? That human drivers have gotten a lot more reckless since three years ago or something else? Where can someone else see these stats?
@elithebearded @yacc143 @robcornelius The human driver statistics are published by the Insurance Institute of America. The automation statistics are reported to the California DMV. They were published until a few years ago but now you have to ask the DMV each year for the data. Human drivers have not gotten a lot worse but automation has stopped getting better and was never as good as average drivers

@MartyFouts @elithebearded @robcornelius

Funny Swiss RE just published a self-driving cars are safer than human-driven cars, purely based on insurance data.

https://arxiv.org/pdf/2309.01206.pdf

@yacc143 @elithebearded @robcornelius There are not enough automated cars on the roads to have sufficient data for such a comparison unless they have a definition of “self driving” that is mainly based on driving assistance automation rather than driver replacement systems.
@Marty Fouts @Eli the Bearded @Andreas K @RobCornelius and if they’re looking at driving assistance systems, then I will note that they routinely try to kill you. It’s just that the driver interrupts them in the act (source: have one)
@elithebearded @yacc143 @robcornelius There are active and passive assistance systems. Passive systems like backup cameras and blind spot warning are safer than no assistance. I don’t know specifics of active systems but they seem to vary widely in quality. But certainly they can all cause problems especially lane following and automatic emergency brakes. I have had a Subaru attempt to counter my steering on ice and nearly crashed as a result. Traction control makes me cringe.

@Marty Fouts @Eli the Bearded @Andreas K @RobCornelius for me the benefit of active systems is that they are immensely valuable in keeping you fresh on a long journey by reducing cognitive load.

This more than compensates for the occasional blip where they try to kill you at 120kph and you have to intervene to stop that.

But they absolutely still do it. Not often, but dead is dead, right?

@yacc143 @elithebearded @robcornelius I’ve had a chance to read the article, which I should have done earlier. The flaw is that they are only looking at Waymo One 3rd party insurance data; but Waymo is self insured and so does not report all of their incidents this way. You have to look at data reported to the DMV for a more comprehensive analysis. Also, Waymo One only represents a fraction of Waymo’s data.

@elithebearded @MartyFouts @robcornelius
COVID/2020 has made statistics/relative references suspicious. (Actually that applies to subjective perception too)

It was such an extreme outlier, that 2021 might have some number at 75% of the 2019 value, and be still perceived as a huge relative raise over 2020.

@yacc143 @elithebearded @robcornelius That may be but it is not relevant here. Covid data, as you say, represents an outlier for various reasons. The driving automation data on the other hand is produced under controlled circumstances and reflects the best that the cars can do. It is part of why Waymo has publicly stated that it is an intractable problem.
@yacc143 @MartyFouts @robcornelius My observations walking around a lot are speeding/ignoring traffic signals (lights and stop signs)/ terrible choices about u-turns etc, got bad in 2020 _and have not improved_. Which has me question comparisons between pre-covid and now. I am just one person walking around one city, so very much ancedote and not conclusive data, but I don't see self-driving cars doing those dangerous driving things. I see them block traffic but that's about it.

@robcornelius Relevant xkcd

https://xkcd.com/1897/

(transcript available at: https://explainxkcd.com/1897/ )

@robcornelius Computers can identify traffic lights.

https://www.digit.fyi/ai-bots-can-beat-captcha-tests-better-than-humans-now/

@robcornelius Does this include our answers to the trolley bus problem?
@seb321
I listened to an excellent podcast on just that about 5 years ago. No chance of finding it again. That's in the equivalent of early Neolithic now.
@Seb
@robcornelius That makes the websites I was designing in the late 90s like primordial slime. Very low bandwidth primordial slime.

@seb321
Same.... though I made this 20+ years ago and its still there and being used http://www.jigsawstaff.com/

I think they finally got rid of my login that had god level permissions a few years ago at least.

@Seb
@robcornelius That’s impressive. I’m sure all mine disappeared into Wayback Machine eons ago.
@seb321 security issues *cough cough*
@Seb

@seb321 @robcornelius side note, just in case it has relevance: some years ago, a buddy of mine wrote a piece about the trolley problem --

https://medium.com/personified-systems/killing-the-runaway-trolley-problem-bf679b59baef

@robcornelius
While true, the way you worded this comment seems condescending
@robcornelius That's indeed the main issue here..
"select every bicycle in this image we ripped from a half loaded GeoCities site from 1996."
or identify bicycles or trains ☠️
They still ask about bikes too. It's been years. If they still can't tell the difference between them and say a lamppost, it's not good.
@Wilhelm Gere Cyclists are expendable in the coming tech utopia.
@Wgere when it comes to techbros, humans full stop are generally considered expendable depending on publicity levels.
@Wgere (both types of bikes), and pedestrian crossings, and buses.
Just last week I heard on a podcast that captchas became obsolete since AI can now read those better than humans.
@MartinSenk
If you successfulky solve the captcha, you are considered a bot and only get the SEO version of the website, that is meant for indexing by search engines. To get the actual hidden amazing content you have to choose wrong answers only!
somebody once said CAPTCHA is the reverse Turing test. Where humans try to convince computers we aren't also computers.
I wonder if this means that if you spend hours intentionally providing the wrong answers round and round in circles, that it will serve to confuse AI learning somewhat 🤔
The great plan:
1️⃣ Give techbros all your data for free
2️⃣ spend your meaningless life watching boring and annoying ads all time, even while farting at WC, giving techbros bucks in the process.
3️⃣ Help techbros to train visual AI via captchas
4️⃣ be smashed in an driving incident by a self-driving car.
5️⃣ PROFIT!
This entry was edited (3 weeks ago)

Sarah Brown reshared this.

gots damb Sarah 🔥

About five years:

"It seems like you're connecting from an unrecognized device. To access your account, please use the controls below to operate the taxi. Your pickup is at Christopher and Washington, and their destination is Columbia Heights and Vine. Be sure to drive on the right side of the street and try not to hit any pedestrians, or you will have to wait 24 hours before trying again."

there is no contradiction. You are only the cheap annotation tool. After we provide enough learning data, there will be another "challenge". 😀

Yes, there is some irony to it 😀

I guess the question for self-driving cars is when and how many.

Personally I don't drive so I'd like cars to drive themselves. And I'd like all cars to be electric.

But most urgently I'd like a *lot fewer cars* traveling *a lot fewer Kms*, regardless of the kind of car and who or what is driving.

Self-driving could even help with that in a small way, but above anything we have to want less car in the world.

I do believe that self-driving capability is inevitable though. It's just a question of time.
Incidentally, when we solve captchas they give us images that are labeled to check it we are "human", but also some that are not or by few people so aren't reliable. So we are contributing to the training data that advances computer vision. In fact, computers can solve most of the captchas already.

So I think self driving is inevitable but I hope cars aren't.

FYI machine learning models are now significantly better than humans at solving capthas

well, in theory you could have the traffic light status communicated to the cars in some other manner

but at the same time, self-driving cars would solve traffic issues no more than adding "just one more lane" every other year

to be fair, there was a study published recently that showed that AI bots solve CAPTCHAs 90% faster and 70% more accurately than humans 😅

I HAVE seen a Tesla's entertainment unit spawn an endless row of traffic lights on both sides of the road when driving behind a truck carrying traffic lights, though, so the original point still stands.

iirc that's the exact reason why captchas include things that can be encountered on roads, to help the AI recognize them.

@Ádám :fedora: 🇭🇺 Ok, 2 problems: 1. American roads. The US has not signed up to the Vienna Convention on traffic signs which most of the rest of the world uses.

2. Those cars are out there today, so if they still need training to recognise things that they aren't supposed to run into, then they're basically murderbots.

I didn't want to imply that I support self-driving cars. I think it's a terrible idea. And in its current form it actually does more harm than good.

the irony is that ML models are now better at solving captchas than humans, making captchas entirely pointless.

https://arxiv.org/abs/2307.12108

BDM reshared this.

@Ariadne Conill 🐰 Swap out the Captchas for ones that use Vienna Convention traffic signals, and see what happens.

I bet the machine accuracy plummets.

@ariadne Please let captchas die. Since my browser doesn't keep cookies I'm constantly solving them. I often refresh until I get the easiest one possible.
@binarypie @ariadne I don't keep cookies and several months ago using ArtStation website was a nightmare. 2-3 captchas every time I had to log in. For some reason it disappeared and is ok now, unless I would try to log in from Tor.
@ariadne worse than useless, even I struggle to solve them and I have no (major) impairments. My son with autism would find most of them impossible. Captchas are inherently exclusionary and I’m guessing also violate laws like the Americans with Disabilities Act. And now they don’t work, in the sense of keeping computers out but letting people in. Even @cloudflare figured this out last year https://blog.cloudflare.com/end-cloudflare-captcha/ before the AIs took over.

@kurtseifried (he/him) @Ariadne Conill 🐰 @Cloudflare And suffer from a terminal case of r/USDefaultism.

The rest of the world has to learn what American street furniture looks like to solve them.

@ariadne @cloudflare I’ve actually noticed I’ve been getting photos from other parts of the world and asked to identify motorcycles where there are scooters? And some very strange looking fire hydrants.
@ariadne which means it's probably reasonable to assume at this point that captchas are now bots training bots.
@ariadne Plot twist: If you can solve a #captcha on the first try, you're now more likely to be a robot than a human.

@ariadne Such weaknesses were pointed out at least as far back as 2009.

To quote Jonathan Wilkins' https://web.archive.org/web/20110723025202/http://bitland.net/captcha.pdf

"For instance, with a 10,000 machine botnet (which would be considered relatively small these days),
given broadband connections and multi-threaded attack code, even with only 10 threads per machine, a
0.01% success rate would yield 10 successes every second, which would provide the attacker with 864,000
new accounts per day if they were attacking a registration interface."

@goatsarah

If I was a techbro I would totally use the data collected from captchas to train AI models.

@Anton Podolsky That's what they're doing.

As a result, they will get very good at spotting things that look like American traffic lights used in Captchas, and probably dreadful at spotting anything else.

That's a straw target. They're training them to be better than _all_ humans, not just as good as _a_ human.

AI automation won't come about from a machine being able to identify some specific traffic light, it'll come when a machine can identify more traffic light signals far quicker than an average human, and on average drive safer per mile - not perfect, but better than an average human. That's maybe already happened. But there are always more traffic lights to make them better.

[ to be clear/more general to the above, "traffic light" of course could be any object requiring identification or assessment in such a situation ]

I understand the repulsion, tho.
I personally suspect we're at least on a cusp of where self-driving machines will likely be better, on average, than an average human, which would therefore be safer for everyone, even if there's mistakes, there'll be fewer mistakes.

AGI automation, well that's a different kettle of very scary fish.

@'ingie Ever seen a Captcha that uses Vienna Convention traffic signs, as used in most of the world?

I’d say they’re training them to be better at people who’ve never been to America working out what American street furniture looks like.

And that’s the problem with this sort of AI training. It is almost certainly not optimising for what you think it is.

Not sure what you meant with the American versus Viennan Convention bits. The training is to allow them to learn things that aren't easily trained on standards.

Tho I totally agree with your latter point. It's not really important for automated driving as it's not AI driving, as the AI model isn't in control, it's part of a feedback system of sensors to a conventional autopilot system.
Like in an airplane... which can already land themselves safely if needed.

... but i fully agree, we should never, at least not in any time scale I can see currently solvable, put an AGI actually in charge of driving *anything*. As that's indeed when the optimisation - the alignment problem, as they call it, rears its very ugly head. (an AGI will never have the same philosophical world view as a human)

We haven't got AGIs yet... but that's a very italicised yet. That's that moment we need to really be very very cautious of.

related: I've found Rob Miles' videos on Computerphile very good on the subjects - he's a researcher into such things at Nottingham Uni.
I found his stuff very eye opening to things I hadn't even considered as "threats" in that sense - his research is a good reference of concerns that really need to be made more widely known. Most of his interviews tend to be a little technical, but not excludingly so. I think this is a good example of such worrying concepts:
https://www.youtube.com/watch?v=3TYT1QfdfsM

@'ingie I mean that they're all trained on American stuff.

5% of the world population.

So when they unleash the result on the other 95% of us, it's going to have no fucking idea what it's doing, because it doesn't look a damn thing like the training data.

Idea: a CAPTCHA reading "select all squares with idiots" and with images of Musk, Zuckerberg, Bezos etc.

Actually this is a misconception. By now AI are much better at passing reCapthchas than we are. How you might ask?

Because we taught them.

https://www.techradar.com/news/captcha-if-you-can-how-youve-been-training-ai-for-years-without-realising-it

fun fact, when you find the pictures that contain traffic lights that actually contributes to the training data to these self drivint cars. Thats why captchas used to be just words since google was working on image to text stuff.
There will be AI that correctly identifies traffic lights and ignores them by choice; like some humans do. ^^
👏🏾👏🏾👏🏾 perfection.
*snort* Perfect. No notes.
@weirdwriter I already find myself, as someone with a reputation for knowing about computers, being told more personal and financial details than I would like to help out friends who are blocked by the complexity of the process from, for example making charitable donations, or sponsoring someone for a good cause.
and always make your self driving cars obey the speed limits, easy to do… make them drive like old old people!
and now, AI-generated traffic lights 🫠
Just like how "the cloud" is somebody else's computer.

Very funny. Truly!

With that said, I think the point is that self driving cars will be here someday, but they are not here yet.
Or maybe they are? Supposedly, AI is better at captchas than humans now.

if self driving cars are inevitable, shouldn't AI be able to identify traffic lights? seems counterproductive.
I saw in a news article musk had to intervene on a live video when his twitter car was going to go through a red light 😂
@monkeyben yep, he was first in line at the intersection, the green arrow came on for opposing traffic to turn left and the car hit the gas
to be fair to Google re-captcha, those traffic lights contain special grainy noise patterns that throw off computer algorithms...
typical re-captcha challenge, many images, some contain pedestrian crossings, some do not.
same people making my thumbs do a massive amount of work for this experience.

@toplesstopics
M Knight Shyamalan twist:

We are all trapped in a simulation inside a self driving Ford Windstar 200 years in the future and the traffic captchas help them drive.

@DavBot or the tech bros remote control the car to drive you to the hit man location of their choice, as happens in many not-so-sci-fi movies I could name 😩
if they were able to take a step back they may feel the pain.
actually Teslas can identify traffic lights.
Just™️ install a transmitter at every traffic light
What could possibly go wrong 😱
The slowly dawning realisation that somewhere in San Francisco a "self driving" car is sat patiently waiting for you to complete a CAPTCHA :tiredcat2:
though there was recent paper showing AI is faster and better than humans in solving CAPTCHA, so I am not sure what's the point of it anymore

Worst part is the computer then considering it incorrect when the human says that a sign with an image of a traffic light is not actually a traffic light.

If that AI is going to be driving cars, I'll pass.

at intersections automated vehicles continue in between crossing traffic without stopping, all vehicles being aware of the others and keeping necessary distance. No lights needed.
@Erik Andersen Sounds like a really dystopian space for humans.
I know this is a funny, but given that your responses suggest you really believe it I should point out that the action of identifying them traffic light (or bridge, or motorcycle, or boat, or whatever) is not in itself sufficient to identify you to the recaptcha as human, it's the way in which you select the items, the movement of the cursor, the time taken to read the page etc.
Nice joke!! I actually learned recently why captcha or recaptcha got so simple, it's because it's not the result that matters 😂 It's the mouse path 🐁
Why do you think "techbros" are a homogenous amorphous mass with a single will? Do you even have a working definition of what a "techbro" is? This sounds like classical prejudice to me - invent a group, then ascribe all opinions by one member to the group as a whole.
@Stephan Schulz I wrote the software that built the chip that powers your phone. Sit down.
Trying to pull seniority on someone you don't know is not really a power move. Nor does it advance the discussion.

@Stephan Schulz "Nor does it advance the discussion", he said, nasally.

Nice fedora. Did your mum get it for you?

Whenever I encounter a captcha like that I imagine there’s somewhere just in this very moment a poor little self driving car needing my assistance.
With enough people solving captchas constantly we'll be able to use them to help cars solve ethical problems on the fly like should I swerve to avoid the small child and hit the old lady