Hacker News Reader

Relying on vision might be good for the advancement of computer vision, but is it good for automated driving?

It seems to be an unnecessary anthropomorphization to me. One of the greatest weaknesses of human drivers is the limited information available to act on. And there’s no possible scenario in which computer vision overcomes information unavailability issues. Fog or obstructions will always defeat it.

Absolutely. Replicating the way humans do it is also going to replicate the failure modes (optical illusions, crappy low light performance, etc.)

> Uber was very effective at galvanizing voters when local markets were threatening legislation via the app

By “very effective”, you mean just throwing around millions [1] of dollars to spam the local population to “galvanize” them to vote for/against certain laws?

Didn’t work in Austin, TX [2].

You are overestimating the power of what FSD can do. The unfortunate mistake most TSLA drivers do in the first place.

TSLA is playing catch up and burning massive amounts of cash. Laying off people left and right. The market and interest for TSLA is shrinking.

[1] https://www.austinchronicle.com/daily/news/2016-04-30/money-...

[2] https://www.businessinsider.com/what-happened-to-austin-texa...

What doesn't work in Austin works everywhere else.

I find it funny to think back to the fervor of comments on HN a year ago about Tesla doubling down on vision only and how this would prove to be dangerous and ineffective. A year has gone by and not only have we not seen the catastrophe that was thought to be inevitable, but Autopilot has gotten leaps and bounds better. I use it every day and over the past few months it's gotten so good that there are few scenarios where I ever intervene, and when I do, it's because it's always because it's being a little too slow/hesitant for me at an intersection. I'm not a Musk fan one bit, but I see this as evidence that the popular narrative has diverged wildly from the reality of Tesla's product.

Exactly. The technical argument about capabilities is largely settled at this point. Vision won, LiDAR didn't "lose", but it's clearly not any better. Yet every time the subject comes up we rehash the same arguments from 2020/21 as if the evidence of the last ~billion miles (or whatever) of road testing doesn't exist. The crashes didn't show up. The bodies aren't there. The disaster didn't happen. The product is "safe" for whatever reasonable definition you want to pick.

And to be clear, my sense is the same: FSD microdriving at this point is better than mine, interventions are about planning and intersection recognition. It'll be slow to take turns sometimes, it'll be in the wrong lane to prepare for a turn, stuff like that.

Vision didn't "win" in any way, shape, or form. The only products on the road that offer actual self driving don't rely on vision only, they have vision + lidar + other sensors.

Teslas are not "self driving", they don't allow you in any way to rely on the car alone, for any length of time, to drive itself: it's the drivers who are preventing accidents, not Tesla's FSD. When drivers of these vehicles lose attention, accidents do happen and people have died.

And exactly as predicted, these are all arguments that could have been made, verbatim, three years ago. And yet...

They are arguments that were made three years ago, and they are arguments that still hold up today.

But they didn't hold up. There are literally two orders of magnitude more vehicle-miles driven under FSD now than there were then, and there's absolutely no data showing a safety effect above the noise floor. Your argument is non-falsifiable nonsense, basically. You can continue making it forever, but that doesn't make it right.

There are exactly 0 miles of FSD driven by the public without constant attention from a driver keeping their hand on the wheel. There are disengagements of FSD in virtually every trip, where the driver feels that they need to intervene or the car would do something wrong. The safety record of FSD is basically due to the human drivers, not the AI. Tesla themselves don't have any confidence that the car can drive without constant human attention.

This is in contrast to Waymo and Mercedes, where the car actually drives itself, using both vision and other sensors, and the driver can actually disengage.

Again, non-falsifiable so I can't argue. Good trick. My car still drives me around every day and no one is crashing, so I guess everyone wins. I have an amazing supercar and you don't need to feel wrong about it.

Do you take your hands off the wheel? Do you take out a book and read it? If not, do you ever intervene on these trips?

What I'm saying is the official line from Tesla: they don't even claim to have achieved level 3, nevermind the level 4 driving you seem to think they have.

I repeat: I cannot falsify your contention. We will both go to our graves "not knowing" whether or not the driver is "really" the magic that makes the car drive itself. But only one of us is actually right.

Of course it's falsifible. I don't own a Tesla, and I wouldn't trust it even if I did, but I could check for myself by setting it to self-drive and then not paying any attention whatsoever. If I get in an accident, I was right.

More realistically, if Tesla believed their car is safe without human intervention, they could, you know, put them out on the streets without drivers, like Waymo did. Or at least tell their customers that it's actually safe to not pay full attention and that Tesla would accept liability if accidents happen, like Mercedes did. Or something in between. Since they are not doing any of these things, I think your opinion that is just obviously wrong.

> I could check for myself

This is disingenuous. And the reason is you clearly refuse to take testimony from those of us who do this every day. I'll tell you right now that it's been well over a year since I had any kind of safety-related disengagement. So, you'll change your mind?

Of course not. You're going to say that that's not enough data, and then go and cherry pick a single accident or testimony to the contrary to invalidate the data that you just said was a valid test.

Like I said, non-falsifiable.

I have seen this "huge improvement" comment about FSD just about in every discussion about FSD ever, since the first version. It keeps getting huge improvements, and it's still not done. At this point I expect version 579 of FSD will still be a "huge improvement" over 578, but just not 100% yet, and will soon be finished.

Each of Tesla's "huge improvements" seem to approximately halve the number of interventions. You can go from one intervention per minute to one per two minutes to one per four and so on. But you'll never hit zero.

There is no way in hell you have any kind of backing for the 50% halving number at each version. Prove me wrong.

It's remarkable how many comments on "hacker news" are basically "I like it, trust me it works great". Line noise.

LOL, I write a comment mentioning an intervention every 4 minutes and you interpret it as Tesla praise.

I've been seeing such comments for 4 years or more. Mercedes and Waymo actually have full self driving on the roads today, with regulatory approval and proven efficiency. Tesla has a nothing burger that disengage if you take your hands off the wheel.

Following a car in front of you in perfect conditions (otherwise it full-on disengages) just to comply with the regulatory labels is whack.

I don't get why people cheer on it and even compare it with Tesla, which goes for ... actual self driving

Because the Mercedes solution works, today, with legal guarantees, under quite precisely specified conditions. It allows you, in those specified and advertised conditions, to play a game or read a book or take a nap while driving.

In contrast, Tesla only has promises. In practice, you can't do anything with FSD that you couldn't do without it, because you have to pay exactly as much attention when it's engaged as when it's disengaged, and you can't take your hands off the wheel or it full-on disengages.

Mercedes has an excellent base to build up from. Tesla has next to nothing.

I’m not on the same boat, NTHSA has found Tesla’s Autopilot linked to increased accidents with “safety issues with Tesla Autopilot contributed to at least 467 collisions and 14 deaths from January 2018 through August 2023”. They have done a recalls and ar still investigating.

https://www.cnbc.com/2024/05/07/tesla-must-provide-nhtsa-wit...

> Tesla doubling down on vision only and how this would prove to be dangerous and ineffective. A year has gone by and not only have we not seen the catastrophe that was thought to be inevitable, but Autopilot has gotten leaps and bounds better.

Maybe it’s better in terms of customer satisfaction or reliability, but in terms of the SAR J3016 standard, it hasn’t improved at all.

It's wild to me that people argue automated driving is great based on individual personal experience.

If I were to argue about manual driving from my personal driving, I'd be way off. I haven't had an accident in decades. Not even any close calls. If I extrapolating my accident rate out to the world, I would assume that human-driven cars are a perfectly safe technology.

In the real world, all accidents are a long-tail phenomenon. They happen pretty rarely, often under conditions that are unusual or unexpected. So somebody saying "in these last few months I hardly ever have to intervene anymore to not die", a standard that would cause a driving instructor to fail a teen driver, not only doesn't give me confidence that this technology is a-ok for road use, but suggests to me they haven't really thought about what is required for safety at global scale.

It's perfectly reasonable for me to make a determination and offer my first hand experience about Autopilot's capabilities relative to that of my own and that of other drivers I see on the road.

Accidents might be long-tail phenomenon but driving habits are observable. Autopilot has a consistent, safe, alert demeanor on the road. It reacts to vehicles I as a driver can't as quickly see, changing situations that it would take me an extra half second or more to react to.

Compare that to the average human driver on the road - a third or more of whom you can expect to see holding a phone - and I think it's a reasonable conclusion to say it's already at least marginally better. And if we are talking about accidents, you can't deny that people are generally terrible at choosing the best possible reaction to an emergency situation that happens in the space of a few seconds.

It's reasonable for you to offer your firsthand experience. It's unreasonable for you to think that you can offer a determination broader than that experience.

For example, I might note that a couple weeks ago a Tesla almost crashed into a train: https://teslamotorsclub.com/tmc/threads/ap-fsd-related-crash...

The dashcam vids are sobering: https://www.dropbox.com/scl/fo/vxlz4vk3rck114145nbp8/AAukGDt...

Solving that problem on machine vision would be nice, but:

"Tesla bought over $2M worth of Lidar from Luminar (theverge.com)" [1]

[1]: https://news.ycombinator.com/item?id=40299991

$2mil is nothing worth of equipment, most likely just some testing or ground truth stuff. Borderline fake news or a lack of critical thinking imo.

Lidar probably makes a lot of sense of small scale accurate articulation of robotics where precision matters, could be for internal use too.

2m is nothing lol

Emphasis on "if it ends up working"

The Laconic "if"

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

not to mention much more scalable than waymos approach, both in terms of rollout, but more importantly in terms of data collection - the gold of the AI era.