Hotz: 3 Problems of Autonomous Driving

[SandiaGrunt]

You probably know a lot, but your attitude leaves a lot to be desired. I get the feeling you're bothered to talk to laymen, so why do you bother?

Because the Dunning-Kruger effect is funny. It's hilarious when amateurs try to teach me stuff about my own specialty, without any self-awareness of the limits of their own knowledge.

Maybe you've noticed that physicians also wear a Mona Lisa smile when their patients start with, "so I was looking at WebMD and I think I might have this condition called...."

 

[wooter]

I understand that. I'm not here to teach you anything; I only think your way of explaining things is crass. There's no reason to swear, is there?

Don't forget you only explained why you think Hotz is a doofus, and only vaguely went over what would be the correct course of action to solve autonomous driving. And you only give a glimpse about your speciality.

Now, I understand everyone likes to maintain a certain amount of anonymity on the internet so I'm not asking for your LinkedIn page to check your credentials, but at the moment nobody has really a clue about what credentials you really have.

For me it's easy: I'm a proper layman and am just interested in the matter, without any actual knowledge to share. I did listen to the interview a few months ago and was entertained, but as a layman I have no clue to actually determine if he's an expert or just vaguely familiar with the subject and just a good people person.

Which leads me back to my first point: maybe you don't want to explain your knowledge to anonymous amateurs on the internet, and maybe you should then just better stop and go do something much more productive, fulfilling or relaxing?

Personally I like to explain my specific expertise to laymen and see how their gears turn and make them click. But my expertise is no in autonomous driving. But I do know being crass and swearing doesn't get your point across.

 

[SandiaGrunt]

but at the moment nobody has really a clue about what credentials you really have

Good.

maybe you don't want to explain your knowledge to anonymous amateurs on the internet, and maybe you should then just better stop and go do something much more productive, fulfilling or relaxing?

I have a productive, fulfilling life, though admittedly not very relaxed. I come here and argue with idiots like @diplomat33 because it's fun.

But I do know being crass and swearing doesn't get your point across.

I'm never going to succeed in teaching hundreds of thousands of doofuses why, e.g. lidar is actually important, anyway.

 

[wooter]

I come here and argue with idiots like @diplomat33 because it's fun.

Yeah, that's the end of the ride for me here.

 

[diplomat33]

I understand that. I'm not here to teach you anything; I only think your way of explaining things is crass. There's no reason to swear, is there?

Don't forget you only explained why you think Hotz is a doofus, and only vaguely went over what would be the correct course of action to solve autonomous driving. And you only give a glimpse about your speciality.

Now, I understand everyone likes to maintain a certain amount of anonymity on the internet so I'm not asking for your LinkedIn page to check your credentials, but at the moment nobody has really a clue about what credentials you really have.

For me it's easy: I'm a proper layman and am just interested in the matter, without any actual knowledge to share. I did listen to the interview a few months ago and was entertained, but as a layman I have no clue to actually determine if he's an expert or just vaguely familiar with the subject and just a good people person.

Which leads me back to my first point: maybe you don't want to explain your knowledge to anonymous amateurs on the internet, and maybe you should then just better stop and go do something much more productive, fulfilling or relaxing?

Personally I like to explain my specific expertise to laymen and see how their gears turn and make them click. But my expertise is no in autonomous driving. But I do know being crass and swearing doesn't get your point across.

Don't waste your time with @SandiaGrunt . He is not here to actually have a constructive discussion with anybody. He has no interest in sharing his expertise. He has made it clear that he is only on this forum to attack and make fun of us.

Case in point:

I come here and argue with idiots like @diplomat33 because it's fun.

 

[wooter]

Wow, first time I used the ignore feature and I can highly recommend it. It's like I wiped him of the face of this site. I can't even see whatever I quoted from him in my own posts!

Oh, and @SandiaGrunt, if you read this... Your digital exhaust matters. Your style was completely off-putting and arguing and trolling is not fun. I choose not to go along in such practices, even if you might be knowledgeable.

@diplomat33, may I suggest you also use this ignore function? I can't tell if you are an actual idiot, but nobody needs this.

 

[diplomat33]

@diplomat33, may I suggest you also use this ignore function? I can't tell if you are an actual idiot, but nobody needs this.

Thanks for the suggestion.

 

[electronblue]

Don't waste your time with @SandiaGrunt . He is not here to actually have a constructive discussion with anybody. He has no interest in sharing his expertise. He has made it clear that he is only on this forum to attack and make fun of us.

Case in point:

I’m not sure that is fair either. @SandiaGrunt seems to offer an educated opinion on what won’t work.

In every sense of the word.

A non-working debate style included.

Yeah, a bit too acidic for my taste as well...

 

[Southpasfan]

I'm not sure what "counterfactual" driving is exactly. I should probably re-listen to the podcast, which was excellent.

There is nothing like owning a Tesla for a few months to make a person really think about driving.

Driving is:

1. Controlling the car. As far as I can see, Tesla (and other carmakers) have this solved. It is not as if the auto controls fail to control the car. By "fail" for example, it is not as if you set the speed for 50mph and the car wobbles between 50 and 70 mph.

2. Controlling the speed. Same as 1.

3. Keeping the car oriented in (i) its lane, or (ii) to the extent there is not a clear lane, on the correct side of the road. Teslas are fantastic at (i), and I have seen improvements in (ii) since I bought the car in June. This is the first area where you need perception. You need to have the car be able to recognize the lane at lease as well as, if not better than, two human eyes. Too many people, from what I read, really discount how far along Teslas are on (i) and (ii). Item (ii) may not be perfect yet, but it seems obvious to me that if you can solve for (i) you can solve for (ii). I don't quite see how the naysayers can argue this won't happen, when it is actually happening.

4. Avoiding moving cars. Regardless of whether you are in a lane or not, you need to avoid other cars. Again, in the context of cars going in the same direction on a road, Tesla may be better than humans at this already. Its only a feeling, because without a deep dive into data to separate accidents generally from accidents caused by not getting out of the way of another moving car, its not provable. But given NAP's ability to change lanes, which involves waiting for the lane to be open, its hard to argue that Tesla has not solved for this as well. Hotz argues in the podcast that Tesla only has "one kind of lane change" - meaning that Teslas do not do the type of lane change where you speed up to get into the lane ahead. That's true, but that begs the question of whether that type of lane change is even needed, notwithstanding the fact that humans do it all the time, sometimes very aggressively. I believe that it may be needed for lane changes where there is simply so much traffic that waiting is not practical or even unsafe. Right now, with the combination of front cameras and radar, it seems to me that Teslas are far, far better than humans at not rear-ending the car in front of you.

I would note that last software release also shows cars coming the other way -- this is how a left and right turn can be done. Tesla already can identify the lane/road. I successful left turn involves that plus cars coming the other way. It also involves calculating safe distances.

5. Avoiding stationary objects, including people. This is why the release of smart summon was such a big deal. Telsas already have automatic emergency braking. But who would actually "test" it? I have noticed in the last software release a great improvement in warning when the car ahead is slowing down and I am not slowing with it. But smart summon attempts to avoid objects and people, on its own, and does so. It does it very slowly. But why would anyone assume its impossible to do it better? The cameras are better positioned than human eye level, and, of course, there are more of them. But the ability to identify and avoid objects is now officially a feature.

6. Process objects, such as stop lights, stop signs, etc. This is next. Teslas may be able to indentify, for example, a temporary stop sign placed between lanes when the traffic light is out of order. But other than the non-released developer software, current software does not attempt to distinguish a "sign" from, say, a trash can. In addition, the type of object (cone v. newspaper) dictates the response.

Five of the six features are already released.

So I guess I don't know that "counterfactual" driving is an actual feature. If you can do all six things above with the flawlessness of a computer, do you need the same amount of anticipation which humans have? Or do we just need the anticipation because we do not have the same level of perception of 8 cameras plus the sensors?

 

[alsetym]

Apple, Google, Movidius, MobilEye, Intel, and NVIDIA

Apple and google for like, what...2 years? The only NVIDIA and MobilEye product I'm aware of in consumer products are in cars and only for the last couple of years (I'm not sure what the original Mobile EyeQ chips were but I'm pretty sure it was only tesla rocking those until more recently). Movidius is a bit longer but I've never even seen any of their products...So I call BS on being late and double BS on being 8 years late. And most of those (apple, google, and most NVIDIA stuff) are still solution agnostic hardware which is not what we are talking about. HW3 is much more narrowly focused, as is the MobileEye hardware per my understanding. Tesla isn't at the bleeding edge, but they aren't late.

PGAs are never faster or more power-efficient than the same logical function in custom silicon. Do you know anything at all about FPGAs?...Are you not aware that FPGAs literally use compilers to translate HDL into the bitstream written into the FPGA's SRAM?

In most of your replies you are too narrowly focused for a future looking discussion. FPGAs are obviously not going to have advantages over the exact same silicone implementation. BUT being able to change your 'silicone' implementation as you go and/or over time to always be inline with what will best solve whatever problem you are looking at...now we have something interesting. If HW3 was a giant FPGA it would give them an opportunity to tweak the hardware with each software release, keeping them in perfect synch. That synchronization could eliminate some of the loss in efficiency compared to a silicone. Not my area, not sure if the numbers would ever actually work out (even on future FPGA hardware) but if the hardware and the software could both support it...Heck, imagine hardware that adapted to the need on every tick. That tick box does not exist. Under the right circumstances, and with efficient FPGA switching array, maybe you can economize at a system level...only solve the problem at hand in the best possible way. You want to talk about why imitation learning on computers sucks so badly? Well, the organic version updates the hardware and the software during that learning process, so its worth considering what the consequences of fixed hardware are. Its at least interesting as a thought experiment, for me anyway. By learning why it won't work you learn about what might work.

I still consider FPGA's fringe for any consumer/front end device. I actually don't know of any, I'm sure there are (I didn't even google this), but they are not common. Using them for development before you print silicon has nothing to do with the present conversation.

And you're talking to 'me' because at some point I might say something that causes your head to pop out of your arse and you might have an original thought roll out from between your ears.

 

[alsetym]

I'm not sure what "counterfactual" driving is exactly. I should probably re-listen to the podcast, which was excellent.

There is nothing like owning a Tesla for a few months to make a person really think about driving.

Driving is:

1. Controlling the car. As far as I can see, Tesla (and other carmakers) have this solved. It is not as if the auto controls fail to control the car. By "fail" for example, it is not as if you set the speed for 50mph and the car wobbles between 50 and 70 mph.

2. Controlling the speed. Same as 1.

3. Keeping the car oriented in (i) its lane, or (ii) to the extent there is not a clear lane, on the correct side of the road. Teslas are fantastic at (i), and I have seen improvements in (ii) since I bought the car in June. This is the first area where you need perception. You need to have the car be able to recognize the lane at lease as well as, if not better than, two human eyes. Too many people, from what I read, really discount how far along Teslas are on (i) and (ii). Item (ii) may not be perfect yet, but it seems obvious to me that if you can solve for (i) you can solve for (ii). I don't quite see how the naysayers can argue this won't happen, when it is actually happening.

4. Avoiding moving cars. Regardless of whether you are in a lane or not, you need to avoid other cars. Again, in the context of cars going in the same direction on a road, Tesla may be better than humans at this already. Its only a feeling, because without a deep dive into data to separate accidents generally from accidents caused by not getting out of the way of another moving car, its not provable. But given NAP's ability to change lanes, which involves waiting for the lane to be open, its hard to argue that Tesla has not solved for this as well. Hotz argues in the podcast that Tesla only has "one kind of lane change" - meaning that Teslas do not do the type of lane change where you speed up to get into the lane ahead. That's true, but that begs the question of whether that type of lane change is even needed, notwithstanding the fact that humans do it all the time, sometimes very aggressively. I believe that it may be needed for lane changes where there is simply so much traffic that waiting is not practical or even unsafe. Right now, with the combination of front cameras and radar, it seems to me that Teslas are far, far better than humans at not rear-ending the car in front of you.

I would note that last software release also shows cars coming the other way -- this is how a left and right turn can be done. Tesla already can identify the lane/road. I successful left turn involves that plus cars coming the other way. It also involves calculating safe distances.

5. Avoiding stationary objects, including people. This is why the release of smart summon was such a big deal. Telsas already have automatic emergency braking. But who would actually "test" it? I have noticed in the last software release a great improvement in warning when the car ahead is slowing down and I am not slowing with it. But smart summon attempts to avoid objects and people, on its own, and does so. It does it very slowly. But why would anyone assume its impossible to do it better? The cameras are better positioned than human eye level, and, of course, there are more of them. But the ability to identify and avoid objects is now officially a feature.

6. Process objects, such as stop lights, stop signs, etc. This is next. Teslas may be able to indentify, for example, a temporary stop sign placed between lanes when the traffic light is out of order. But other than the non-released developer software, current software does not attempt to distinguish a "sign" from, say, a trash can. In addition, the type of object (cone v. newspaper) dictates the response.

Five of the six features are already released.

So I guess I don't know that "counterfactual" driving is an actual feature. If you can do all six things above with the flawlessness of a computer, do you need the same amount of anticipation which humans have? Or do we just need the anticipation because we do not have the same level of perception of 8 cameras plus the sensors?

I think the lack of counterfactual is a large part of what makes smart summon look so ungainly (well, other than the fact its not always sure what something is and if it is moving...) and will ultimately be a choking point. If you can't figure out where other drivers want to go, it makes any non-standard even very difficult. You can't just hard code in EVERY scenario, it's not plausible.

However, without counterfactual, it's probably possible to make something that gets you point A to point B eventually and safety. People just won't accept it and cars will get stuck at unacceptable rates.

 

[LowlyOilBurner]

I like diplomat33’s post’s/threads. They are optimistic and interesting.

 

[SandiaGrunt]

BUT being able to change your 'silicone' implementation as you go

Imagine you’re a medical doctor — a head surgeon at major hospital. You invented a new surgical technique, and you’ve worked on it tirelessly for nine years. Tens of thousands of younger doctors are now using your work, all around the world. What used to be a tiny, almost irrelevant specialty is now front-page news. Every major news network and magazine is reporting on it. There are subreddits and entire forums bursting with people who are ostensibly interested in the subject. People are getting tattoos celebrating it.

Every now and then, you decide to join the conversation. Maybe it’ll be fun? You’re hopeful at first, though maybe a bit too eager to steer the conversation your way, because you’re used to doing that. In your professional life, people pay you lots of money just to hear what you have to say.

But somehow, online, you always end up in a random argument with an amateur who invariably tries to teach you about your very own work. This person looks you right in the eye, so to speak, and tells you a bunch of falsehoods and misconceptions. This person gets adamant and pushy and demands that you respect them, then tells you you have your head up your arse.

But this person is so ****ing ignorant that they cannot even spell the word “surgery!”

This seems to be the reality of being an expert on social media today. As soon as people figure out you’re an expert, a mob forms to try to tear you down. It’s gone beyond the “democratization of knowledge,” beyond the “death of expertise,” all the way to outright hostility towards anyone claiming greater knowledge of any subject.

Today, everyone on social media — no matter their actual qualifications or experience — looks around for opportunities to teach. They can’t stand the passivity of merely learning, or the shame of admitting the limits of their own knowledge. They need the dopamine rush of not just impersonating an expert, but the sublime satisfaction of believing they got the better of one.

 

[electronblue]

I'm not sure what "counterfactual" driving is exactly. I should probably re-listen to the podcast, which was excellent.

Looking beyond the Tesla world, this is actually a hot debate. One of the reasons why Waymo has issues with certain situations is that they have so far been driving fairly ”factually” and carefully. For example MobilEye pointed out their own robotaxi prototypes in Israel had to be made a lot more aggressive and perhaps creative about ignoring some free space limitations just to survive the local traffic culture.

Still, credit where credit is due: Waymo is driving passenges out there while many others are still mostly talking. (There are smaller scale robotaxi/robobus projects in other countries too of course.)

There is nothing like owning a Tesla for a few months to make a person really think about driving.

...and nothing like owning Tesla for a few years to make a person really think in general, as many find out.

 

[EVNow]

I have a productive, fulfilling life, though admittedly not very relaxed. I come here and argue with idiots like @diplomat33 because it's fun.

Reported. I suggest others to report this guy as well.

 

[SandiaGrunt]

Reported. I suggest others to report this guy as well.

Look! Look everyone! There’s an expert in our midst who’s completely exhausted by people literally arguing with him for twelve pages over basic statistics! He said some rude stuff because he gets exasperated by the actions of all these other people, who all eventually admit they were wrong, anyway! Let’s form a mob! Let’s gang up on him! Protect our echo chamber! Protect our overconfident amateurs and their right to post endless falsehoods! RUN THE EXPERT OUT ON RAILS, NOW!

 

[EVNow]

Look! Look everyone! There’s an expert in our midst who’s completely exhausted by people literally arguing with him for twelve pages over basic statistics! He said some rude stuff because he gets exasperated by the actions of all these other people, who all eventually admit they were wrong, anyway! Let’s form a mob! Let’s gang up on him! Protect our echo chamber! Protect our overconfident amateurs and their right to post endless falsehoods! RUN THE EXPERT OUT ON RAILS, NOW!

Dude, people like you with highly inflated egos are dime a dozen in SV. You aren't doing any of us great favors. If you can't adhere to TMC terms, we don't need to listen to your cr*p.

 

[SandiaGrunt]

Dude, people like you with highly inflated egos are dime a dozen in SV. You aren't doing any of us great favors. If you can't adhere to TMC terms, we don't need to listen to your cr*p.

Jesus, the irony is so strong. I’m one of the most knowledgeable people who has ever posted here, but you’re now going to discredit me by saying I’m “dime a dozen.” Another poster told me I have my head up my arse. Other people have argued with me for literally twelve pages about basic statistics.

You have created a environment here where it’s okay for a freshman student to waste a professor’s time for hours. An environment where it’s okay for a kid who speaks volumes of falsehoods about a subject to tell a professor to his face that he has his head up his arse.

And if the professor gets snippy about it? Call the forum nanny to kick his ass out!

Do you realize how delusional, how counterproductive, how stupid that attitude is?

Maybe the greatest irony in this layer cake of irony is that Elon Musk is one of the snarkiest, most aggressive, most egoistic people in the history of tech, but most people here excuse him, and many people exalt him for it. You can’t even get your opinions straight.

 

[enixneo]

I’m one of the most knowledgeable people who has ever posted here.

You have created a environment here where it’s okay for a freshman student to waste a professor’s time for hours. An environment where it’s okay for a kid who speaks volumes of falsehoods about a subject to tell a professor to his face that he has his head up his arse.

Here's the problem SandiaGrunt, no one here knows who you are. In the real world you may or may not be an expert in some relevant field. But here you are just a random forum user like everyone else. An anonymous voice. In your case, an anonymous voice throwing a tantrum because people don't respect your unvetted years of experience.

 

[SandiaGrunt]

Here's the problem SandiaGrunt, no one here knows who you are. In the real world you may or may not be an expert in some relevant field. But here you are just a random forum user like everyone else. An anonymous voice. In your case, an anonymous voice throwing a tantrum because people don't respect your unvetted years of experience.

You are correct! And it’s not worthwhile to identity myself here, because there are NDAs and lots of money involved.

But how do you justify 12 pages of debate with @diplomat33 about basic statistics? How do you explain @aselytm saying I have my head up my ass because I said “FPGAs are 4-10x less power-efficient than ASICS?”

I think the reality is that social media is impossible for me. There’s no benefit to me, and it’s frustrating as hell.

I do enjoy seeing the public excitement around the technology. But the misconceptions and falsehoods and confusions just drive me crazy.