Elon Musk is a nice chap

[james_s]

My only disagreement comes because I suspect that few drivers of non-automated cars are 100% effective.  They fail on edge cases too, and also mainstream cases because of boredom, distraction , medical issues and a host of other problems.  In an ideal world the automated driving system doesn't have to be perfect just better than human drivers.  These systems have an advantage.  With no ego they should have no problem pulling safely to the side and stopping when it is raining too hard, too foggy, too slick or any number other problems where human drivers readily proceed beyond their skills

Calling the current automated systems "not 100% effective" is extremely generous, right now I would say they're at the level of "15 year old who just got a learner permit and has had a cumulative total experience of 2 hours behind the wheel." They are still making rookie mistakes left and right, and it's going to take a massive leap in AI tech to get it anywhere close to where it does not need constant supervision. Then there is the issue of liability, if a human driver has a habit of breaking the law or causing accidents we punish them or revoke their driving privileges. If the manufacture is liable for accidents there will be enough of those happening every day to completely sink a company that has widespread deployment of these vehicles. Yes human drivers readily proceed beyond their skills, but the autonomous systems are showing a habit of doing this as well. How many cases have we seen already of Teslas on autopilot slamming into the back of stopped emergency vehicles at highway speed, turning into an oncoming semi, smashing into a highway barrier, and how many other accidents would have taken place had a human not managed to jump in and intervene? They don't stop and pull over safely when they are outside of their capabilities, they just disengage and toss the ball to the human right then and there.

[james_s]

Remember, Tesla FSD is called Full Self Driving, but that's in name only. It's not actually full self driving. It's actual name is AutoPilot, and that's what it officially is, and advanced autopilot that can lane change and brake etc. And it has a cute auto-summon mode where it find you in a carpark or something. They are very clear about this on their website. It's Musk that has been personally talking up "Full Self Driving" and fully autonymos for years now, or at least implying it. There are countless FSD beta video out there showing how dumb the system is in even basic almost ideal condition scenarios.
He's either had a technical awakening or the suits have tapped him on the shoulder to calm it down, they aren't even remotely close to full autonymous self self driving which Musk now has essentially admitted.

Publicly calling it "Full Self Driving" was a monumental mistake, it is not an exaggeration, it is an outright misrepresentation and there will continue to be people who abuse it and lawsuits will result. They can warn people all they want that "Full Self Driving" cannot actually drive itself, but that is not going to stop people from doing stupid things and claiming "It's called Full Self Driving!"

[james_s]

These are placed to let autonomous vehicles know their position. In addition to that they are also building a digital infrastructure to support vehicle to vehicle communication.

They've been there for a week without being vandalized by graffiti taggers? Don't get me wrong, I loath graffiti and would be perfectly ok with drastic measures to remove those who do it from society, but that doesn't change the fact that it is rampant. On top of the idiots with spraypaint there are other idiots slapping political and other stickers onto signs. I would honestly be shocked if a sign like that placed near almost any major city in the USA didn't get vandalized in some way within weeks.

[bdunham7]

right now I would say they're at the level of "15 year old who just got a learner permit and has had a cumulative total experience of 2 hours behind the wheel.

Actually I think they are more like a wide-awake drunk driver that is trying to fake it by staying in the lane, etc but then occasionally makes a huge mistake or can't react properly to something.  Perhaps self-driving systems could be rated by how they compare with drunks with certain blood alcohol levels--the eBAC rating.  NHTSA could set up a proving ground with obstacles and tests and volunteers can get drunk and then compete against the FSD vehicles.  I think Tesla is probably blowing at least a 0.15-0.20 at this point. 

[nctnico]

right now I would say they're at the level of "15 year old who just got a learner permit and has had a cumulative total experience of 2 hours behind the wheel.

Actually I think they are more like a wide-awake drunk driver that is trying to fake it by staying in the lane, etc but then occasionally makes a huge mistake or can't react properly to something.  Perhaps self-driving systems could be rated by how they compare with drunks with certain blood alcohol levels--the eBAC rating.  NHTSA could set up a proving ground with obstacles and tests and volunteers can get drunk and then compete against the FSD vehicles.  I think Tesla is probably blowing at least a 0.15-0.20 at this point.

There is an upside to it: everytime I notice a Tesla car doing something really stupid (like I witnessed on my trip to the swimming pool today) I blame Musk's auto-pilot / FSD (Fully Sober and yet Drunk) and put less blame on the driver.

[bson]

They probably will have to be at least an order of magnitude better than humans and not make any of certain classes of errors in order to be widely accepted.

As an autonomous autopilot yes, but certainly not as a driver assist.  It's exceedingly useful to have a vehicle that is completely incapable or running into a pedestrian in a crosswalk, or drift off into a cyclist on the shoulder.  Or with some added instrumentation like beacons, incapable of rolling through a stop sign, running a red light, will always stop at yellow lights when possible, will warn when trying to stop in a no-stop zone, will always properly yield in intersections, and prevents many other thoughtless and dangerous behaviors or outcomes.  When the emergency flashers are activated the vehicle can emit a beacon that makes it practically impossible for another vehicle to collide with it or drive too close (where it might hit an open door or injured person).  Bicyclists could mount beacons (similar to existing radars like the Garmin Varia RTL515) that no matter how stupid or illegal they are it's impossible for a car to hit them - it will simply stop when on a collision course.  This is absolutely the only way we can make any progress towards a zero goal.  It's also IMO a whole lot more immediately useful than cars that drive themselves, that's just a novelty.

[BrianHG]

I'm pretty sure this point is known and acknowledged as a challenge. 10:27, 13:09. Sandy is blaming poor road markings for difficulties but Elon is asserting that an autonomous driving system must remain safe under any condition whatsoever.

Elon is slowly coming to realise this, and it's why he recently tweeted that, essentially, true autonomous cars will require a dramatic leap in AI (that we don't currently posses).

  And how I was railed at in the other thread saying that Elon's Full Self Driving required a ton more processing and added layers to it's capabilities.  Yes it can be done.  No, currently it would require too much neuronet processing with specialist sections dealing in not just the current generic shape (including signs, humans, vehicles, bikers, ect..) and geometric representation of what the cameras see with motion tracking, but a complete set layers dealing with how to interpret what is happening and predictions on what the environments should look like over time as you move, including items which are visually damaged, hidden by obstructions or missing all together.  In isolated circumstances, many of these functions already do exist as seen on Dr. Károly Zsolnai-Fehér YouTube A.I. channel 'Two Minute Papers'.  The problem is integrating all those specialized functions + a neuronet to link them together, and properly train everything.  Today, this would make an FSD AI computer consume a few kilowatts and be the size of 2-3 huge PCs.

  This figure will improve and we will get there as the latest 2nm IC eventually filter down to specialized dedicated ASICS like the Tesla FSD chip, though, you would still need a few of them.  There is too much money in FSD to just drop the ball, so it is guaranteed to eventually happen.

  Anyone who designs a FSD relying on future improvements/changes to the road environment for their FSD to function properly will be superseded by the the corporation which develops a complex and robust enough AI to drive as 'competently' as a human who is paying proper attention to the road at all times without fatigue or disruption, even if the AI takes a few hundred watts to operate in the beginning.

(Adding Lidar on-top of the camera system is a big plus, but the FSD should be able to operate on sight alone...)

[nctnico]

  Anyone who designs a FSD relying on future improvements/changes to the road environment for their FSD to function properly will be superseded by the the corporation which develops a complex and robust enough AI to drive as 'competently' as a human who is paying proper attention to the road at all times without fatigue or disruption, even if the AI takes a few hundred watts to operate in the beginning.

Don't think so. The FSD which relies on infrastructure will be cheaper and useful sooner. In the end simplicity wins and in this case simplicity is achieved by adapting the road.

[BrianHG]

  Anyone who designs a FSD relying on future improvements/changes to the road environment for their FSD to function properly will be superseded by the the corporation which develops a complex and robust enough AI to drive as 'competently' as a human who is paying proper attention to the road at all times without fatigue or disruption, even if the AI takes a few hundred watts to operate in the beginning.

Don't think so. The FSD which relies on infrastructure will be cheaper and useful sooner. In the end simplicity wins and in this case simplicity is achieved by adapting the road.

Which FSD tech will be ready first?  The one which requires every road, suburb road, in all countries, all of them to have an infrastructure set-up, maintained and regulated across all states and countries, and you need to get investment to handle these logistics and government regulation?  Or, a company getting investment to make a vehicle which can just do it on it's own?  A vehicle which can operate just like a taxi or delivery vehicle anywhere in any state/country on any road anywhere on earth?  What about country roads or dirt roads or emergency repair crews?  The technical prowess and value to investment to make the later work does exists and it will quickly make such specialized adaptation of every road in every country obsolete.

[bdunham7]

As an autonomous autopilot yes, but certainly not as a driver assist.  ....    It's also IMO a whole lot more immediately useful than cars that drive themselves, that's just a novelty.

I agree and so do some car companies, but you can't sell those systems for $10,000 and people won't show them off on social media, so no free advertising. 

[bdunham7]

The technical prowess and value to investment to make the later work does exists and it will quickly make such specialized adaptation of every road in every country obsolete.

That's yet to be proven on both counts.

[bdunham7]

Don't think so. The FSD which relies on infrastructure will be cheaper and useful sooner. In the end simplicity wins and in this case simplicity is achieved by adapting the road.

Consider this accident:



This is not an uncommon situation and I've seen similar myself.  You can see the truck in front is slow and has its hazard blinkers on while the truck involved has its turn signal on and is closing on the front one.  That truck may technically be at fault, but I would also consider the driver of the car to have performed poorly.  I certainly would not have gotten into this accident myself and would expect most sober drivers would be able to avoid it as well.  Even after the truck moves over, a single hard brake application would have bailed the Model 3 out--no need to go flying off the road.  So the Model 3 Autopilot gets at least a 0.10 - 0.15 eBAC rating here.  What would it take to modify the environment--road, other vehicles, ,  to prevent this?  What would it take to modify the Autopilot to be able to see the turn signals and relative motion of the vehicles and predict what they are going to do in a reasonably sophisticated way, similar to the way an attentive human would?  I think both of those are pretty high hurdles that we are nowhere near overcoming.

Perhaps current tech is as close to comprehensive FSD as Zoltar is to human-level AI.

https://zoltar.org/

[james_s]

They probably will have to be at least an order of magnitude better than humans and not make any of certain classes of errors in order to be widely accepted.

As an autonomous autopilot yes, but certainly not as a driver assist.  It's exceedingly useful to have a vehicle that is completely incapable or running into a pedestrian in a crosswalk, or drift off into a cyclist on the shoulder.  Or with some added instrumentation like beacons, incapable of rolling through a stop sign, running a red light, will always stop at yellow lights when possible, will warn when trying to stop in a no-stop zone, will always properly yield in intersections, and prevents many other thoughtless and dangerous behaviors or outcomes.  When the emergency flashers are activated the vehicle can emit a beacon that makes it practically impossible for another vehicle to collide with it or drive too close (where it might hit an open door or injured person).  Bicyclists could mount beacons (similar to existing radars like the Garmin Varia RTL515) that no matter how stupid or illegal they are it's impossible for a car to hit them - it will simply stop when on a collision course.  This is absolutely the only way we can make any progress towards a zero goal.  It's also IMO a whole lot more immediately useful than cars that drive themselves, that's just a novelty.

Frankly I think it makes the situation worse rather than improving it. All this automation does is act as a crutch which enables people to be less engaged and pay less attention while driving. It's like giving a student a tool that does their homework for them. Then on top of that if they ever drive a vehicle that doesn't have those features they are accustomed to then they are even more of a menace. Personally I would rather cars had no automation at all, and no automatic gearboxes, force people to use both hands and feet to drive the car so they can't easily focus on other things while they're supposed to be watching where they go.

[BrianHG]

The technical prowess and value to investment to make the later work does exists and it will quickly make such specialized adaptation of every road in every country obsolete.

That's yet to be proven on both counts.

Please don't think along the lines of my parents watching me play a video game back in 1982 on my 48 kilobyte Atari 800 believing this is all computers could and would ever achieve.  IBM/Intel and others are developing high density neuronet IC which operate in the analog domain as well as new research in the optical processors which would take my so-called multi-hundred watts specialized processor down to a tiny device consuming a few watts which will run circles around it.

If you think we reached an end apex, I am here to tell you we have just begun and we will progress further.

Today's AI is only at such a level of infancy that what we have today is like an ant compared to what's coming down the line.

[bdunham7]

If you think we reached an end apex, I am here to tell you we have just begun and we will progress further.

Today's AI is only at such a level of infancy that what we have today is like an ant compared to what's coming down the line.

I once had the opportunity to meet and attend a talk by famed physicist Arno Penzias, who by that time was older and more of a scientific philosopher and had taken up the topic of artificial intelligence.  This was in the mid-80s.  He said something interesting which I've never forgotten.  He brought up a supposed statement by Albert Einstein regarding computers, apparently made in the ENIAC days, that in the future there would be dozens or even hundreds of computers, they would each fit in a single room and that they would control almost every aspect of our lives, like some benevolent (or not) overlords.  Obviously this prediction was wrong in two ways--computers got a lot smaller and a lot faster than predicted but didn't control our lives, at least not at that point.

The point of all that, he said, was that in regards to technological advance, we tend to underestimate the degree that technology will advance while simultaneously overestimating the effect those advances will have.

If I compare using Microsoft Word on a PC in 1991 with using Office 365 on a PC in 2021, the technological differences are astounding.  The modern PC will have a thousand times the memory and I don't know how many times the processor power.  Office 365 is connected online and has a myriad of features that were only hinted at if even that in 1991.  And yet the way I use it the most has hardly changed at all. 

With AI-type applications, there's a similar phenomenon that I think can be explained by repurposing Penzias' statement.  With AI, we may underestimate the advances in technology, speed/memory/neural nets and so forth, and simultaneously overestimate how effective those advances will be in solving our problem.  The advances may be stupendous, but what is still unknown is how hard the task is to accomplish.  Computers have only fairly recently (20 years, OK, recent by some standards) gotten to the point of being equivalent to humans at chess, which is really a game much better suited to a CPU than a human brain.  And that was only accomplished with a very expensive many-year effort and involved a computer that was the size of a small car and used a lot more than a few hundred watts.  Nobody knows how hard it might be to make a car FSD with only camera images.   I suspect it is going to require a lot more specific programming and a lot less machine learning, but we'll see.  Quite a few companies that supposedly can hire the best talent have been working on this for years and frankly I don't think any of them have made any real progress, just show-time crap that is more dangerous than helpful.

Consider chat-bots that companies use for customer service.  Can any of them pass Turing muster?  Can they go three back-and-forth responses without the user realizing they are a bot?  Isn't there a huge financial incentive to develop a chat-bot that could replace all the people in a call center?  The gigabytes and gigaflops are there, the money is there, but no solution has been found so far.

[nctnico]

  Anyone who designs a FSD relying on future improvements/changes to the road environment for their FSD to function properly will be superseded by the the corporation which develops a complex and robust enough AI to drive as 'competently' as a human who is paying proper attention to the road at all times without fatigue or disruption, even if the AI takes a few hundred watts to operate in the beginning.

Don't think so. The FSD which relies on infrastructure will be cheaper and useful sooner. In the end simplicity wins and in this case simplicity is achieved by adapting the road.

Which FSD tech will be ready first?  The one which requires every road, suburb road, in all countries, all of them to have an infrastructure set-up, maintained and regulated across all states and countries, and you need to get investment to handle these logistics and government regulation?  Or, a company getting investment to make a vehicle which can just do it on it's own?

The first. Investments & research in smart roads pre-date self driving cars and also support non-self driving cars. IOW: there is a much bigger market & benefit from smart roads than just self driving cars.

[sandalcandal]

  And how I was railed at in the other thread saying that Elon's Full Self Driving required a ton more processing and added layers to it's capabilities.  Yes it can be done.  No, currently it would require too much neuronet processing with specialist sections dealing in not just the current generic shape (including signs, humans, vehicles, bikers, ect..) and geometric representation of what the cameras see with motion tracking, but a complete set layers dealing with how to interpret what is happening and predictions on what the environments should look like over time as you move, including items which are visually damaged, hidden by obstructions or missing all together.  In isolated circumstances, many of these functions already do exist as seen on Dr. Károly Zsolnai-Fehér YouTube A.I. channel 'Two Minute Papers'.  The problem is integrating all those specialized functions + a neuronet to link them together, and properly train everything.  Today, this would make an FSD AI computer consume a few kilowatts and be the size of 2-3 huge PCs.

  This figure will improve and we will get there as the latest 2nm IC eventually filter down to specialized dedicated ASICS like the Tesla FSD chip, though, you would still need a few of them.  There is too much money in FSD to just drop the ball, so it is guaranteed to eventually happen.

  Anyone who designs a FSD relying on future improvements/changes to the road environment for their FSD to function properly will be superseded by the the corporation which develops a complex and robust enough AI to drive as 'competently' as a human who is paying proper attention to the road at all times without fatigue or disruption, even if the AI takes a few hundred watts to operate in the beginning.

(Adding Lidar on-top of the camera system is a big plus, but the FSD should be able to operate on sight alone...)

AFAIK actually running a NN is extremely computationally cheap (see ML, MV and NN algorithms running on 8-bit Arduinos). The computationally intensive part is doing the training. So you wouldn't need a several kilowatt computer to actually run an FSD AI, but it would be very helpful for training it. Apparently that's a common misconception when talking about processing power required to do ML. (This is according to my friends that work on such tech, I need to do some of my own research)

Tesla's strategy has been to use data collection via their current on road vehicles doing partial, distributed processing then sending data back to a mothership. On top of this there is their more recently planned "Dojo" supercomputer using Tesla inhouse ASICs to do more centralised NN training.

[Marco]

The expensive part is the insurance when it can't recognize a stationary fire truck.

MLPs will never do general intelligence.

[james_s]

(Adding Lidar on-top of the camera system is a big plus, but the FSD should be able to operate on sight alone...)

The issue with operating on sight alone is that cameras, at least the sort we are seeing installed in cars, are nowhere near as capable as the human eye. The eye is a remarkable image capture system, it has an enormous dynamic range that can function from the light of a few stars in the night sky to full direct noon sunlight, very high resolution, extremely wide field of view, high frame rate, and on top of all that it can be rapidly and accurately aimed. Cameras have the advantage of being capable of seeing beyond the wavelengths the eye can detect and it's possible to have a larger number of cameras than a person has eyes but this is not even close to enough to make up for where they fall short.

[helius]

Did you some how overlook every other word of CatalinaWOW's post? "no ego" is talking about "automated driving systems" not "Elon Musk" unless you are somehow suggesting autonomous driving systems are somehow magically imprinted with the human personality traits of a CEO ... hang on ... I think I hear the whisper of Steve Job's ghost from my iPhone

Nothing magic about it; if those personality traits, like ambition or grandiosity, influence product development schedules or priorities.
Machines are artefacts of human intelligence and the only "personality" they have is imbued in them by their designers.

[EEVblog]

If I compare using Microsoft Word on a PC in 1991 with using Office 365 on a PC in 2021, the technological differences are astounding.  The modern PC will have a thousand times the memory and I don't know how many times the processor power.  Office 365 is connected online and has a myriad of features that were only hinted at if even that in 1991.  And yet the way I use it the most has hardly changed at all. 

And it's arguably just as slow to load and do stuff.
Some things just don't benefit from improvements.
Solar panels for example, no matter how much innovation you have, it's just best to put them on the roof.
No matter how much toaster innovation happens, it's just best to have a heater coil, a timer and a pop-up lever.
WiFi light bulbs? A switch on the wall works just fine.
Don't mention Juicero.

Autonomous cars either have to be perfect without any driver at all, or you need a driver. Anything else is just glorified smart assistance features we've had for decades. And while that's useful, it's not game changing. Games changing requires a level of advancement we aren't even close to attaining yet.

[JustMeHere]

Autonomous cars either have to be perfect without any driver at all, or you need a driver. Anything else is just glorified smart assistance features we've had for decades. And while that's useful, it's not game changing. Games changing requires a level of advancement we aren't even close to attaining yet.

Agree.  Passenger jets can fly totally autonomously. Probably better than humans.  But we still let the pilots do the work.  They can do a better job (zero visibility landing capable) and we still don't consider them advanced enough yet.

Airbus low visibility, automatic landing:


F-18 (always) automated takeoff:  (The pilot grabs that handle so they will not interfere with the launch)

[andy2000]

I still think practical cars that can self drive in regular traffic, on any road, in any weather conditions are still a LONG way off.  Starting with cars that can be fully autonomous under certain conditions would be a good stepping stone. 

I could see having dedicated express lanes on major highways that are reserved for autonomous cars.  The roads and cars would be designed to work together which would make the cars a lot simpler.  It would be great to be able to join the 120 mph express lane and let the car take over for the next 3 hours while I relax.  We could do this with current technology. 

[sandalcandal]

So I was having a think about issues with LIDAR/RADAR based systems. One of the main "false triggers" for radar is apparently increased returns due to things in the environment that aren't actually in the way of the vehicle e.g. an overpass. You could probably solve this by using PN (pseudo noise) radar or RM (random modulation) CW LIDAR (which have been a concepts since the late 70s) to over come range ambiguity, Doppler ambiguity, and clutter. You could even use multiple antennas then process correlations to get pretty robust 3D maps. Needless to say such systems introduce much more complexity and I'm not an expert in these technologies.

I'm not really sure what people have actually implemented but it doesn't seem like these approaches are being used? Are they potentially restricted tech?

Most LIDAR looks like single point disco balls. Haven't seen anything that looks like multiple radar antennas either (can you get wavelengths small enough for multiple points?).

Impossible to tell from pictures if PN/RM is being used and they aren't mentioned in press releases I could find. It is mentioned in some fairly recent papers though
https://arxiv.org/pdf/1909.01729.pdf
https://www.researchgate.net/publication/338506276_Pseudo-LiDAR_From_Visual_Depth_Estimation_Bridging_the_Gap_in_3D_Object_Detection_for_Autonomous_Driving

I'm guessing at a certain level of complexity in hardware and data processing, it becomes more efficient (cost and development wise) to simply use multiple visual image sensors and neutral net based algorithms to perform processing?

Answering myself. Aurora is a self-driving tech startup that appears to use multi LIDAR to do their environmental sensing. Unlike most other self-driving development programs, they appear to be targeting a commercial road freight application rather than private/personal transport (robo-taxi). This application is probably less cost sensitive than the private transport application Tesla and others are targeting and the level of risk is arguably higher so the multi LIDAR based approach seems more justifiable. Aurora also acquired (was given) Uber's Automated Driving Unit late last year (Uber's self-driving fatal accident happened in early 2018).
https://aurora.tech/

[DrG]

Autonomous cars either have to be perfect without any driver at all, or you need a driver. Anything else is just glorified smart assistance features we've had for decades.

It is an interesting position. I am not sure I disagree, except that it, IMO, would need to be qualified in the sense of autonomous vehicles sharing all of the same roadways as human-driven vehicles. It is not likely that is going to happen soon and not without restricted use cases that achieve some acceptable level of performance.

We could, conceivably, outright reject autonomous vehicles that had a significantly lower incidence of causing injury than human-driven vehicles.

I remember a few years back, autonomous robots were in use at a hospital...motoring around the same hallways as humans. All of a sudden, they disappeared. I asked several people about that and they told me that they were removed because they did not work well, but never caused any injures. I noticed that they were basically wall-huggers and were programmed to stop if anything was anywhere near them - consequently, their performance sucked. Not sure I get it all straight though, but that is what I ended up concluding.