Fully automated vehicles - 'not in our lifetime'?

kabayiri

Clifford_Pope wrote: »

...
But robots will be judged much more harshly. Just wait until autonomous cars kill a mere 100 people, and the project will be killed stone-dead by angry public opinion.
...

They really won't in other parts of the world though.

Look at the fatalities incurred amongst migrant workers to build the Qatari world cup stadii. The same numbers would generate outrage here.

Over there, it's an acceptable cost.

GreatApe

Malthusian wrote: »

From which we are still a million miles away.

We can program a computer that can beat humans at chess, but only if humans shoot themselves in the foot by telling it that checkmating is a good thing and being checkmated is bad.

Human and animal intelligence is arrived at by pain and death so the environment tells animals what is and is not intelliget. There is nothing wrong with defining success and failure.

We cannot program a computer that can run through a maze to find some cheese, unless we tell it beforehand that finding cheese is desirable. A rat can work out for itself in its pea brain that cheese is desirable. A computer cannot.

A rat didn't evolve into a rat out if nothing it went from simpler things to become a rat and on the way it learnt food would help it survive while hunger will kill it. The environment over many thousands if generations fought the rat food is good for it. It didn't just come into being and magically know cheese is good for it.

This is the nut we have to crack. We can't make a rat level intelligence and we have no idea how to go about making a rat level intelligence, although we're trying really hard and we'll probably get there eventually.

We already have specific AIs at rat level intelligence or more

(Let us say I am playing a chess novice with an IQ of 180, and I sportingly point out that if he sacrifices his rook on g7, after the forced capture he can fork my king and queen. Who's better at chess after he wins? Me, because he couldn't have won if I hadn't told him how. Who's better at chess between AlphaZero and its human programmers? The humans, for exactly the same reason. If you say "The beginner / AlphaZero because they beat you and they're really clever", you are ignoring who was responsible. When they manage to win without their opponent shooting themselves in the foot by showing them how to win, we can acknowledge that they are superior. Their superior brainpower is neither here nor there until they can use it to replace external instruction. AlphaZero is superior at the calculations of chess but until it can work out for itself whether checkmating is good or bad without a human feeding it the answer, it is not superior at playing the game.)

What the hell are you on about we define to each other what is a win and a loss if you and me like we can define taking the knight with a pawn is a win and that's our chess game. Defining what is and is not a win is not part of intelligence that is an agreement in a game. In the real world nature makes that decision. A rat doesn't come out of nowhere and know cats are bad news. A rat learnt that over many generations of death and failure. The environment told it cats are bad

If you want you could define longevity as success just as nature does that and money as food. Then let the millions of self drive cars figure out how to get money and stay alive as long as possible. Over time they would learn to be a super efficient taxi fleet. But we don't need to do that were not trying to create general AI to drive taxis we want a specific AI and we know what we want it to do so we can tell it from the begining.

GreatApe

HAMISH_MCTAVISH wrote: »

Good point.

I remember TV shows in the 70's promising us all Flying Cars in our driveways by the year 2000.

The technology theoretically exists for such a thing but mass production and adoption has not proven viable yet and we're already almost two decades behind where the techno-cheerleaders said we'd be.

We might in the next few years see the larger quadcopter drones currently being developed doing very limited passenger work to and from fixed locations in a few places such as Dubai.

But that's a long, long way from everyone having a flying car... Or even from a flying car being available and legally useable for anyone in developed nations such as the UK.

The TV shows miss the economics part of it. A product needs to be technically workable but also economically workable. Concord failed on the economics part of the equation

A car costs $30,000 while a helicopter costs $1 million
The primary reason we don't use flying cars (helicopters) is because it is much more costly

There are also fundamental reasons why a helicopter can never cost as little as a car.
There are some exceptions the super rich do use flying cars some of the time.

Self drive tech has a very strong economy reason to exist it frees up huge amounts of people and capital and would improve productivity and costs by a massive amount

GreatApe

HAMISH_MCTAVISH wrote: »

Good point.

I remember TV shows in the 70's promising us all Flying Cars in our driveways by the year 2000.

The technology theoretically exists for such a thing but mass production and adoption has not proven viable yet and we're already almost two decades behind where the techno-cheerleaders said we'd be.

We might in the next few years see the larger quadcopter drones currently being developed doing very limited passenger work to and from fixed locations in a few places such as Dubai.

But that's a long, long way from everyone having a flying car... Or even from a flying car being available and legally useable for anyone in developed nations such as the UK.

One of the problems are that things are already good enough for a comfortable life and a lot of the advancements most people think of are on the fringe and wont be all that important anyway

Take airplanes one side of Europe to the other in about 3-4 hours. Halving the time by doubling the speed isn't really going to change our lives much especially when you consider the total time for the trip including getting to the airport and waiting etc is not halving but is maybe being cut by 20%

The next revolution is flight will be spacecraft. Leave New york go up 50km and then across the globe and then back down. Half way around the world in half an hour rather than 12 hours. Now that would be a big improvement in travel speed but again its virtually irrelevant to productivity as a whole because how frequently does the average world citizen currently or in the future want to travail half way across the world? Maybe a couple of times in a lifetime? The savings are thus almost trivial even if we had planes or spaceships 10x father than today its not all that important or useful

This is where AI is so different, not only will it work but it can have a huge impact.

Lets take for example something as mundane as getting your hair cut.
It takes maybe 20 minutes and the cost is more or less proportional to the time it takes.
Someone at some point will invent an AI clipper or an AI robot so the job goes from 20 minutes to effectively a 2 minute but the clippers or robot automatically adjust the size and cut and do all the hard work. Sounds boring very boring but that tech would actually be more useful and bring more productivity to the world than super fast spaceships replacing 747s

Electricity was the last mega productivity jump. The period 1950-1980 was defined by electricity making huge improvements in working practices and productivity. That was the real industrial revolution. A strong stable grid + electric motors made mass manufacturing possible. The 1700s steam engine revolution was mostly useless (apart from a stepping strong to the next thing) as all the way even upto the 1900s we had kids go to school bare foot in the UK they were so poor it was still 3rd world conditions for most people in the west the steam revolution didn't do much. The electricity revolution took us from near 3rd world to 1st world in just one generation. As is happening in China today. China began the electricity revolution around 1990 and fast forward 30 years it is close to first world nations. India is about 10-20 years behind and africa (apart from a few nations) is still not in the electricity age

Near AI is going to do what electricity did but it will likely be quicker and more powerful
The first mega change is going to be self drive vehicles. Globally It is worth in the region of $5 trillion annually because it allows so much productivity and labor/capital savings.

Many areas will be hit by AI and collectively they will make us much much richer.
It will start off slowly but compound over time. China eletricity usage from 1990-1995 went up modestly (even if it was 10% a year) but its usage 2013-2018 went up massively (even if it was still 10% a year). It will be the same for near AI

Herzlos

GreatApe wrote: »

Let's say a robot EV costs $40,000 but does 500,000 miles over 5 years. The electricity cost of 2.5 cents a mile plus the capital cost of 9 cents a mile gives 11.5 cents

So the comparison is 40 cents vs 11.5 cents per mile. The biggest saving is from capital and capital interest per mile.

If the fleet owners charge 20 cents a mile the cost of using a robot fleet would be half that of ownership.

You're comparing apples and oranges here. You've got personal car fuel as being $.10/mile and EV fuel at $0.025/mile, so are you comparing ICE and electric?


What you should be comparing is the cost of using a robotaxi as a taxi, and a cost of using the same robotaxi as a private car. The only distinction there is between the mileage/cost/availability. Assuming people only buy new private robotaxis and not older ones, and keep them on the road as long, then:


A private car will cost about 3x as much in depreciation, but about 0.3x as much in terms of tyres, maintenance, tax, etc.



There will reach a mileage point where the private car breaks even, since the taxi company are taking about a 50% profit margin.



Somewhere before that point, users will happily pay a bit more for the convenience of having their own car - it'll be as they left it, no-one else will be using it, they can leave stuff in it or pre-load, they can get upgrades.

GreatApe wrote: »

The sensor hardware already exists

True, but it's still very bulky.

and is very affordable.

Not true. The current sensor equipment for an autonomous car is expensive, though that will come down somewhat with scale. It's also pretty heavy.

A $200 smartphone already has all the sensors needed.

Not true in any meaningful way. You could probably frankenstein something onto a dismantled $200 smartphone, and it could 'read' from the sensors but it'd be completely incapable of doing anything useful with it, even without the operating system overhead.

The processor might not be powerful enough but given another 12 years processors will be 100x the power of today. Application specific chips will be used so that's another 10-100x up which brings us towards 1,000 - 10,000 x the power of a smartphone today which is probably more than enough.

I'm not so sure, at some point Moores Law is going to hit problems with phsyics, we're already down at 3nm test chips, with stuff around the 10nm being in production. I'm not sure how much smaller and thus more dense we'll be able to get.



They'll be drastically more powerful than today for sure, and at some point we'll hit a series of ASICs which will do everything needed for a reasonable power and dollar cost.



But none of that addresses the human element - folk are going to be really reluctant to let a robot drive them on the open road.

kabayiri

Herzlos wrote: »

...
True, but it's still very bulky.

Not true. The current sensor equipment for an autonomous car is expensive, though that will come down somewhat with scale. It's also pretty heavy.
...

Where do you get this heavy/bulky stuff from?

One thing sensors do is track acceleration/deceleration and orientation.

Have you seen the size of the 3 axis accelerometers today compared to what they were?

Which sensors are you referring to?

Herzlos

Sure, 3-axis accelerometers are tiny now, but there's a lot more needed for self driving cars, like LIDAR:





And then there's the rest of the sensors:





Each of which will need some kind of communication node and wiring (you can't use wireless for reliability or security reasons), which will take it back to whichever processor is doing the work.




This stuff will all get smaller, but it's still at least 3/4 generations away from being small enough.

kabayiri

Herzlos wrote: »

...
Each of which will need some kind of communication node and wiring (you can't use wireless for reliability or security reasons), which will take it back to whichever processor is doing the work.


This stuff will all get smaller, but it's still at least 3/4 generations away from being small enough.

This is where I differ from the fully autonomous vehicle lot.

You can bolster the intelligence and sensor awareness of the road network, and reduce the sensor load on the vehicle.

I'm more interested in getting cars to work as a collective, particular around difficult bottlenecks on our road network.

I mentioned before about individuals gaming the current system, to the detriment of the whole.

The problem with my argument is that doing anything to the road network here is ridiculously expensive....which is why it might have more traction in relatively green field environments.

Herzlos

That'd only work once there were no human-controlled cars on the road, though, but would likely be the better option in the future.


In the first few iterations the cars would need to be self sufficient, beyond potentially sharing sensor information with neighbouring cars (so you can see what's happening 10+ cars in front),

GreatApe

Herzlos wrote: »

You're comparing apples and oranges here. You've got personal car fuel as being $.10/mile and EV fuel at $0.025/mile, so are you comparing ICE and electric?

What you should be comparing is the cost of using a robotaxi as a taxi, and a cost of using the same robotaxi as a private car. The only distinction there is between the mileage/cost/availability. Assuming people only buy new private robotaxis and not older ones, and keep them on the road as long, then:

A private car will cost about 3x as much in depreciation, but about 0.3x as much in terms of tyres, maintenance, tax, etc.

There will reach a mileage point where the private car breaks even, since the taxi company are taking about a 50% profit margin.

I compared a $32,000 ICE to a $40,000 EV. If you want to stick to EV on both it would be 2.5 cents a mile for the power. But the capital cost the biggest cost will still be much less for the EV taxi as it might do 500,000 miles in five years vs 150,000 miles over 15 years for the private EV. Not only is there a massive milage advantage but also a capital advantage.

For the robo taxy it is 9 cents a mile for the human EV it is 38 cents a mile so the cost differential is close to 30 cents a mile. The insurance and maintenance is assumed roughly the same.

Over a 80 year lifetime a person might do what 400,000 miles?
The saving is thus on the order of $0.30 x 400,000 = $120,000
This isn't per family it is per person.

Yes still the profit margin to take into account but it wont be 50% almost no business has a 50% margin most operate closer to the 5-15% mark. Also we havent talked about shared trips which massively reduce the cost per person per mile further

Somewhere before that point, users will happily pay a bit more for the convenience of having their own car - it'll be as they left it, no-one else will be using it, they can leave stuff in it or pre-load, they can get upgrades.

Sure if you do lots of miles or need more convenience you can own your own self drive car and you will pay more for it.

True, but it's still very bulky.

I dont think we will need radar lidar and the like for superhuman levels of driving ability. Just cameras. All you have are your eyes and you drive ok. a smartphone already can operate at 30 x the frame rate of your eyes and at a central resolution 10x higher and have 20 cameras vs your two eyes and they are so cheap most $200 smartphones come with 2-3 cameras and we are talking about 2018 by 2030 cameras will be much better and cheaper still

Not true. The current sensor equipment for an autonomous car is expensive, though that will come down somewhat with scale. It's also pretty heavy.

A 2030 smartphone will probably be enough so <200 grams and 1 watt of power consumption.
Sounds silly but a smartphone today already has GDP, accelerometer, microphone, speaker, CPU GPU multiple HD cameras capable of 10+ MP and 500+ fps. Already the sensors are far more capable than your sensors its just the information processing isn't there yet and application specific processors by 2030 will be >1,000 x more powerful than today

Not true in any meaningful way. You could probably frankenstein something onto a dismantled $200 smartphone, and it could 'read' from the sensors but it'd be completely incapable of doing anything useful with it, even without the operating system overhead.

The only thing you need are cameras as the sensors and cameras are already superhuman and super cheap. Im not talking about the processing hardware just the sensors. you dont need radar to drive I assume you dont have radar or sonar or lidar but you can drive right?
Already smartphone cameras are higher resolution than the central part of your eye and they can do close to 1000 frames per second your eye and processor are closer to 10 frames per second.

I'm not so sure, at some point Moores Law is going to hit problems with phsyics, we're already down at 3nm test chips, with stuff around the 10nm being in production. I'm not sure how much smaller and thus more dense we'll be able to get.

Yes but there is still time we are not quite there yet I suspect it will continue until 2030 but there will be problems post that point. But you dont necessarily need more transistors. Application specific chips can get magnitudes improvement using the same transistor count.

There is also likely to be another step in processing tech. Transistors are great but before them we had vacuum tubes and before them magnetic mechanical switches and before them cloggs etc. There is no reason to assume this is the final medium of information processing.

They'll be drastically more powerful than today for sure, and at some point we'll hit a series of ASICs which will do everything needed for a reasonable power and dollar cost

If my understanding of AI is roughly right I think once you train an AI and have the 'map' you could build a very powerful very specific AI chip that does just that one task but does it so so fast at very low power. We dont have anything like that right now but with 200 million cars produced annually its a big enough market to build super specific chips for that market

But none of that addresses the human element - folk are going to be really reluctant to let a robot drive them on the open road.

They wont care the 17 year old kid that doesn't have the £5k up front for a 10 year old banger and insurance is going to use the 20p a mile self drive taxi.

All the taxi miles now done at £1 a mile will swith to 20p a mile robo taxis.
At the same price point maybe you would be correct people would opt for a human driver but not at a huge price differential