Fully automated vehicles - 'not in our lifetime'?

GreatApe

Herzlos wrote: »

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.

A good human driver is a lot better than the average human driver and he achieves this without a LIDAR or RADAR sticking out of his head

All that is needed for super safe driving are cameras and processing the imagine

The sensors already exist and are tiny and already surpass human sensors.
My phones cameras are higher resolution than my eye
I can determine at most 10 frames a second and just barely while a new camera can do closer to 1,000 frames in perfect vision and just look how far digital cameras have come over the last 15 years the next 15 will be further improvements and make todays cameras look crap

The sensors are done its the software and perhaps the processing power of the ASICs

GreatApe

A human is about a gigabyte of information and some self learning on the way. We start off a two cells roughly 0.001 square millimeters. Sure there is a lot of compression of the data and some additional data needed along the way but we are not made of anything special or mystical we are more or less made of the same matter in the same proportions as the universe is. When you have a baby you turn inanimate matter into a human.

We can't by pure chance be the peak of all possible intelligence or ability (which is obviously correct as humans vary in ability and intelligence let alone possible other biologies or non biological information processing systems) we will be surpassed the only question is when and the exponential nature of information processing progress means the when is probably a lot sooner than you would imagine.

GreatApe

Malthusian wrote: »

From which we are still a million miles away.

The thing about the exponential function is that a million miles away is very close.

If I walk one mile today, two miles the day after, four miles the day after that and so on i would cover the million miles in about 20 days. So lets say its not a million miles away its a trillion miles away. Well that just means 40 days

Information processing has been improving by around 40% annually for the last 50 years if this exponential continues by the end of this century you will have in your pocket a computer that is a million times as powerful as every single computer on earth in 2018 combined.

It is also likely when we get near AIs the annual 40% improvement could be compressed to maybe say 40% improvement every quarter then every month then every week then every day.

Humans probably wont exist beyond this century maybe what comes next will call itself human but it wont look or think anything like us

westernpromise

GreatApe wrote: »

Humans probably wont exist beyond this century maybe what comes next will call itself human but it wont look or think anything like us

You mean like Scotland?

Malthusian

GreatApe wrote: »

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.

We can but we don't. Taking the other guy's king is the game. Why? Because as humans we understand the concept of "our pieces" and "their pieces". That gives us the game of draughts, and if you add the concept of a leader who must be protected, you have chess. As humans we understand why these concepts are important.

Why do we play chess and not misere chess (where the first person to lose all their pieces wins)? Partly because misere chess is a silly game which you can lose on the first move. But that's not the real reason, because if we'd been playing misere chess for 1,500 years, we'd've solved that problem with rule changes by now.

The real reason we don't play misere chess is because games are a simulation of conflict or problem-solving. A game in which you simulate getting your whole army killed is unsatisfying to play because getting your whole army killed is not something you would want to do in real life. The only reason anyone plays misere chess (on rare occasions) is as a sort of deliberately contrary, absurdist thought-experiment.

Chess is an abstract simulation of a fight. A toddler wrestling another toddler in a sandpit can work out that in a fight you want to win. So can a rat. A computer cannot, without human intervention.

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.

A computer does not understand that death or failure is bad. Unless a human programs it into them. That is the fundamental problem that you are failing to grasp, that is the gap between what computers can do and what intelligence is.

We are not capable of making a computer that can work out for itself that failure and death is bad. We are nowhere near rat level intelligences.

We already have specific AIs at rat level intelligence or more

That's like saying I have an abacus at human level intelligence. Look, the abacus can solve complex problems way faster than I can on my fingers, as long as a human moves the beads in the right way. It's not anywhere close to rat level intelligence because it's incapable of numerous things that the rat can do.

You are making the classic mistake of thinking that all we have to do is make better and better abaci and suddenly presto! the abacus will start moving the beads around by itself. It never, ever will. There is something fundamental missing which is the Holy Grail of AI research. No scientist working on the problem of AI claims they know when they will find it. Why do you think they are wrong?

The specific AIs you talk about are inadequate for the task of solving the problem of how to create an algorithm that can successfully navigate a road environment, because as we've seen, if you tell them "work out how to not kill the humans", it will kill a load of humans anyway because it didn't consider them human, and was incapable of working out that it was wrong.

Herzlos

GreatApe wrote: »

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.

You said it'd cost 9p/mile but taxi companies would charge 20p/mile, which means the differential is 18p/mile. You're also assuming that insurance and maintenance will be the same for 10k miles/year as 30k/miles year, though the difference is going to be less than the amortization over the miles.

I dont think we will need radar lidar and the like for superhuman levels of driving ability. Just cameras.

And you'd be wrong, otherwise google wouldn't be spending the money on LIDAR.


Our stereo vision helps us judge distance brilliantly in a way you can't get with a pair of cameras, and even then our accuracy isn't that good. LIDAR, radar etc can give much more accurate distances which allows them to potentially do more.


You'd have to do a lot of image processing to get results that just aren't as good as using specialist range sensing equipment.

GreatApe wrote: »

All that is needed for super safe driving are cameras and processing the imagine

Just no.

The sensors are done its the software and perhaps the processing power of the ASICs

The sensors will need to become smaller and cheaper to hit mass market. The processing power of the ASICs will need to increase drastically, along with their power consumption.




Have you tried reading any papers on how this stuff works?

GreatApe

Herzlos wrote: »

You said it'd cost 9p/mile but taxi companies would charge 20p/mile, which means the differential is 18p/mile. You're also assuming that insurance and maintenance will be the same for 10k miles/year as 30k/miles year, though the difference is going to be less than the amortization over the miles.

And you'd be wrong, otherwise google wouldn't be spending the money on LIDAR.

Our stereo vision helps us judge distance brilliantly in a way you can't get with a pair of cameras, and even then our accuracy isn't that good. LIDAR, radar etc can give much more accurate distances which allows them to potentially do more.

You'd have to do a lot of image processing to get results that just aren't as good as using specialist range sensing equipment.

Just no.

The sensors will need to become smaller and cheaper to hit mass market. The processing power of the ASICs will need to increase drastically, along with their power consumption.

Have you tried reading any papers on how this stuff works?

You don't need anything more than vision (eg a camera) to do better than Hunan driving.

The reason google is trying lidar and radar it is it makes the computation easier much easier

It is a race you have companies like Tesla and mobile eye working primarily on vision system. They will need huge amounts of data and AI to get to level 5

On the other hand you have google and uber doing more lidar based approach. They need to get the lidar kits much smaller more accurate but don't need as much AI or data or processing power

I don't know which one will win this race but both can work.

Personally I think the Tesla and mobile eye method is more likely it is passive and ultimately it can be shrunk down to the size of a smartphone physically and power draw wise. So 1 watt of power plus about 200 grams in size. The lidar route too will shrink but I'm not sure it can match vision only on power and price size and mass.


Have you seen project tango?
A google effort to do 3D mapping with dual camera smartphones
Looks amazibf already. Go forward 10 years and you can see how with just a smartphone and smart software you can do 3D world imagining including distance and location. Its already amazing but like with all tech give it tens years and it will be 100x better.

GreatApe

Malthusian wrote: »

We can but we don't. Taking the other guy's king is the game. Why? Because as humans we understand the concept of "our pieces" and "their pieces". That gives us the game of draughts, and if you add the concept of a leader who must be protected, you have chess. As humans we understand why these concepts are important.

Why do we play chess and not misere chess (where the first person to lose all their pieces wins)? Partly because misere chess is a silly game which you can lose on the first move. But that's not the real reason, because if we'd been playing misere chess for 1,500 years, we'd've solved that problem with rule changes by now.

The real reason we don't play misere chess is because games are a simulation of conflict or problem-solving. A game in which you simulate getting your whole army killed is unsatisfying to play because getting your whole army killed is not something you would want to do in real life. The only reason anyone plays misere chess (on rare occasions) is as a sort of deliberately contrary, absurdist thought-experiment.

Chess is an abstract simulation of a fight. A toddler wrestling another toddler in a sandpit can work out that in a fight you want to win. So can a rat. A computer cannot, without human intervention.


A computer does not understand that death or failure is bad. Unless a human programs it into them. That is the fundamental problem that you are failing to grasp, that is the gap between what computers can do and what intelligence is.

We are not capable of making a computer that can work out for itself that failure and death is bad. We are nowhere near rat level intelligences.

That's like saying I have an abacus at human level intelligence. Look, the abacus can solve complex problems way faster than I can on my fingers, as long as a human moves the beads in the right way. It's not anywhere close to rat level intelligence because it's incapable of numerous things that the rat can do.

You are making the classic mistake of thinking that all we have to do is make better and better abaci and suddenly presto! the abacus will start moving the beads around by itself. It never, ever will. There is something fundamental missing which is the Holy Grail of AI research. No scientist working on the problem of AI claims they know when they will find it. Why do you think they are wrong?

The specific AIs you talk about are inadequate for the task of solving the problem of how to create an algorithm that can successfully navigate a road environment, because as we've seen, if you tell them "work out how to not kill the humans", it will kill a load of humans anyway because it didn't consider them human, and was incapable of working out that it was wrong.

You only really have two options

Either humans are created by a supernatural power which gives us our ability to think and be intelligent or there is no supernatural power and matter can get together and form intelligent systems. If you believe in the latter then why would you assume humankind is the maximum possibility of information processing and decision making?

Forget cars. This is the century humans go extinct. Whatever replaces us might call itself human but it will be as different from today's humans as we are from the single cell organisms that eventually created us.

kinger101

GreatApe wrote: »

This is the century humans go extinct. Whatever replaces us might call itself human but it will be as different from today's humans as we are from the single cell organisms that eventually created us.

A century is not an evolutionary timescale, when it's between 3-5 generations.

GreatApe

kinger101 wrote: »

A century is not an evolutionary timescale, when it's between 3-5 generations.

What replaces us won't be biology