robots are all around though most don't have andriod bodies or do very many different things. Smart appliances are robots.
Robots wash my dishes, robots wash my clothes, robots cook the dinner, robots wash the floor, robots vacume, robots mow the lawn, robots filter my email or answer it, robots answer the phone, robots offer a route through traffic, robots interpret my voice commands, and some will bring water or a beer and tell a joke or do a dance.
Next they will track the supplies and place the orders and open the packages and take out the trash, drive me to work, etc. There are none that do all of that and are cheap, and we won't see that for a while.
Do people want those jobs done, or do they want to see an overpriced andoid struggling to do all those different things at once and failing?
I want an overpriced android that can learn to do those things and and not fail. Cheap specialized machines optimized for a specific task will always be around but there is value in having a humanoid shaped machine with human like learning skills. The value is that they can operate, and live in, a world built for humans. They would be able to use all the tools designed for humans to use - like driving a car or pushing a lawn mower, or walking up and down stairs, coking meals in a kitchen, clean up a house, build things with normal human tools.
Before we can create a general purpose learning android we have to solve AI. I tend to be optimistic about how long that's going to take (I've recently made another 10 year bet on the subject), but it's possible it will still take another 50 or 100 years to duplicate human level learning skills. This is the technology that's holding robotics back and no one really knows how long it's going to take to solve. All we can do is make some wild guesses. (and work hard to solve it).
Around the turn of the 19th to 20th century the new thing was electric appliances that were run by motors. Motors were expensive, so people who were lucky could afford one motor which powered all their appliances. They had to multitask the motor attaching/detaching it from each appliance and running them one at a time to get assistance with their work.
Bad engineering, good economics while motors were so expensive. Economics drove down the price and people got more convenient appliances and people could get assistance on more work more easily or even in parallel. They no longer had to wait for one motor.
Around the turn of the 20th to 21st century the new thing was electronic appliances that were run by embedded processors. Processors had been expensive, so people who were lucky had been able to afford one processor and if it was powerful enough it could multi-task and do more
than one computing job at a time. But around the 21st century processors became cheap and powerful enough that people owned dozens of them in their laptop, cell phone, ipods, and automobiles. They no longer had to wait for one processor.
The machines got smarter and began doing most things previously defined as AI, having conversations, monintoring conversations, looking up answers on the Internet, searching databases, solving math and logic proglems, playing games, doing new research, precision positioning, group activity planning etc. and most of this became possible when collections of computers were interconnected.
Is Santa Claus brining you one?
All of the problems except economic and engineering ones have been solved. AI has been more advanced than most hobbiests realize for a long time.
You are not taking into account the fact that human learning skills are not a fixed quantity. Average human skills are dropping about as fast as computer capabilities are rising.
Even so a million dollar humaniod robot is closer to a goldfish than to a minimum wage human worker in practical functionality today.
Decent vision, voice, reflexes, navigation, planning, and general purpose learning and reasoning are out of the range of computing toys today but we are pretty close. They are way out of the range of most toy robot.
Consider that if a neuron can be simulated with 100 instructions per second a thousand dollar laptop can simulate 10^7 interconnected neurons and 10^4 of them interconnected could do much of what a person does with a couple pounds of grey matter. Now once you figure out how to get that ten million dollars worth of computers and megawatt power supply inside the head of that humanoid android that you want to replace your minimum wage servant you have solved an important remaining problem. Then you just have a few other engineering problems to get those other manufacturing costs and maintanence costs down below those of a jumbo jet.
I heard that a lot fourty years ago before better and better solutions to most problems were demonstrated.
Of course, remember that people argued that controled heavier than air flying machines were simply impossible for years after the Wright Brothers demonstrated it, and people still argue whether space flight is possible, or whether old AI problems are solvable. But this sort of thing creates more new problems than the number of old problems that it solves.
The real problems are neither AI nor econcomics. They sort themselves out. The real problems are social just as was predicted fifty years ago.
My main interest is in AI, not robotics. I am a hobbyist but yet a still have I have a good understanding of the field of AI. No one has invented a machine than can learn, on its own, to do all those things. The current learning algorithms are very limited. They tend to only be applied, and work with, very specific toy problems.
I've never heard that reported. What data is there to support that idea?
Right. There are reasons I made the bet that human level machine intelligence would be here in the next 10 years. (9 now). I think we are closer than most people believe.
Well, those calculates are very questionable (as are all attempts to estimate the amount of computing power to duplicate human intelligence since no one knows what computer power is needed - it's still an apples to oranges comparison at best).
How many instructions per second does it take to simulate a transistor which can switch a billion times per second in a digital circuit? Neurons fire around 1000 times per second max and you said it takes 100 instructions per second to simulate one so I assume this means you would estimate 10^8 instructions per second to simulate a transistor. Using this logic, it would take a building full of interconnected lap tops just to duplicate the power of a single CPU chip. And that of course is absurd because we know it only takes one CPU chip to duplicate a CPU chip and not a building full of lap tops.
We don't need to simulate transistors to create a computer, and we won't create intelligent machines by simulating neurons in software. Cortex simulation projects are great research tools, but they are not going to be the basis of real intelligent machines any more than SPICE circuit emulators are used to build radios.
The problem with estimating the amount of hardware that's needed to duplicate human level intelligence is that we don't yet know how to build it, and if we can't build it, all estimations of how much hardware it's going to take are likely to be way off base - in either direction.
For example, with neurons switching at 1000 times per second and 100 billion of them in a brain, you get hardware with a max speed of around
10^14 operations per second. But transistors can switch 10^9 times per second so you only need 10^5 transistors to get the same max switching performance. A lap top with 1 GB of memory has over 8x10^9 transistors in it (one per every bit of memory). That means a single lap top has 80,000 times more computing power than a human brain.
Now, I'm not trying to argue that this is a valid way to estimate the amount of hardware required, but I am tying to argue that it's probably no less valid than your numbers. A lap top sized board full of custom chips has more information processing power than than a human brain. But once we understand how to build human level intelligence into a machine, will we be able to do build it in that form? We just don't know yet since no one knows how to build it in any form.
I suspect however that as we finish mastering the technology, that is exactly what we will end up with - custom hardware that if built with today's electronics technology, would only be the size of desk top computer.
Neurons range from 4 to 100 microns in size. Transistors are below .25 microns now. Large parts of the brain is filled with interconnects (white matter), computers reduce interconnects by using high speed switching and sharing interconnects. Evolution didn't have access to high speed switching devices so to keep response time low, it had to use massive interconnects. When we redesign with a different technology (transistors instead of neurons) we will end up with a very different architecture. It's likely a design with more switching devices (transistors) and fewer interconnects will be the optimal solution when building with the higher speed transistors. In the end, I expect our machines will be substantially smaller than a human brain in order to duplicate it's same power.
Yeah, I was just reading about the amount resistance the idea got. Interesting stuff.
Well, the social issues are sorting themselves out as well. There's a lot of momentum to be overcome. A surprisingly huge percentage of the population still doesn't believe in evolution after 150 years. But like the Wright brothers, once you build something that works, it doesn't take long to make society believe. I believe it was only a few years to get rid of the doubters in the case of the Wright Brothers work. With the speed of communication we have today, word would spread very fast - if only we had something to show them.
I design and program new computers. My background includes AI too.
Algorithms are not the problem. Toy computers do well on toy problems but serious computers are expensive. I accept that you can't afford a supersonic plane, I don't accept that that means that they are impossible and don't exist at all.
I could point you to a lot of books and scientific papers. I would recommend Jaques Barzun's From Dawn to Decadence or The Twilight of American Culture by Morris Berman as excellent references to trends like the decline of literacy. You know, over half of American's don't know if the sun goes around the earth or the earth goes around the sun etc.
I was just using a rule of thumb offered here by Hans Moovec a few years ago. Read his book Mind Children for more details..
Your arguments seem to have mixed up the idea of simulating neurons on a digital computer with simulating a digital computer using neurons.
I design and program new computers for a living. They are optimized for things like AI and robotics. This is a mixture of abstraction, large computations, realtime performance, low cost, and lower power use.
Designing new computers is mostly about simulating transistors so I could not disagree more that simulating transistors is not important.
Who is the "WE" you refer to creating new computers? That certainly doesn't include me or the people I work with or the people that they worked with at Intel or General Dynamics or whatever.
Who is the "WE" you refer to as creating intelligent machines? That doesn't include me or the people who I have worked with who did that with your tax money.
In a lot of cutting edge science people are using computers to simulate neurons, or using genetic algorithms to make new discoveries for them. Some of the hardware and software that most impress people was not really designed by people.
We? Who is this we?
Is that the end of time, or just the end of humans?
Well it depends on whom you asked and when. You know there are people who don't think man actually went to the moon.
Judging what technology exists by what the public is told is a very bad idea. If you are part of the 'we' you know that the trick is getting close to, but not crossing the line of what the public is allowed to know. What can be made public is often a few decades behind what is classified. The trick to commercializing technology is to not get too far ahead of what the public is allowed to know about.
Here is a view of history: When _____ was still classified it was common knowledge that what it did was impossible.
My favorite is when amatuer scientists tell you that you can't break the laws of physics. It's funny, I thought that is what physicists try to do every day, break the old laws, develop new ones. And they are not interested in only things that the average person already accepts.
Sure, I agree that some things require very large and expensive computers these days. And some AI tasks might be doable with today's technology, but simply haven't been done because the project can't be cost justified. But there are still many things no one knows how to do, no matter how much money they have to buy or build hardware.
Show me for example a robot that acts like real dog. I don't care if it's a billion dollar computer connected wirelessly to a robot or if it's just a simulation of robot dog. I've never seen any software that acts anything like a real dog. Have you? If I gave you unlimited funds, could you build one without first developing new technologies?
Ah, you were talking about how average human knowledge is declining (at least in the US). I thought you were talking about how learning skills were being evolved out of us. I see those as two very different things. You seem to want to lump it all together as one.
I see the problem of AI as being a problem of building a strong reinforcement learning machine. So I tend to use the word "intelligence" to describe the raw learning power of a machine instead of using it to describe it's base of accumulated behaviors (aka knowledge). The behaviors do in fact improve it's learning skills, but still, I see only the core learning skills of the machine as its real intelligence.
Cool. Which ones? Anything that I would have heard about?
You seem to have missed the point. You are talking about using simulation tools to aid in the design of new computers. I was talking about how we don't build transistors simulators into the new computers we build. We instead, just put real transistors into the computers.
I did say "build" and the design process is generically all part of what we do to build new computers, so I understand your reaction. But the point I was making is that the most optimal design for a new computer isn't created by including a transistor emulator in the computer - we just use transistors. And likewise, when we better understand how to create intelligent learning controllers, we not build them by using transistors to emulator a neuron based learning controller. We will simply build a transistor based learning controller.
Are you talking about neural nets or actual neuron simulations? Neural nets logically might be thought of as some type of neuron simulation but they are not in fact anything like a neuron simulation. They would not do anything useful if you took a neuron out of your head and put a node from a neural net in it's place.
All the real neural simulations I've seen reference to are only tools for helping to understand what real neurons are doing and work just like all our simulation tools.
Neural nets of different types, which are sometimes incorrectly talked about as neural simulators, are in fact many times just research into the behavior of new algorithms and the only connection to real neurons is that their design was inspired by the existence of neurons. I've spent a lot of my time working with these types of things, but I never made the mistake of believing it was a neural simulator.
You seem to be jumping to places I just wouldn't go. There are a few people that have worked for years to build real neural simulators in order to understand both what individual neurons are doing, or to understand what networks of neurons are doing. These are all driven by real data collected from real neurophysiology studies and their results are compared to real neurophysiology studies.
Neural nets, which are not designed from real neurophysiology data, and who's results never get compared to the operation of real neural networks, are not neural simulations even though the word "neural" is part of their name. They are not simulations at all. They are data processing algorithms optimized for their specific purpose.
Human level machine intelligence will no doubt in my view be created by neural network like algorithms. But they won't be neuron simulators.
You, me, and everyone else working on some part of the AI problem.
Oh, 50 to 100 years order of magnitude I'd guess (to make a machine with the power of a human brain smaller than a real human brain). Humans I think will be evolved out of existence by the machines (not wiped out - just replaced by attrition) but that will probably take thousands of years. It's fun to speculate about but most likely what actually happens will be different from what everyone was expecting.
Yeah, well, there will always be a fringe that believes all sorts of odd things.
So, you design classified technology for a living? I worked as a contractor for the Navy for many years but I wasn't connected with any classified work.
Yes. It was not a physical dog, or an organic dog, and it didn't reproduce sexually with real dogs. But it was designed to allow AI professionals to demonstrate to non-professionals what a dog's neurons do. It was a software simulation of virtual dog built with simulated neurons and studied in a simulated reality.
After I took the money I would tell you to look for it on google.
I am working on SeaForth24 now. Scalable Embedded Arrays. I doubt if you have heard of it, just announced, prototypes, but not for sale to the public yet.
Gerald M. Edelman won a Nobel Prize for that kind of research. He went on to create his theory of neuronal group selection and to apply it to the real world. They did large scale implementations of large self-programming learning neural societies demonstating a continuum of intelligence.
To demonstate it to the public one of the people at the institute wrote a small example, a simulation of a dog in a simulated world. The dog would learn. Your job was to train it. Like a real dog it could learn the wrong things if you were not good at training it. It acted like a dog.
Alan Alda demonstrated his skills at training an artificial dog to do tricks on an episode of the Scientific American Frontiers program. The dog demo was just a toy to show a small example of Neural Dawinism to the public who are not well educated about such things. Google on the theory of Darwinian Neuronal Group Selection or old computer programs like Darwin II and the dog demo demonstrating the theory in action.