Hello. I'm new here. Sry for my bad English, I hope you can understand me though. :) When I think about A.I. that could pass the Turing-Test, I have some ideas how to create it. Of course I'm not that naive, people surely had these ideas before I had. People who work their entire life on that problems and in contrast to me have a budget that allows them to try out their ideas. We are still far away from from concious machines, but an imitation of intelligent behavior would do the job anyways. I would like to share some of my ideas with you to find out, why it doesn't work. A computer is kind of an artificial "inverted" human. The CPU, cache, RAM and harddisks are essential parts of the brain. Graphicscards and soundcards are kind of too, but instead of their analog parts of the brain they don't analize graphics and sound, they create it. This is what I mean by inverted. So take a computer, attach arms and legs (or wheels), cameras and microphones. Program an OS with native support for all this devices. Now you have a hull, like a human body, with a brain, but an "empty" one. Now the A.I. part. I think the general idea should be to create a servant. You program several threads, one for vidio-analizing, one for sound-analyzing, one for controlled movement. Now the most important and hardest part I think is the speech. The problem is, that our language and the way we think is different from how computers think and how they communicate. So there has to be a layer between the A.I. and the human. Now the idea is to set tasks and rules and let the robot accomplish the tasks regarding the rules. For example I tell the robot: "Get me the book from the shelves." The sound-analyzing thread now has to analize my spoken words and convert them into a text message the robot can handle. Now "Get me the book from the shelves" is in the memory. The most important part is to correctly analize the sentences. Propably we even need some linguists. However, the robot has to recognize that the sentence is a new task. He needs a database to look up, what a book is and what shelves are, so he can recognize them. Now he analizes his position relative to his master, his position relative to the object he has to get. Some pathfinding and he goes for the book. What I am trying to say is, that I think, intelligent behaviour can be cropped down to some basic rules. The A.I. has to analyze the task, several positions and points in time for timecritical tasks. You implement pathfinding and automatically created map. Generally you just implement the things a human would thing when he does something. Like I hear something, I analyze who said that (ah, it was my master ;) ) by face recognition which can already be done today, then I analyze the best way to accomplish a task and so I behave in an intelligent way. There are some additional things to implement, like a thread that forecast the future based on physical laws and warns the A.I. to prevent the robot from being killed by a car or fall down a hatch. Then you implement the obvious rules (no harm to people) and then you need a score system. Some scorecounters to imitate feelings and desires a human being has. Then you just need to set permanent tasks to maximize or minimize those scores and set rules for them to be count. If you add the possibility to learn like adding words to database, setting new rules, you should have a robot that behaves intelligent. And after some time you could port parts of his A.I. and use it for any tasks you want by specializing it a bit. So, wheres the problem by doing this?

You have a gift for understatement, my friend. Your plan may work, but describing what it would take to accomplish any given step of your description would fill a large and expensive book. :) And even then, after decades of studying and understanding all of that, you'd still have a robot that is as intelligent and graceful as a retarded siamese goat.

For example: Machine Vision : Theory, Algorithms, Practicalities

In the mean time, watch these robots dance:

">Hell yes.

Each of these little things (voice analysis, computer vision, robot kynematics, etc) are so immensely complex that they each have thousands upon thousands of people researching them everyday. Just tracking a god-damn point reliably in a video sequence is still an open problem.

Try to make an program that can detect and track "a book" in a video sequence. I dare you :P Also, the best face recognition I ever heard of had an about 85% success rate, and thats with a lot of input and pre-processing and darpa money.

The Forsyth and Trucco-Verri books are much better. Forsyth's book not for begginers.

The problem is context. Look up NLP (natural language processing).

If i say "get the book from the desk" it doesnt mean much on its own. The computer needs to look at the previous context of the conversation to find out what I meant.

Perhaps the whole conversation was :
"I need a book called 'Bobs adventures'"
"Get the book from the desk"

In this case, looking at past sentaces reveals that "the book" is actually "a book titled 'Bobs adventures'".

However, from these sentances, the computer still can't figgure out what "the desk" is. It's video processing module would have to be able to classify every object that it sees.

It would need to be able to see a picture of a desk and think "that is a cup, that is a key-board, that is a pen, that is a book..." then "those objects are on a desk", and "that desk is in a room".
Then it can think "I am in the room, the room also contains a desk, perhaps this desk is 'the desk'".

Also, what if you call the desk a table instead of a desk? The computer needs to categorise the object as "that is a desk, which is a kind of table, this desk is also a wooden structure, it is used to place objects on top of". As you can see, you need a way to construct very complex matricies/webs of information that is all interconnected and automatically categorised...

Basically, you're creating a model of cognition and a model of human memory/knowledge.

Reading up on some psycology books where people have tried to explain how human memory and cognition work might help you understand exactly how complicated organising all this information is.

Also, the speach processing is pretty hard, but many people have figgured it out pretty well, but the video processing is very very hard. Theres only a few computer systems in the world that can successfully categorise objects, and they require a lot of processing power, and a lot of training.

What you're describing is more of a programmable robot, not an intelligent being. You have to attack the problem of intelligence from bottom-up, not top-down, otherwise the thing you create will never be able to learn on its own or develop new skills. Of course then the problem becomes making the right connections and giving the thing goals, especially in the early stages.

2nd the recommendation of Trucco-Verri as a good machine vision text.

The problem with human level AI is that we're not even sure what the challenges are. In fact, "human level AI" is a strange notion to begin with. A human level AI would be able to recognize one face out of thousands with perfect accuracy while simultanously being unable to do more than 10 arithmetic operations per second. Such a machine is unimaginable to me. If/when thinking machines are born, they will be nothing like us, but their intelligence will be undeniable.

Just my 2c.

A computer cannot think. It can only compute. It can add, subtract, multiply, and divide. It can tell if two numbers are the same or if one is greater or less than the other. It can move numbers from one device to another. That's really all. Everything else is a combination of these. You describe a series of steps that an automataton can follow but you did not say how it would compute how to do these things, and that is exactly what information is missing and what people are trying to research.

some stuff

Human level AI is in-part computations and moving numbers and everything, our brains do the very same thing. Take this example: You are an ordinary person and you are presented with a car. Someone tells you to build a car that is similar to it. The only way you would be able to do this is by first understanding how the car works, breaking it down into the parts and eventually the nuts and bolts (the most basic parts of the car). Once you understand how it works then you will have the ability (not necessarily the skills or supplies) to build a similar car.

It's the same thing with our human brain. The only problem is we are starting out with very complex brains and we are trying to understand how they work. But we can't do this without understanding the most basic functions and how they interact with each other.

So to say that human level AI is too basic to compare to our brains isn't quite right. What you should be saying is that we don't yet understand how our brains work so that makes it difficult to make an equal AI.

One of the biggest problems even if you got past making the mechanism -- processing power memory storage, all the program aspects of the AI needed to simulate a humans thinking processes (all of them), you still have to 'teach' the machine all the data - the facts, thinking patterns and interrelations of data.

In complex AI the program part is only 10% or less of the project. The data that makes up the intelligence has to be added to the machinery and that is most of the work.

There is much theory about how the learning might work. Some say that to think like a human the machine would have to live like a human -- be exposed to all the human experiences and situations so that it could absorb what it needs to know.

If you try the hand teaching method, imaging how many people and how many years it would take to convert all those facts and interrelations into equations. There was a 'common sense' (big) AI project done a few years ago. They took something like 10 years and still hadn't come close to what they realized was required.