November 2, 2011 at 21:54Thoughts aloud about humanoid AI
I have long dreamed that the computer was a smart assistant. So that he could describe the task with words, and he completed it - as in many stories. I talked to Alice, but you yourself understand that she is far from the goal. Probably because he does not know how to think, feel emotions, and so on - this is a standard set of AI problems. And I got an idea how to solve this problem, and I want to tell you about it.
The idea is to grow AI as a small child, and do it as detailed as possible. A machine does not understand a person simply because from its “point of view” everything “looks” completely different - on the example of Alice, its world is a set of symbols, incoming and outgoing flows.
I propose to show the human world to a computer - in the most similar light. So that he can see, hear, touch, he can even smell. To give him a human body, but not physical - it will cost too much and difficult, but simulated.
Imagine a toy, like Sims (I didn’t play it myself, judging by the screenshots), in which AI controls one body. I do not think that there is a need to describe this body in great detail - but some nuances are important. For example, vision: on the AI input, send pictures from two eyes (not the player’s eye, the “camera”, but the body’s eye, that is, the picture that you would see if you were in its place) to be able to get a three-dimensional representation. There’s a minimum of sensors from the arms, legs, and other parts of the body - so that he can feel if he hits the back of his head, or understand when he grabbed his glass with his fingers so as not to squeeze his fingers further and break it. The way out, as you already understood, is to control various muscles - arms and legs, eyes (to simplify the neural network, you can lower the picture resolution, but add the “zoom” feature), and so on.
Immediately make a reservation about a broken glass - I do not claim that the physics of destruction is a necessity in this model, but I believe that its presence will allow the computer to better understand a person.
Of course, if you take it this way and immerse the neural network in such a body, the result will be zero. Too much chaotic data, and one generation will not play any role here. Therefore, evolution is necessary - to begin with, at least teach a “person” to eat (one of the sensors will be a hunger sensor, AI will receive data from it, but initially will not know how to react to it). I think many generations will die while learning to eat (at this period, generations will have to be propagated manually). Then you need to teach at least a little control of your body - for example, move to food (at least crawling). I believe that after such "primitives", we will get a creature that looks like a baby. It should have basic needs - food, comfort, reproduction. And this baby needs to be educated further - to teach him to walk, talk. Give it to some kind of kindergarten, then schools. More and more to introduce him to the human world. The result is close, now what?
When it will be possible to communicate with him at the level of a schoolboy, one can take his body from him and leave his mind. AI inputs and outputs can be connected to other things - return to the same streams of characters so that he can read books and chat. Give him the choice of what will serve him with his eyes - whether it be a movie, or the webcam of your laptop. In general, after the "school" of the "man" it is already possible to form the necessary result. At the “schoolchild” level, the computer should understand the person himself well enough, and not just his speech.
I believe that in this way it will be possible to "teach" the computer to experience emotions. And the more detailed the simulated world, the more the result will be similar to a person. Of course, you can give not a human body, and not human needs - say, instead of food - the need to grow a tree, but instead of the conditions of your own comfort - the comfort of a neighbor. As a result, a certain “alien” will come out - it will be difficult for him to understand us, if at all possible. But if you leave the human communication skill - it will be possible to communicate.
Of course, the need to eat arose probably even before the same monkey formed, if you believe that all living things came from unicellular ones. You can start modeling with unicellular ones - the result will be better (the problem of nutrition, comfort and reproduction will be sewn in the "natural" way), but it will require much more time, and generally does not guarantee that it will look like a person (hello UFO).
These are such thoughts out loud. I myself will not pull such a project, therefore, I give the idea to you. Perhaps among you there will be those who catch on this idea and assemble a team for its implementation. Maybe someone will even see financial potential in this project. Of course, the result will come not in a week, and not in a month - it can take years to educate. But the "schoolboy" can be saved, and used as a base for the "descendants".
If you have questions, I am ready to share my thoughts on this topic.
Idea
The idea is to grow AI as a small child, and do it as detailed as possible. A machine does not understand a person simply because from its “point of view” everything “looks” completely different - on the example of Alice, its world is a set of symbols, incoming and outgoing flows.
I propose to show the human world to a computer - in the most similar light. So that he can see, hear, touch, he can even smell. To give him a human body, but not physical - it will cost too much and difficult, but simulated.
Engine
Imagine a toy, like Sims (I didn’t play it myself, judging by the screenshots), in which AI controls one body. I do not think that there is a need to describe this body in great detail - but some nuances are important. For example, vision: on the AI input, send pictures from two eyes (not the player’s eye, the “camera”, but the body’s eye, that is, the picture that you would see if you were in its place) to be able to get a three-dimensional representation. There’s a minimum of sensors from the arms, legs, and other parts of the body - so that he can feel if he hits the back of his head, or understand when he grabbed his glass with his fingers so as not to squeeze his fingers further and break it. The way out, as you already understood, is to control various muscles - arms and legs, eyes (to simplify the neural network, you can lower the picture resolution, but add the “zoom” feature), and so on.
Immediately make a reservation about a broken glass - I do not claim that the physics of destruction is a necessity in this model, but I believe that its presence will allow the computer to better understand a person.
Birth
Of course, if you take it this way and immerse the neural network in such a body, the result will be zero. Too much chaotic data, and one generation will not play any role here. Therefore, evolution is necessary - to begin with, at least teach a “person” to eat (one of the sensors will be a hunger sensor, AI will receive data from it, but initially will not know how to react to it). I think many generations will die while learning to eat (at this period, generations will have to be propagated manually). Then you need to teach at least a little control of your body - for example, move to food (at least crawling). I believe that after such "primitives", we will get a creature that looks like a baby. It should have basic needs - food, comfort, reproduction. And this baby needs to be educated further - to teach him to walk, talk. Give it to some kind of kindergarten, then schools. More and more to introduce him to the human world. The result is close, now what?
"Schoolboy"
When it will be possible to communicate with him at the level of a schoolboy, one can take his body from him and leave his mind. AI inputs and outputs can be connected to other things - return to the same streams of characters so that he can read books and chat. Give him the choice of what will serve him with his eyes - whether it be a movie, or the webcam of your laptop. In general, after the "school" of the "man" it is already possible to form the necessary result. At the “schoolchild” level, the computer should understand the person himself well enough, and not just his speech.
Result?
I believe that in this way it will be possible to "teach" the computer to experience emotions. And the more detailed the simulated world, the more the result will be similar to a person. Of course, you can give not a human body, and not human needs - say, instead of food - the need to grow a tree, but instead of the conditions of your own comfort - the comfort of a neighbor. As a result, a certain “alien” will come out - it will be difficult for him to understand us, if at all possible. But if you leave the human communication skill - it will be possible to communicate.
Detail
Of course, the need to eat arose probably even before the same monkey formed, if you believe that all living things came from unicellular ones. You can start modeling with unicellular ones - the result will be better (the problem of nutrition, comfort and reproduction will be sewn in the "natural" way), but it will require much more time, and generally does not guarantee that it will look like a person (hello UFO).
Afterword
These are such thoughts out loud. I myself will not pull such a project, therefore, I give the idea to you. Perhaps among you there will be those who catch on this idea and assemble a team for its implementation. Maybe someone will even see financial potential in this project. Of course, the result will come not in a week, and not in a month - it can take years to educate. But the "schoolboy" can be saved, and used as a base for the "descendants".
If you have questions, I am ready to share my thoughts on this topic.