What is intelligence?
A model is a system whose study serves as a means for obtaining information about another system; representation of some real process, device or concept.Wiki
The model simulates the simulated system with some accuracy. The greater the accuracy, the more objects will be in the model. The information model has information objects. At maximum accuracy, for each simulated object of the system there will be a corresponding informational one. That is, there will be a one-to-one mapping. In the future, under the word “model” I will mean the mapping of some objects into others, a given correspondence between the real and the information.
An information system is a system for processing information. Information in some way arrives at the input of the information system. To properly handle it, the system has an information model.
Where does the information come from? From some external source. I call it "reality."
Reality is all that is outside the information system.
So, in order to interact with reality, an information system must have some model of this reality. For each real object there must be an informational corresponding to it, that is, a model of this object. Another information system is also a real object. Accordingly, if we want to somehow interact with it, we also need a model for it.
The need for such a model is clearly visible when developing an API. On the server, there is a certain structure on the output; on the client, the same structure on the input. In PHP and in JS, the same classes. This is not redundancy; it is a simulation of one system to another.
Natural intelligence is also an information system. He has sensors that give incoming information. Based on this information, he builds a model of what happens outside this information system, and on the basis of the model makes decisions about actions. Thus, we can give a definition of what intelligence is.
Intellect is the ability of an information system to build a model of reality based on incoming information.
A computer program is also an information system. Does she have intelligence? No, because the model is programmed into the program. Consider for example a program for partitioning a hard disk. Information about possible formats, the order of command execution, response codes - this is all embedded in the source code. The program has a hard disk model; it is updated when receiving responses to commands, but the program does not build it, cannot change it.
Artificial intelligence must build the correct model of reality. In particular, he must respond to it correctly. At least the same as a man. First of all, this is a feedback for us, according to which we can come to the conclusion that the system has intelligence. This can be checked in the debugger, but you need to know what should be there. Secondly, because one of the goals is the implementation of human activities in some tasks.
All animals somehow react to reality. Therefore, with regard to them, it is impossible to talk about whether there is intelligence or not intelligence. You can talk about whether it is more or less, whether it is enough for a certain type of tasks to build certain models.
Unicellulars have almost no intelligence, because there are no elements that store and process information about the environment. There are only chemical interactions that allow them to function. They can also be considered a model, but not informational, but hardware. Due to its small size and chemical laws, it cannot be complicated.
It is also worth noting that intelligence is closely related to the concepts of object and memory. An object is something that has a state and a behavior, which we define as one and the same at different points in time. Memory is needed to store the attributes of objects between these moments; here the model begins. But objects are a topic for another conversation.
People experience pleasant and unpleasant feelings. Consider the sensation of pain. How can you describe it? Pain is an unpleasant feeling, if it is present, then the brain constantly pays attention to it. Usually, a person cannot feel or not feel pain at will. In this case, people have several unpleasant feelings, for example, there is still fear.
For an arbitrary information system you can give such a definition.
Sensations are a process of perception of data that enter the information system constantly, they require their processing, and the system is designed so that it cannot but react to them. Perception is how data influences a model. Their presence can be perceived as positive or negative feedback. It tends to increase the positive relationship, to reduce the negative.
Any signals from the sensors are sensations.
Can artificial intelligence experience pain?
First, we note that the natural intellect experiences several different sensations for it. Pain is just the name of one of them. It signals damage. Thus, if a certain intelligence information system gets a sensation with damage information, which it defines as unpleasant for it, seeks to minimize its influence and cannot do it by turning it off, then this is analogous to pain. It is an analogue, we can only define the similarity that “Sensation 1” is similar to human pain, and “Sensation 2” is like pleasure from something. The main thing here is how the system itself reacts to this data. How to determine that the AI is not lying? Check in debugger.
The more signs, the greater the similarity, including chemical and electrical processes. Most likely, animals feel sensations similar to human ones, since they react in the same way as humans, they also have nerve cells, which also consist of organic compounds.
That is, if we make an information model of a person with an accuracy of atoms and electric fields, as well as the surrounding reality with which he interacts, and simulate a cut on a finger, then yes, in this reality he will experience pain. If with the help of sensors it will interact with our reality, then in ours. As in terms of information processing, it will not be different from other people.
Is it ethical to make him feel pain against his desire in any reality? I think no. This applies to any other creature and any unpleasant sensation. But since this is an information model, it is possible to simulate a different reaction or another way of perception, and then the level of ethical will be different.
Consider a database management system with monitoring. If a failure occurs, the monitoring subsystem begins to send a notification with an alarm level to the administrator at some intervals. Can we say that such a system is in pain? Not. First, it does not have intelligence, the model is embedded in it by the programmer, and there is no concept similar to pain. Secondly, although the signal is rated as unpleasant, it does not have the desire to minimize it. In general, even on the contrary, it is made in order to provide the most accurate model of what is happening with the system. It can be called an analogue of nerve cells, not the creature that has them.
Consider a bot with AI in a computer game. The game for him is a reality, albeit a very limited one. Damage information is required for proper action. If the bot receives it at will - performed some actions, paid attention to the wound, began to receive signals about the damage, remembered their parameters, got distracted, stopped receiving - then this is not an analogy of pain. If the bot perceives it as additional information, and not as a negative sensation, the level of which should be minimized, if it does not affect actions taken for other purposes (for example, there is no desire to withdraw the hand, if it received a blow to the wound on the arm), then not.
What is consciousness? It can be defined as the process of functioning of the intellect, the process of obtaining information and updating the model of reality. It is closely related to the concept of "I". “I” is an element in the model that represents the information system itself. Is there a notion of "I" in dogs? Yes, this is the object with which they associate the "Sit-Lie" commands. This means that dogs have consciousness and intelligence. But for example, dogs cannot multiply. Perhaps this is due to the fact that for higher levels of abstraction additional information elements are needed, which will observe lower levels and build their model.
Consciousness is associated with self-observation. That is, there is not just the receipt of information about the world and the updating of the model, but also the registration of the facts of receipt themselves, their reflection in the model, the connection with the object “I”. "If I poke under the ribs, I tickled." "Yesterday I did not know this, today I know." That "I think - therefore I exist." It turns out a closed system, part of the elements from the output of the analyzing unit is transmitted to its input. It can be assumed that only complex beings can possess consciousness, who have enough information elements to organize such a connection.
What is understanding? What does it mean when we say that a person understood something? This means that he has the right model of what he analyzed. The information elements of the model correspond to the elements of the real system. Understanding is building the right model. Why is experience so important? This is a model. It allows you to simulate the processes in the information form, and therefore plan and select the desired option.
There are neural networks, they themselves build a model according to the incoming information. Do they build the right model? The results show that somehow not very. They usually require a special sample, this is also a prepared model. But they can build some kind of model, so they have some level of intelligence.
Update: An important point. Objects in the model are objects with a state and behavior that are recognized as the same at different points in time. I think if the information system does not have a mechanism for the notion of the “same”, then it cannot be said that it has intelligence.
I decided to add an analysis of the Chinese room in order to show the applicability of this approach to the resolution of incomprehensible questions.
So, there is a room, a person is sitting in it, he has an instruction, guided by which he shifts the hieroglyphs. The room answers Chinese questions, but the person does not know Chinese. The answer is simple and has already been given earlier - the Chinese language is known to the whole system. But there are arguments against it. Consider some points.
“The instruction is composed in such a way that after applying all the steps to the hieroglyphs of the question, they are converted to the hieroglyphs of the answer.”
Strictly speaking, this is impossible, since the answer requires additional data that are not in the question. If the system does respond correctly, they are stored somewhere. This is her model of reality.
“Formally, the argument can be represented as follows:
1. If the hypothesis of a“ strong ”AI is correct, then there is such a program for Chinese writing, when launched, in the computing system this system will understand Chinese writing.
2. I can execute a program for Chinese writing without understanding this writing.
3. Consequently, the hypothesis of a "strong" AI is incorrect. "
Clause 1 should look like this: " there is such a program for Chinese writing, when launched, on the computer system, this program will understand Chinese writing . " The system sends data to the program, the program processes it.
Therefore, from 2 it does not follow 3. Understanding Chinese script will be an executable program, not the one who executes it.
In order for a person to begin to understand Chinese, it is necessary to expand his model’s model from the program. And when this data is inside a virtual machine, it is inaccessible.
“Searle’s more general conclusion is that any manipulation of syntactic constructions cannot lead to an understanding.”
True. Understanding is building the right model. Syntactic constructions only designate elements of the model.
“The system from the rule book, the person and the room, in their opinion, is reasonable and understands Chinese. As a counter-argument, Searle suggests making a person remember the formal answer rules for the system. ”
And when he performs the program, she will understand Chinese.
Article in English with a few additions: What is intelligence?