Why am I sure you can create a strong AI?
In the wake of AlphaGo's victory and robotic demonstrations , the topic of artificial intelligence became popular again, and I want to share my thoughts - why I’m sure that soon we will be able to create a "strong AI" (AI, fully replacing the person).
I graduated from the cybernetics department - I had applied mathematics with a bias in statistics. I started programming when it was not yet “fashionable”, and only those who really liked it programmed. My hobby is recursive algorithms, parsers, static code analysis and much more associated with a high level of abstraction.
By nature, I am the one who closely watches other people (a little later about this). At some point, I seriously took up practical psychology: I passed and conducted trainings, was engaged in NLP, and coaching. My systems thinking helped me a lot in understanding myself and other people.
')
I know that not only such is at the junction of “physics” and “lyricism” - but still this is quite a rare combination: “techies” are often poorly versed in psychology, and psychologists usually become humanists who, to put it mildly, are not used to operate with strict by categories.
One of the key differences between people from each other is "sorting the world." We are born with different initial settings for perceptual filters, and over the years we accumulate different subjective experiences from the same world around us. Roughly speaking: one in the head has the files sorted by name, the second by size, and the third by creation date. Therefore, when the third one says the “obvious” fact for himself that there have been fewer files lately, it turns out to be completely unobvious for the first and second. And you can not prove it.
This moment alone often explains why it would seem that smart people say short-sighted things, such as the man from the US Patent Office in 1899, who said that "everything that could be invented was invented." I don’t know exactly what happened specifically in his case, but I can easily imagine a modern investor who receives two dozen offers every day to finance another social network, another aggregator of something, or another fitness bracelet. He can also decide that "ideas for startups have been exhausted."
In this article - a lot of my subjective conclusions from the observations of people and my understanding of what is happening in our heads.
I will begin with the most important conclusion: the complexity of the human psyche is greatly overvalued.
Most practicing psychologists will confirm that people come to them with typical problems. Quite rarely there are people with significant disabilities or pathologies, and Billy Milligans are almost never. But there are many clients with questions “why I do not have the strength”, “how do I keep the relationship” and other popular requests, and every second one is sure that his case is unique.
For many years, two or three thousand “recipes” have accumulated in my head, which “cover” 99% of cases of appeals to a psychologist.
The late Roman Trachtenberg led a radio show in which people started telling jokes, and he finished. Of course, he did this in part because of his memory and its features. But at the same time, most of the jokes are about situations in relations between people, and there are just not so many of them. So they are stored in my head, and for almost any situation I can pick up an anecdote "in the subject." I think that among the readers of this article, many also can.
In typical everyday situations, people are quite predictable. Plots of works (read: stories about people) are well grouped into seven main groups . Some connoisseurs of people (for example, swindlers) live off the pattern of human behavior. Once a gypsy described me my character in the twenty seconds that I passed by (and I, frankly, was too lazy to answer her character, although it would be funny :))
The general scheme of such "magic" is simple - to determine the "type" of a person and the "weights" of his goals, using the "database of people" in his head. You will be surprised to know how quickly this can be done and how accurately it turns out to predict human behavior in the usual "domestic" environment.
Summing up: the behavior of people in most life situations is a simple and predictable thing that can be set by the program.
The most successful operating model for the psyche that I have met is a model about “parts”. Imagine that many independent processes are running in our heads at the same time: “be in comfort”, “learn new things”, “save energy”, “multiply”, etc. Some of them are “system”. Priorities are floating (“Wow! New iPhone! To go nuts! I want, I want, I want!” -> new excitation center -> priority: the highest). Many processes - conflict with each other (stability - risk; harmful, but fun, etc.). Some processes - started and forgotten, but they exist (“Do not talk to unfamiliar uncles and aunts on the street!”). Processes can be realized and completed, and these are engaged in various spiritual and psychological practices.
For any of our actions, you can dig out one or more of these "parts" that stand behind this action. The task of coaches and psychologists is usually just to determine which particular “parts” stand behind a particular human problem, using logic and observation. The process itself is very much like debugging.
For a meaningful functioning of the system, it needs a set of goals. In humans, heuristically found the main goals (“self-preservation”, “reproduction”, “preservation of community / species”) and secondary goals (“desire for comfort”, “craving for change”, “craving for stability”, etc.). Our AI can, for example, not set “self-preservation” and “kill all humans” as targets, but add “thirst for knowledge”.
I will skip the story “why do we need emotions,” but I will say that, unlike stereotypical opinion, the system of emotions is fairly well formalized , even with diffusions ! For example, “I expected little, I received a lot - joy” or “I have little, he has a lot - envy”. Even in The Sims have embedded emotions .
Animals also have emotions, so it will be possible to practice on robo-cats.
With these categories are not very advanced. There is an opinion that “funny” and “beautiful” are some emergent properties of our psyche.
One of the main qualities of our brain - we are able to simulate situations and find relationships. I have been “wearing” for a long time with the idea that humor is a kind of minimax in our brain. We are “given” the situation - our brain builds some kind of model, its own understanding of this situation - the “salt” of the joke goes on: it turns out that we all understood correctly or almost everything (the minimal change in the initial data), but our understanding is our model turned out to be very far from reality (maximum difference). The more minimax - the more emotion. If we are safe - we are funny, if not - scary.
And someone is building a more serious theory.
With “beauty,” there are similar observations based on optimality of movement / form. But the concept of “beauty” is very subjective and varies greatly even among people.
In any case, these things can be taught already functioning AI mathematically, neural networks or something else.
Here will be a little tricky.
How do we teach AI to create fundamentally new things, incl. make inventions?
I see several approaches:
- Peeped in nature, figured out, did;
- Applied a combination of known methods, as in TRIZ ;
- Applied a combination of meta-methods (methods for creating methods), created a new method, created something new;
- Random search.
I think that this is exactly what people use in their heads to create something new.
So, we create a computer with sensors and the ability to operate in the real world, give it a basic algorithm for modeling the world and ourselves in it (that's you and self-awareness), set goals in some primitives, launch parallel processes to strive for these goals, and we train, we train, we train ...
The most serious objection to the possibility of creating an AI is with Roger Penrose . It seems to me that his work on AI is a vivid example of the saying “in theory there is no difference between theory and practice, but in practice it is.” Realbtr wrote very well: "... Penrose idealizes the intellect, and moves away from the real task into the realm of pleasant ideas."
So, the famous "No-go" theorem of Penrose says that "No matter how powerful a device having the architecture of a finite state machine has, human thinking has some possibilities that are inaccessible to such a device." It is based on the Gödel incompleteness theorem: “Within the framework of a formal system built on axioms, we can formulate statements that cannot be proved within this system” (I simplified it a little without losing meaning).
So, Penrose for some reason believes that people can cope with the situation when they encounter such unprovable claims, but computers do not. And from this it concludes that people have a tricky system in their heads that helps to solve “incomprehensible” problems.
And I am sure that people here are not fundamentally different from robots. At the moment there is a set of axioms, on the basis of which many theorems are proved. And many other theorems have not yet been proved or disproved.
According to Gödel's incompleteness theorem, some of them cannot be proved without the introduction of new axioms that people successfully and gradually do. The AI ​​could do the same.
For the sake of interest, I went and refreshed my knowledge of the axioms in modern mathematics:
- The axioms of arithmetic (Peano)
- Axioms of real numbers
- Axioms of set theory
- ...
If all axioms in arithmetic and in the theory of real numbers are a formalization of more or less obvious things that coincide with my life experience, then already in set theory there is “the same” axiom of choice , which is not at all obvious, but it had to be introduced to substantiate part of the theoretical results.
The main question is: can the AI ​​add axioms to its formal system? And I see no reason why not. The question boils down to creativity, which I sorted out a little higher.
Another “problem” from Penrose’s work is the famous “stopping task”: a computer cannot always tell whether a particular program will work forever or ever stop.
Again, Penrose writes that people "can cope with this, but no AI can."
I worked a lot with the code of strangers (on legacy projects) and worked a lot on reverse engineering. So, very often I can only say something about someone else's code. The RIPEMD-160 algorithm in the firmware of the car navigator, after optimization by the compiler (with the cycles expanded), is a couple of thousand lines of assembler, very similar to each other. To complicate things a little bit, I won’t be able to determine in the near future whether it will get stuck or not. But I can give some kind of conclusion with a certain degree of confidence. And I am sure that the computer could already cope with this task better than me - to issue such conclusions with a certain degree of confidence.
Alexander Panov, in his articles, gives an example of a task that will certainly someday stop, but very soon - “to type 2 ^ 2 ^ 2 ^ 16 ones”.
But my task is to prove that there is no such byte sequence S for which:
md5 (S) = 1bc29b36f623ba82aaf6724fd3b16717 (= md5 ('md5') - 1)
To do this, you must either go through a lot of combinations in search of a counterexample, or open some new method to analyze the function md5. I wonder how and why “the person will cope with this, but the AI ​​will not.”
For me, indirect evidence that you can create a device better than our brain - these are other examples from the structure of the human body, for example - the eyes . Helmholtz wrote about the structure of our eyes: “If an optician would like to sell me an instrument that would have so many such defects, I would consider it fully justified to accuse him of negligence in strong expressions and would bring the instrument back.”
When I do really intellectual work - I get distracted, forget details, confuse facts, fall asleep. I am sure that all this can be done more efficiently.
Many believe that quantum computations occur in our brain. I think so too. At the same time, quantum computations, as I understand them, although they represent a qualitatively different way of solving problems, do not fundamentally affect the ability or impossibility to create a strong AI. Traditional computers are already somehow coping with speech recognition - the same network with the same parameters recognizes both English and Chinese speech equally well .
My subjective experience of a programmer and a psychologist tells me that most of the tasks in the project of creating a strong AI by people have already been done in one form or another. The rest are either simpler than they seem, or people can deal with them in the course of work.
Although maybe I'm just a naive amateur, but somewhere serious professional uncles have already created a strong AI or have proven that this cannot be done.
A few words about me
I graduated from the cybernetics department - I had applied mathematics with a bias in statistics. I started programming when it was not yet “fashionable”, and only those who really liked it programmed. My hobby is recursive algorithms, parsers, static code analysis and much more associated with a high level of abstraction.
By nature, I am the one who closely watches other people (a little later about this). At some point, I seriously took up practical psychology: I passed and conducted trainings, was engaged in NLP, and coaching. My systems thinking helped me a lot in understanding myself and other people.
')
I know that not only such is at the junction of “physics” and “lyricism” - but still this is quite a rare combination: “techies” are often poorly versed in psychology, and psychologists usually become humanists who, to put it mildly, are not used to operate with strict by categories.
I'll start from afar
One of the key differences between people from each other is "sorting the world." We are born with different initial settings for perceptual filters, and over the years we accumulate different subjective experiences from the same world around us. Roughly speaking: one in the head has the files sorted by name, the second by size, and the third by creation date. Therefore, when the third one says the “obvious” fact for himself that there have been fewer files lately, it turns out to be completely unobvious for the first and second. And you can not prove it.
This moment alone often explains why it would seem that smart people say short-sighted things, such as the man from the US Patent Office in 1899, who said that "everything that could be invented was invented." I don’t know exactly what happened specifically in his case, but I can easily imagine a modern investor who receives two dozen offers every day to finance another social network, another aggregator of something, or another fitness bracelet. He can also decide that "ideas for startups have been exhausted."
In this article - a lot of my subjective conclusions from the observations of people and my understanding of what is happening in our heads.
I will begin with the most important conclusion: the complexity of the human psyche is greatly overvalued.
People's behavior
Most practicing psychologists will confirm that people come to them with typical problems. Quite rarely there are people with significant disabilities or pathologies, and Billy Milligans are almost never. But there are many clients with questions “why I do not have the strength”, “how do I keep the relationship” and other popular requests, and every second one is sure that his case is unique.
For many years, two or three thousand “recipes” have accumulated in my head, which “cover” 99% of cases of appeals to a psychologist.
The late Roman Trachtenberg led a radio show in which people started telling jokes, and he finished. Of course, he did this in part because of his memory and its features. But at the same time, most of the jokes are about situations in relations between people, and there are just not so many of them. So they are stored in my head, and for almost any situation I can pick up an anecdote "in the subject." I think that among the readers of this article, many also can.
In typical everyday situations, people are quite predictable. Plots of works (read: stories about people) are well grouped into seven main groups . Some connoisseurs of people (for example, swindlers) live off the pattern of human behavior. Once a gypsy described me my character in the twenty seconds that I passed by (and I, frankly, was too lazy to answer her character, although it would be funny :))
The general scheme of such "magic" is simple - to determine the "type" of a person and the "weights" of his goals, using the "database of people" in his head. You will be surprised to know how quickly this can be done and how accurately it turns out to predict human behavior in the usual "domestic" environment.
Summing up: the behavior of people in most life situations is a simple and predictable thing that can be set by the program.
How to implement?
The most successful operating model for the psyche that I have met is a model about “parts”. Imagine that many independent processes are running in our heads at the same time: “be in comfort”, “learn new things”, “save energy”, “multiply”, etc. Some of them are “system”. Priorities are floating (“Wow! New iPhone! To go nuts! I want, I want, I want!” -> new excitation center -> priority: the highest). Many processes - conflict with each other (stability - risk; harmful, but fun, etc.). Some processes - started and forgotten, but they exist (“Do not talk to unfamiliar uncles and aunts on the street!”). Processes can be realized and completed, and these are engaged in various spiritual and psychological practices.
For any of our actions, you can dig out one or more of these "parts" that stand behind this action. The task of coaches and psychologists is usually just to determine which particular “parts” stand behind a particular human problem, using logic and observation. The process itself is very much like debugging.
Goals
For a meaningful functioning of the system, it needs a set of goals. In humans, heuristically found the main goals (“self-preservation”, “reproduction”, “preservation of community / species”) and secondary goals (“desire for comfort”, “craving for change”, “craving for stability”, etc.). Our AI can, for example, not set “self-preservation” and “kill all humans” as targets, but add “thirst for knowledge”.
Emotions
I will skip the story “why do we need emotions,” but I will say that, unlike stereotypical opinion, the system of emotions is fairly well formalized , even with diffusions ! For example, “I expected little, I received a lot - joy” or “I have little, he has a lot - envy”. Even in The Sims have embedded emotions .
Animals also have emotions, so it will be possible to practice on robo-cats.
Humor and beauty
With these categories are not very advanced. There is an opinion that “funny” and “beautiful” are some emergent properties of our psyche.
One of the main qualities of our brain - we are able to simulate situations and find relationships. I have been “wearing” for a long time with the idea that humor is a kind of minimax in our brain. We are “given” the situation - our brain builds some kind of model, its own understanding of this situation - the “salt” of the joke goes on: it turns out that we all understood correctly or almost everything (the minimal change in the initial data), but our understanding is our model turned out to be very far from reality (maximum difference). The more minimax - the more emotion. If we are safe - we are funny, if not - scary.
And someone is building a more serious theory.
With “beauty,” there are similar observations based on optimality of movement / form. But the concept of “beauty” is very subjective and varies greatly even among people.
In any case, these things can be taught already functioning AI mathematically, neural networks or something else.
Creation
Here will be a little tricky.
How do we teach AI to create fundamentally new things, incl. make inventions?
I see several approaches:
- Peeped in nature, figured out, did;
- Applied a combination of known methods, as in TRIZ ;
- Applied a combination of meta-methods (methods for creating methods), created a new method, created something new;
- Random search.
I think that this is exactly what people use in their heads to create something new.
Putting it all together
So, we create a computer with sensors and the ability to operate in the real world, give it a basic algorithm for modeling the world and ourselves in it (that's you and self-awareness), set goals in some primitives, launch parallel processes to strive for these goals, and we train, we train, we train ...
Theoretical limits
The most serious objection to the possibility of creating an AI is with Roger Penrose . It seems to me that his work on AI is a vivid example of the saying “in theory there is no difference between theory and practice, but in practice it is.” Realbtr wrote very well: "... Penrose idealizes the intellect, and moves away from the real task into the realm of pleasant ideas."
So, the famous "No-go" theorem of Penrose says that "No matter how powerful a device having the architecture of a finite state machine has, human thinking has some possibilities that are inaccessible to such a device." It is based on the Gödel incompleteness theorem: “Within the framework of a formal system built on axioms, we can formulate statements that cannot be proved within this system” (I simplified it a little without losing meaning).
So, Penrose for some reason believes that people can cope with the situation when they encounter such unprovable claims, but computers do not. And from this it concludes that people have a tricky system in their heads that helps to solve “incomprehensible” problems.
And I am sure that people here are not fundamentally different from robots. At the moment there is a set of axioms, on the basis of which many theorems are proved. And many other theorems have not yet been proved or disproved.
According to Gödel's incompleteness theorem, some of them cannot be proved without the introduction of new axioms that people successfully and gradually do. The AI ​​could do the same.
For the sake of interest, I went and refreshed my knowledge of the axioms in modern mathematics:
- The axioms of arithmetic (Peano)
- Axioms of real numbers
- Axioms of set theory
- ...
If all axioms in arithmetic and in the theory of real numbers are a formalization of more or less obvious things that coincide with my life experience, then already in set theory there is “the same” axiom of choice , which is not at all obvious, but it had to be introduced to substantiate part of the theoretical results.
The main question is: can the AI ​​add axioms to its formal system? And I see no reason why not. The question boils down to creativity, which I sorted out a little higher.
Another “problem” from Penrose’s work is the famous “stopping task”: a computer cannot always tell whether a particular program will work forever or ever stop.
Again, Penrose writes that people "can cope with this, but no AI can."
I worked a lot with the code of strangers (on legacy projects) and worked a lot on reverse engineering. So, very often I can only say something about someone else's code. The RIPEMD-160 algorithm in the firmware of the car navigator, after optimization by the compiler (with the cycles expanded), is a couple of thousand lines of assembler, very similar to each other. To complicate things a little bit, I won’t be able to determine in the near future whether it will get stuck or not. But I can give some kind of conclusion with a certain degree of confidence. And I am sure that the computer could already cope with this task better than me - to issue such conclusions with a certain degree of confidence.
Alexander Panov, in his articles, gives an example of a task that will certainly someday stop, but very soon - “to type 2 ^ 2 ^ 2 ^ 16 ones”.
But my task is to prove that there is no such byte sequence S for which:
md5 (S) = 1bc29b36f623ba82aaf6724fd3b16717 (= md5 ('md5') - 1)
To do this, you must either go through a lot of combinations in search of a counterexample, or open some new method to analyze the function md5. I wonder how and why “the person will cope with this, but the AI ​​will not.”
Other arguments FOR
For me, indirect evidence that you can create a device better than our brain - these are other examples from the structure of the human body, for example - the eyes . Helmholtz wrote about the structure of our eyes: “If an optician would like to sell me an instrument that would have so many such defects, I would consider it fully justified to accuse him of negligence in strong expressions and would bring the instrument back.”
When I do really intellectual work - I get distracted, forget details, confuse facts, fall asleep. I am sure that all this can be done more efficiently.
Many believe that quantum computations occur in our brain. I think so too. At the same time, quantum computations, as I understand them, although they represent a qualitatively different way of solving problems, do not fundamentally affect the ability or impossibility to create a strong AI. Traditional computers are already somehow coping with speech recognition - the same network with the same parameters recognizes both English and Chinese speech equally well .
Summarizing
My subjective experience of a programmer and a psychologist tells me that most of the tasks in the project of creating a strong AI by people have already been done in one form or another. The rest are either simpler than they seem, or people can deal with them in the course of work.
Although maybe I'm just a naive amateur, but somewhere serious professional uncles have already created a strong AI or have proven that this cannot be done.
Source: https://habr.com/ru/post/392165/
All Articles