What to feed AI (end)
In this post I would like to consider only typical mistakes, when thinking about how to start doing AI. The topic does not claim to the truth in the last resort, only trying to indicate obvious errors.
By AI, I mean a program that can perform different types of tasks, leading them to the same type, which they can already solve.
A person receives by sight millions of bits per second, touch gives a person hundreds of thousands of bits per second, and with the help of hearing a person receives tens of thousands of bits per second.
The remaining channels of sensations provide an incomparably smaller amount of information, are very difficult to implement, slow and do not have the opportunity to communicate (it is hardly possible to build an AI on the basis of smell).
')
Let me explain: human children raised in a pack of wild animals have a markedly lower level of development, as a rule, cannot speak. People who grew up in urban settings are more likely to have a high IQ than the villagers. From this we can make the assumption that the AI for its normal development will be very voracious to the information filled with meaning.
For example, only reading books can hardly be compared with communicating with people, because books do not provide feedback. A more subtle example: inscriptions in foreign languages: even after reading them a hundred times to understand their meaning is impossible. Although if you try to stroke the snake, and your guide-aboriginal shouts something, you will quickly understand what he meant.
It is desirable, of course, in real time. Any delay reduces the chance to draw an analogy and link two events. Especially if the intervals are different. This rule is not necessary, but rather simply follows from the previous two. Since the information can have feedback and be extensive only if it is relevant. Of course, you can communicate with letters the size of a couple of gigabytes once a day, with a conscious interlocutor, but it will be extremely difficult to identify the relationship of honey with concrete actions.
In my opinion, these rules are common to any AI, both algorithmic and neural. Both for real and for consciousness-in-the-matrix.
Based on the theory, it can be assumed that the AI will need information in the amount of at least 1% -10% of what a person receives (the conclusion is based on the fact that blind people can still think). If there is less information, then the AI will most likely not be intelligent enough (a blind-deaf-dumb child, broken by paralysis and without tactile sensations, is unlikely to become a full-fledged person). Also, the AI should be able to get an answer to almost all of its actions (as an example, the fact that a person many times a second receives an extensive account of his own movements with the help of touch). A delay in the answers can complicate the identification of relationships and reduce the speed of learning, so it should be within seconds.
Consider the theoretical developments of the participants in the Habr:
OpenMinded - Creating AI based on vector graphics. Since there is no vector graphics in the real world, this AI will have serious difficulties to adapt to reality, if it can do it at all. As for the flow of information, it will take hundreds of people to train such an AI to provide at least megabits per second, each of which will have to quickly draw a response when receiving a message from the AI. And at this pace, the years until the AI progresses a bit.
P.S. This topic made me write this post.
The most common option, of course, is text. They wrote about him:
alizar - Alzmatchbotz suzette chat
krestjaninoff - Algorithmic analysis of morphology
gorcer - Algorithmic chatbot Marishko
pablozer - A program for writing algorithmic AI
The main problem of textual AI is information poverty. To feed the AI chat, you need to communicate with approximately one hundred thousand people at a time. Of course, you can feed him all the textual heritage of humanity, but this is just a static text, they can be taught the program to make sentences, but not understand their meaning.
The chess AI about which Latobco wrote:
This AI has the same problems with the saturation of information, during the course of every 2 minutes you need 300 thousand opponents all the time. Not to mention that the world in which AI exists is extremely limited in the number of available options for action.
To give the necessary amount of information can:
Video - if you, of course, can process such a flow of information, provide feedback, and give it some meaning. Your AI wallpaper is unlikely to inspire,
Tactile sensations - if you can build a tactile sensor,
Sound - here, too, there must be some difficulties
Or any other constant, fast, extensive, varying source of answers.
Before you start creating an AI, consider carefully whether you can provide the AI with your information.
Vision - 10 ^ 6 bps
Touch - 10 ^ 5 bps
Hearing - 10 ^ 4 bps
The first 3 years of life (12 hours of sleep) - 5.25 * 10 ^ 13 bits
Blind child, the first 3 years - 5.2 * 10 ^ 12 bits = 650 GB of information
Approximately it is necessary to feed the AI as a whole in order for it to have some chance of self-awareness.
P.S. If someone has any ideas, additions, or you notice a mistake, you are welcome to discuss
The ideas and additions of the collective mind follow below.
(Question group: Man: “his quote ...” is my subjective opinion):
igudym:
"At the level of thinking comes already structured information of a much smaller volume."
Yes, much smaller amounts of information will reach deep layers of artificial intelligence, but the ability to systematize information and sift out unnecessary will necessarily be present.
OpenMinded:
“The main thing is that the information is logical and consistent, otherwise there will be indigestion.”
Yes, for artificial intelligence the saturation of information, its complexity is just as important.
Unnamed:
“Intellect is a property of social organisms. "Adyn thing" create fail. Only community. ”
Artificial intelligence needs an environment with adequate, in terms of development, interlocutors.
dersd:
“The most primitive way of learning is solving problems to eliminate discomfort at the current time.”
Yes, artificial intelligence must have a certain set of organs (capable of sensing = sensors) that may feel harmful effects (discomfort), the avoidance of which is the simplest action.
twdragon:
“So, the AI will have to not only be trained, but also conditioned, giving the input a mixture of useful information with“ water ”in order to train it to concentrate resources to solve a particular task.”
Yes, artificial intelligence does not need all the incoming signals critically, and many of them can be replaced with “water”, noise, but you need to carefully consider how to do this and leave the opportunity for the AI to concentrate on the details of interest.
OpenMinded:
"The laws of the world are much more important than the specification of its objects."
Biga:
“I believe that the creation of AI is possible without robotics, and without simulating the real world, but it will be necessary to communicate with such AI in an inhuman language. „
Yes, artificial intelligence can understand the laws of the physical world on the basis of a simplified model of reality such as the 3D world, but this is not enough to learn how to talk with a person on an equal footing.
Unnamed:
“Personally, I would generally extend the concept of intelligence to the entire nervous system, to the smallest ganglia and endings. And I would include more chemical subsystems of regulation. ”
Artificial intelligence is not only a square neural network, but in general, all neurons are either involved in processing or transmitting information. Systems that can have any sufficient effect on the neural network can also be considered as its parts (for example, humoral regulation).
OpenMinded:
"Complicated environment -> complex mental model -> complex behavior."
Artificial intelligence, to be complex enough, must exist in a complex world and be complex enough in structure to have complex behavior.
grokinn:
"Olga Ivanovna Skorokhodova was a blind and deaf and dumb woman, while a scientist and a writer ..."
Despite this, the intellect was formed before its loss of basic feelings (5 years), besides it is one of the few people with disabilities who have become known. Such a rare exception rather proves the rule about the need for a huge amount of information for consciousness.
mechmind:
"Ants pheromones ensure the coordinated work of their many thousands of" communities "."
klakhman:
“The communication of insects, for example, in a very large part occurs due to the secreted enzymes ...”
Yes, artificial intelligence based on smell is possible in theory,
but not feasible with our level of technology and due to the lack of possible interlocutors.
I thank everyone who has expressed his opinion.
I will look forward with interest to the response articles from twdragon about metaprints and from ami.
By AI, I mean a program that can perform different types of tasks, leading them to the same type, which they can already solve.
A bit of theory
A person receives by sight millions of bits per second, touch gives a person hundreds of thousands of bits per second, and with the help of hearing a person receives tens of thousands of bits per second.
The remaining channels of sensations provide an incomparably smaller amount of information, are very difficult to implement, slow and do not have the opportunity to communicate (it is hardly possible to build an AI on the basis of smell).
')
Argument one: the AI must have access to the content information.
Let me explain: human children raised in a pack of wild animals have a markedly lower level of development, as a rule, cannot speak. People who grew up in urban settings are more likely to have a high IQ than the villagers. From this we can make the assumption that the AI for its normal development will be very voracious to the information filled with meaning.
Argument Two: The AI Must Have Feedback
For example, only reading books can hardly be compared with communicating with people, because books do not provide feedback. A more subtle example: inscriptions in foreign languages: even after reading them a hundred times to understand their meaning is impossible. Although if you try to stroke the snake, and your guide-aboriginal shouts something, you will quickly understand what he meant.
Argument Three: AI must have access to timely information.
It is desirable, of course, in real time. Any delay reduces the chance to draw an analogy and link two events. Especially if the intervals are different. This rule is not necessary, but rather simply follows from the previous two. Since the information can have feedback and be extensive only if it is relevant. Of course, you can communicate with letters the size of a couple of gigabytes once a day, with a conscious interlocutor, but it will be extremely difficult to identify the relationship of honey with concrete actions.
In my opinion, these rules are common to any AI, both algorithmic and neural. Both for real and for consciousness-in-the-matrix.
Some practice
Based on the theory, it can be assumed that the AI will need information in the amount of at least 1% -10% of what a person receives (the conclusion is based on the fact that blind people can still think). If there is less information, then the AI will most likely not be intelligent enough (a blind-deaf-dumb child, broken by paralysis and without tactile sensations, is unlikely to become a full-fledged person). Also, the AI should be able to get an answer to almost all of its actions (as an example, the fact that a person many times a second receives an extensive account of his own movements with the help of touch). A delay in the answers can complicate the identification of relationships and reduce the speed of learning, so it should be within seconds.
Consider the theoretical developments of the participants in the Habr:
OpenMinded - Creating AI based on vector graphics. Since there is no vector graphics in the real world, this AI will have serious difficulties to adapt to reality, if it can do it at all. As for the flow of information, it will take hundreds of people to train such an AI to provide at least megabits per second, each of which will have to quickly draw a response when receiving a message from the AI. And at this pace, the years until the AI progresses a bit.
P.S. This topic made me write this post.
The most common option, of course, is text. They wrote about him:
alizar - Alzmatchbotz suzette chat
krestjaninoff - Algorithmic analysis of morphology
gorcer - Algorithmic chatbot Marishko
pablozer - A program for writing algorithmic AI
The main problem of textual AI is information poverty. To feed the AI chat, you need to communicate with approximately one hundred thousand people at a time. Of course, you can feed him all the textual heritage of humanity, but this is just a static text, they can be taught the program to make sentences, but not understand their meaning.
The chess AI about which Latobco wrote:
This AI has the same problems with the saturation of information, during the course of every 2 minutes you need 300 thousand opponents all the time. Not to mention that the world in which AI exists is extremely limited in the number of available options for action.
Few morals
To give the necessary amount of information can:
Video - if you, of course, can process such a flow of information, provide feedback, and give it some meaning. Your AI wallpaper is unlikely to inspire,
Tactile sensations - if you can build a tactile sensor,
Sound - here, too, there must be some difficulties
Or any other constant, fast, extensive, varying source of answers.
findings
Before you start creating an AI, consider carefully whether you can provide the AI with your information.
In numbers
Vision - 10 ^ 6 bps
Touch - 10 ^ 5 bps
Hearing - 10 ^ 4 bps
The first 3 years of life (12 hours of sleep) - 5.25 * 10 ^ 13 bits
Blind child, the first 3 years - 5.2 * 10 ^ 12 bits = 650 GB of information
Approximately it is necessary to feed the AI as a whole in order for it to have some chance of self-awareness.
P.S. If someone has any ideas, additions, or you notice a mistake, you are welcome to discuss
UPD3
The ideas and additions of the collective mind follow below.
(Question group: Man: “his quote ...” is my subjective opinion):
Input structure:
igudym:
"At the level of thinking comes already structured information of a much smaller volume."
Yes, much smaller amounts of information will reach deep layers of artificial intelligence, but the ability to systematize information and sift out unnecessary will necessarily be present.
OpenMinded:
“The main thing is that the information is logical and consistent, otherwise there will be indigestion.”
Yes, for artificial intelligence the saturation of information, its complexity is just as important.
Education and the world:
Unnamed:
“Intellect is a property of social organisms. "Adyn thing" create fail. Only community. ”
Artificial intelligence needs an environment with adequate, in terms of development, interlocutors.
dersd:
“The most primitive way of learning is solving problems to eliminate discomfort at the current time.”
Yes, artificial intelligence must have a certain set of organs (capable of sensing = sensors) that may feel harmful effects (discomfort), the avoidance of which is the simplest action.
twdragon:
“So, the AI will have to not only be trained, but also conditioned, giving the input a mixture of useful information with“ water ”in order to train it to concentrate resources to solve a particular task.”
Yes, artificial intelligence does not need all the incoming signals critically, and many of them can be replaced with “water”, noise, but you need to carefully consider how to do this and leave the opportunity for the AI to concentrate on the details of interest.
OpenMinded:
"The laws of the world are much more important than the specification of its objects."
Biga:
“I believe that the creation of AI is possible without robotics, and without simulating the real world, but it will be necessary to communicate with such AI in an inhuman language. „
Yes, artificial intelligence can understand the laws of the physical world on the basis of a simplified model of reality such as the 3D world, but this is not enough to learn how to talk with a person on an equal footing.
Internal structure:
Unnamed:
“Personally, I would generally extend the concept of intelligence to the entire nervous system, to the smallest ganglia and endings. And I would include more chemical subsystems of regulation. ”
Artificial intelligence is not only a square neural network, but in general, all neurons are either involved in processing or transmitting information. Systems that can have any sufficient effect on the neural network can also be considered as its parts (for example, humoral regulation).
OpenMinded:
"Complicated environment -> complex mental model -> complex behavior."
Artificial intelligence, to be complex enough, must exist in a complex world and be complex enough in structure to have complex behavior.
Other:
grokinn:
"Olga Ivanovna Skorokhodova was a blind and deaf and dumb woman, while a scientist and a writer ..."
Despite this, the intellect was formed before its loss of basic feelings (5 years), besides it is one of the few people with disabilities who have become known. Such a rare exception rather proves the rule about the need for a huge amount of information for consciousness.
mechmind:
"Ants pheromones ensure the coordinated work of their many thousands of" communities "."
klakhman:
“The communication of insects, for example, in a very large part occurs due to the secreted enzymes ...”
Yes, artificial intelligence based on smell is possible in theory,
but not feasible with our level of technology and due to the lack of possible interlocutors.
End
I thank everyone who has expressed his opinion.
I will look forward with interest to the response articles from twdragon about metaprints and from ami.
Source: https://habr.com/ru/post/113015/
All Articles