This article is reproduced from the public account @Meet the Future series author Teacher Wang. The full text is as follows:
The core concepts of the universe in which we live include space, time, and the arrow of time.
Space is defined by forces, gravity, electromagnetism, strong interactions, weak interactions define different spatial scales. Gravity defines large-scale space in the universe. Strong and weak interactions define the space within the nucleus. And the electromagnetic force defines most of the phenomena in our human living space.
Time is a measure of how quickly or slowly it changes. There is no time without change. There is no absolute time in the universe, and all changes are relative. Therefore, time also needs to be related to a fixed change, such as the day, month, and year are all related to the laws of the earth's motion.
The change itself does not define the direction of time. Reversible changes cannot define the direction of time, only irreversible changes can define the direction of time. The arrows of cosmic time in which we live come from increasing entropy, the second law of thermodynamics. Since the increase in entropy is inherently difficult to reverse, it causes the passage of time and forms an arrow of time.
The essence of productivity is to combat entropy increase, that is, the process of converting energy into an entropy reduction ability. The history of human civilization is the history of more efficient energy acquisition and more efficient conversion of energy into entropy reduction ability.
In this sense, money is a measure of entropy reduction capacity. Investment, on the other hand, is to exchange today's entropy reduction capacity for tomorrow's more efficient entropy reduction capacity.
The essence of increasing entropy is to increase uncertainty in the universe, while the essence of intelligence is to reduce uncertainty. So intelligence itself is a kind of productivity.
The essence of intelligence is to understand space, time, and use the understanding of space and time to combat entropy increase.
Single-celled organisms appeared on Earth 3.5 billion years ago, vertebrates appeared in the Cambrian period about 500 million years ago, and around this era, animals with eyes also appeared in large numbers. Erect walking hominins appeared about 5 million years ago, while the ancestor of modern humans, Homo sapiens, only appeared about 300,000 years ago.
Homo sapiens evolved language about 70,000 years ago, and relying on the strong organizational ability that language brings, Homo sapiens defeated other species on Earth, and humans on Earth today are descendants of Homo sapiens.
Human language represents universal intelligence. Human language is a very peculiar symbol system. The rules for this symbology are often called grammars. Grammar rules include syntactic rules, semantic rules, and speech rules. Unlike the way animals communicate, human language can express abstract concepts and events. For a signal communication system to be called a language, it needs to satisfy the arbitrariness of the association between symbols and meanings, the transfer of space and time, and the discreteness and generativeness.
There is a very interesting link in the evolution of animals, that is, mammals evolved into thermostatic animals, but the price paid is that the speed of metabolism is much higher, and mammals of the same weight need several times the food of reptiles to be able to maintain. To find more food, the mammalian brain underwent an evolution: the connection of memory and navigation systems. Their brains encode visual landscape features in the outermost neocortex and navigation in the entorhinal cortex. The entire system is connected to each other through a brain structure called the hippocampus.
Groups of neurons in the cerebral cortex encode memories of these objects and past events. Recalling an event or an experience reactivates the neurons that originally encoded it. All mammals may recall and re-experience previously encoded objects and events by reactivating these groups of neurons.
The next step is the ability to construct a "memory" that hasn't happened yet.
According to modern medical research, the most primitive form of human imagination of new objects and scenes is dreaming. This vivid, bizarre, uncontrollable fantasy all occurs during rapid eye movement (REM). Scientists speculate that animals with REM also dream.
The equivalent of language ability is the ability to deliberate, responsive, and reliable to combine and reconstruct mental objects, also known as prefrontal synthesis. It relies on the ability of the prefrontal cortex, located at the very front end of the brain, to control the rest of the neocortex. When did humans begin to acquire this ability? According to many archaeological sources, almost all human remains older than 70,000 years ago, such as petroglyphs, are completely realistic, depicting everything they see with simple lines. After that, many fictional images and objects appeared, such as lion-headed people, bone needles, bows and arrows, ornaments, and even totems, which are things that do not exist in nature.
Therefore, human imagination and creativity appeared almost at the same time as language. Since then, human civilization has entered a stage of rapid development.
The height of human intelligence can be represented by a work, that is, Wang Bo's Tengwang Pavilion Order. This ancient text is not only an outstanding representative of ancient Chinese literature, but also a top expression of human intelligence. This article shows the two main capabilities of intelligence. Wang Bo compressed the extremely large amount of data on historical events, geography and humanities, the atmosphere of the scene, and the beautiful scenery of water and sky into an article with a very small amount of information, which is an efficient compression of complex information. Later people imagined various scenes at that time based on this article, which was the process of restoring and upgrading the compressed information. Both of these processes are the main manifestations of intelligence.
Broadly speaking, language is a structured and logically consistent way of describing knowledge and theory. Whether the underlying data is text, speech, pictures, or video, it can be understood as a language.
Much of the work can be understood as the understanding of the language and the translation process from one language to another:
1. Translation: Chinese to English, etc
2. Programming: Translation from Chinese or English to a programming language
3. Painter: Translation from text to image
4. Composer: Translation from text to sheet music
5. Mathematician: Translation from text to mathematical language
So when Homo sapiens evolved language, humans entered the age of intelligence. Mathematics, physics, computers, and even general artificial intelligence are all inevitable things that will happen after the advent of language, but it is only a matter of time.
So how does artificial intelligence mimic human intelligence?
Artificial Intelligence (AI) - A theory and development of computer systems used to study systems capable of performing tasks that would normally require the participation of human intelligence.
The core problems of AI include reasoning, knowledge, planning, learning, communication, perception, the ability to move, and manipulate objects.
In addition to the ability to perceive and execute in physical space, artificial intelligence contains human intelligence, which can be divided into two kinds of capabilities: intuitive and logical.
Intuition is the ability to analyze the correlation between things, and humans have learned a correlation between input and output through a lot of experience, which is intuition. Logic pays attention to the cognitive process of using concepts, judgments, reasoning and other types of thinking to reflect the essence and laws of things.
The basic logical reasoning ability of human beings, since the emergence of computer systems, has been largely simulated by software programming.
The intuitive ability of human beings is gradually simulated by artificial intelligence algorithms after the emergence of neural networks, Internet big data, advanced GPU parallel computing frameworks, and deep learning network models.
Human intuition, from the mathematical model, is a nonlinear function of a high-dimensional space, usually a very complex nonlinear mapping from a high-dimensional space to a low-dimensional space. Let's see how a neural network simulates such a nonlinear map.
Perceptron is one of the simplest neural network models. It is the calculation of human neurons using a very simple mathematical formula. Perceptron has only two operations in each neuron, the point multiplication (inner product) operation of the input vector and the neuronal weight vector, and a nonlinear mapping function (activation function) of the point multiplication result. The point multiplication operation of vectors is essentially a similarity measurement, and the activation function gives the neural network the nonlinearity it needs. Since many human sensory organs can perceive a very large dynamic range (such as the brightness of light, the size of sound), this nonlinearity can effectively simulate the ability of humans to process wide dynamic input.
As you can see, each neuron is the simplest nonlinear filter, which selects the signal closest to the weight vector from the input and compresses its dynamic range. When countless of these neurons are interconnected, most of the complex nonlinear functions can be simulated, that is, human intuition can be simulated.
Large models represented by Transformer can already understand the internal structure of multimodal data, whether it is text, voice, pictures, or video, they can achieve very good results.
Emergence occurs when the number of parameters of a large model exceeds 100 billion. The generative large model represented by GPT4 can understand human language very well, and even has produced preliminary logical reasoning ability. This phenomenon can be understood from two points: 1. Linguistic intuition is a very complex nonlinear function, and it takes enough neurons to accurately simulate this nonlinear function. 2. The brain capacity of modern humans is about 1500ml, the brain capacity of Homo sapiens is about 1400ml, while the apes are only 400~500ml, and the Javan apes are about 900ml, and human intelligence begins to emerge after the brain capacity reaches a certain standard.
The degree of intelligence is determined by the degree of linguistic abstraction. The top math papers in the world can only be understood by a few people in the world. Large models can already understand basic human languages such as Chinese and English, and they also have a good understanding of programming languages. A preliminary understanding of very abstract languages such as mathematics began.
Transformer was born from the problem of translation between languages, and it has been shown to be good at understanding the internal structure of language, as well as the mapping problem between languages. This ability has reached the essence of intelligence, and the next step is how to improve the ability to understand more abstract languages.
Large models currently represent a weak general-purpose AI capability. But more than 99% of human jobs do not require particularly strong intelligence, and most of them can already be done by artificial intelligence derived from large models.
The human brain has 100 billion neurons, and each neuron averages about 1,000 connections, so the upper limit of human brain connections is about 100 trillion. GPT4 has about 1 trillion parameters, which is about one percent of the human brain. It is expected that within 10 years we will have large models with connectivity coefficients that exceed the upper limit of the human brain, and the potential capabilities of such large models are limitless.