How far is today’s AI from understanding the world like a human being？

Press:ChatGPT a few days agoAnnounced the launch of plug-in function,GiveChatGPTUsing tools, networking,Ability to interact with the outside world,Some developers said that this is the most important iterative update after the launch of ChatGPT, which basically provides an app store-style service to make itEnough to become a new generation of all-around digital assistants.

ChatGPT has been launched for more than a month, and every few weeks there will be functional expansion that shocks the world. Some people are excited, some people are afraid,《AI 3.0At the beginning of the book, Hou Shida’s fear is put forward: not that artificial intelligence is too smart, but that it is too easy for artificial intelligence to replace what we cherish. The author of this book, melanie Michel, a complex system scientist, focuses on the question in this book: What is the difference between human and machine understanding?How far is today’s machine from truly understanding the world like a human being?

She believes that there is still "between machines and people."Obstacles to meaning"(barrier of meaning): Humans can understand the situation they face in a profound and essential way. Although the most advanced artificial intelligence system is being completed.Some tasksIt has surpassed human beings, but these systems lack the ability to understand the rich meanings endowed by human beings in perception, language and reasoning.“Obstacles to meaning"Still exists today.

Author | Translated by melanie Michel | Wang Feiyue

Understanding any situation is essentially an ability to predict what may happen next. When you predict that people crossing the street will continue to walk in their original direction; The woman will continue to push the stroller, lead the dog and hold the mobile phone at the same time. You can also predict that the lady will pull the dog leash, and the dog will resist and want to continue to explore the smell of that place. The lady will pull the dog leash harder, and then the dog will follow her and walk into the road. If you are driving, you need to be prepared for it! On a more basic level, you must want a lady’s shoes to stay on her feet, her head to stay on her body, and the road to stay on the ground. You predict that the man will come out from behind the stroller, and he will have legs, feet and shoes, which will support him to stand on the road.

In short, you have what psychologists call important aspects of the world.“Mental model"Based on your knowledge of physical and biological facts, causality and human behavior, this model shows how the world works and enables you to simulate the corresponding situation psychologically. Neuroscientists still don’t know how this mental model or mental simulation running on it comes from the activities of billions of interconnected neurons. Some famous psychologists put forward that it is through these mental simulations that a person’s understanding of concepts and situations activates his previous personal experience and imagines the actions that may need to be taken.

A mental model not only allows you to predict what might happen in a specific situation, but also allows you to imagine what will happen if a specific event happens. For example, if you honk the car horn or shout out the window.“Get out of the way!"This lady may be startled and turn her attention to you; If she stumbles and her shoes fall off, she will bend down and put them on; If the baby in the stroller starts crying, she will take a look at what is going on. The key to understanding a situation is to be able to use mental models to imagine different possible futures.

Psychologist LawrenceBaslau (Lawrence Barsalou) yes“Understanding is simulation."(understanding as simulation) One of the most famous supporters of the hypothesis. In his view, our understanding of the situation we encounter is contained in the mental simulation we perform subconsciously. In addition, Baslau suggested that this mental simulation also constitutes our understanding of situations in which we are not directly involved, such as what we see, hear or read. Baslau wrote:“When people understand a text, they construct a simulation to represent its perception, movement and emotion. Simulation seems to be the core of meaning expression. "

I can easily imagine such a scene.——A woman had a car accident while crossing the street on the phone, and I understood it through my mental simulation of this situation. I may substitute myself into the role of this lady, and imagine through the simulation made by my mental model what I feel like holding a mobile phone, pushing a stroller, holding a dog leash, crossing the road, being disturbed, etc.

For images“truth”“exist”“infinite"How do we understand such very abstract concepts? Baslow and his colleagues have argued for decades that even the most abstract concepts are understood by mental simulation of the specific scenes in which these concepts occur.

According to Baslau, we use pairs of feelings.–Exercise (sensory-motor) state repetition (that is, simulation) to deal with concepts, and to represent their categories, even for the most abstract concepts. Surprisingly (at least for me), the most convincing evidence of this hypothesis comes from the cognitive study of metaphor.

A long time ago, in an English class, I learned "metaphor"The definition of, its general content is as follows:

Metaphor is a rhetorical device that describes an object or action in a way that is not completely true, but helps to explain an idea or make a comparison.……Metaphor is often used in literary genres such as poetry, and when people want to add some literary grace to their language..

My English teacher gave us some examples of metaphor, including Shakespeare’s most famous poem:

“What’s that light in the window over there? That’s the east, and Juliet is the sun."

“Life is just a walking shadow, a clumsy actor who gives directions on the stage. After a short appearance, he quietly retires in silence."

My understanding at that time was that metaphor was only used to add some literary grace to the original unremarkable works.

Many years later, I read a book by linguist George.·Leikauf (George Lakoff) and philosopher Mark·Johnson (Mark Johnson) co-authored Metaphor We Live by (Metaphors We Live By), after that, my understanding of metaphor has completely changed. Leikauf and Johnson’s point of view is that not only our daily language is full of metaphors that we are not aware of, but our understanding of almost all abstract concepts is realized through metaphors based on core physical knowledge.

Leikauf and Johnson cited a large number of language examples to prove their arguments, showing how we conceptualize abstract concepts such as time, love, sadness, anger and poverty with concrete physical concepts.

For example, Leikauf and Johnson pointed out that we will use concrete concepts, such as money, to talk about abstract concepts, such as time. For example, we often say, "You"cost"or“save"Time; You often don’t have enough time to come“cost"; Sometimes you“cost"The time is“worthy"And you have reasonably“use"Lost time; You may know someone who is“Borrowed time"The living people in the park.

These metaphors put forward by Leikauf and Johnson reveal the physical basis of our understanding of concepts, and support Baslau’s theory that people understand by constructing the simulation of mental models derived from our core knowledge.

Psychologists have explored the above ideas through many interesting experiments. A group of researchers pointed out that whether a person feels physical warmth or social warmth.“warmth"It seems that the activation is the same area of the brain. In order to study this possible psychological effect, the researchers conducted the next experiment on a group of volunteers, and each participant was accompanied by an experimenter through a short elevator trip to the psychology laboratory. In the elevator, the experimenter asked the subjects to take a cup of hot coffee or iced coffee for a few seconds, so that the experimenter could record the names of the subjects, but the subjects did not know that this was actually part of the experiment. After entering the laboratory, each subject needs to read a short description about the same fictional character, and then be asked to evaluate some personality characteristics of the character. The results show that the evaluation of the subject with overheated coffee in the elevator is obviously warmer than that of the subject with iced coffee.

Although these experiments and their explanations are still controversial in the field of psychology, their results can be understood as supporting the views of Baslau, Leikauf and Johnson: we understand abstract concepts through core physical knowledge. If in the physical sense,“warmth"Concepts are activated psychologically, for example, by holding a cup of hot coffee, which will also activate the more abstract and metaphorical level.“warmth"Concepts are like experiments to evaluate a person’s character, and vice versa.

It is difficult to talk about understanding without consciousness. When I started writing this book, I was going to completely avoid the question of consciousness, because it was so controversial from a scientific point of view, but somehow, I was still interested in some speculations about consciousness. If our understanding of concepts and situations is achieved by constructing mental models to simulate, then perhaps consciousness and all our concepts of self come from our ability to construct and simulate our own mental models.

I can not only mentally simulate the situation of crossing the street by phone, but also mentally simulate my own thoughts.And predict what you might think next, that is, we have a model about our mental model. Build a model for the model and simulate our simulation.——Why not? Just like the physical perception of warmth can activate the metaphorical perception of warmth, and vice versa, the concepts related to physical perception we have may activate the abstract concept of self, which produces a physical perception of self through the feedback of the nervous system. You can also put the“self"It’s called consciousness This circular causality is similar to what Hou Shida called consciousness.“vicious circle"：“Symbols interact with the physical level and reverse the causal relationship. Symbols seem to have free will and acquire the contradictory ability to promote particle motion."

So far, I have described the core intuitive knowledge that human beings are born with or acquired in the early life from the psychological point of view, and how this knowledge has become the basis for constructing our mental models of various concepts. The construction and use of these mental models depend on two basic human instincts: abstraction and analogy.

Abstraction is the ability to identify specific concepts and situations into more general categories. Let’s describe the concept of abstraction more concretely. Suppose you are a parent and a cognitive psychologist. For convenience, let’s call your child.“S”. Before you observeSIn the process of growing up, you keep a diary to record her growing and complex abstract ability. Now, let me imagine what you might write down over the years.

threeMonth:SI can distinguish my facial expressions expressing happiness and sadness and generalize them to other different people who communicate with them. She has abstracted“A happy face"and“A sad face"The concept of.

sixMonth:SNow she can recognize the meaning when people wave goodbye to her, and she can wave back. She abstracted“wave one’s hand"Visual concept, and learned how to respond with the same gesture.

18Month:SHas been abstracted“cat"and“dog"And many other concepts, so she can identify different kinds of cats and dogs in pictures, paintings and cartoons, as well as in real life.

3 years old:SYou can identify individual letters in the alphabet from different people’s handwriting and printed fonts. In addition, she can distinguish between uppercase and lowercase letters. In short, her abstraction of concepts related to letters is quite advanced! In addition, she also summarized her knowledge of carrots, broccoli and spinach into more abstract concepts.——Vegetables, and now she equates vegetables with another abstract concept.——It tastes terrible.

eightYears old: I overheardSBest friendJtellS, once.JOur mother forgot to pick her up after her football match.SResponded that:“Well, exactly the same thing happened to me. I guess you must be angry and your mother feels very guilty."However, this“Exactly the same thing"It’s actually a quite different situation:SOur nanny forgot to pick her up from school and take her to piano class. whenSsay“Exactly the same thing happened to me."At that time, it was obvious that she had constructed an abstract concept, similar to a situation where a caregiver forgot to pick up the children before or after an activity. She can also map her experience toJandJMother, to predict their reaction.

The purpose of my mentioning this imaginary diary of parents is to explain some important points about abstraction and analogy. In some form, abstraction is the basis of all our concepts, even from the earliest infants, even from the earliest infancy. Such a simple thing as recognizing a mother’s face under different lighting conditions, angles, facial expressions and different hairstyles is the same abstract feat as recognizing a musical style or making a convincing legal analogy. As the above diary shows, our so-called perception, classification, recognition, generalization and association all involve our abstract behavior of the situations we have experienced.

Abstract and“Make an analogy"(analogy making) is closely related. Hou Shida has been studying abstraction and analogy for decades. In a very general sense, analogy is defined as: the perception of the common essence between two things. This common essence can be a named concept, such as“smiling face”“Wave goodbye”“cat”“Baroque music"We call it a category; Or concepts created in a short time that are difficult to express in words, such as a caregiver forgetting to pick up the children before or after the activity,“Or an owner who is not responsible for the content created by users in the public writing space, we call it analogy. These psychological phenomena are two sides of the same coin. In some cases, such as“Two sides of the same coin"Our idea is to start with an analogy, but eventually it is integrated into our vocabulary in the form of idioms, which makes us treat it more like a category.

In short, analogy is often made unconsciously by us, and this ability is the basis of our abstract ability and concept formation. Just as Hou Shida and his co-author and psychologist Emmanuel·Sandel (Emmanuel Sander) elaborated in the "appearance and essence":“No concept, no thought, no analogy, no concept. "

I certainly don’t mean that abstraction and analogy cover all the components of human understanding. In fact, many people have noticed "understand"and“meaning"Terms such as equality are just ill-defined terms that we use as placeholders, let alone consciousness, because at present we have no accurate language or theory to discuss what really happens in the brain.“Marvin, the pioneer of artificial intelligence·Minsky said this:“Although modern science has some budding ideas, it makes‘believe’(believe),‘know’(know),‘mean’Such words have become very common in daily life, but strictly speaking, their definitions seem to be too rough to support a strong theory.……Just like at present.‘self’(self) or‘understand’Words like (understanding) are the same to us, and they are still in the initial stage to a more perfect concept."Minsky went on to point out:“Our confusion about these concepts stems from the fact that traditional ideas are not enough to solve this extremely difficult problem.……We are still in the formative period of a series of concepts about mind."

In order to complete its work reliably and stably, to what extent does the artificial intelligence system need to have the understanding ability like human beings? Or to what extent? No one knows the answer, but researchers in the field of artificial intelligence agree that mastering core common sense and complex abstraction and analogy ability is an indispensable part of the future development of artificial intelligence.

This article is excerpted from AI 3.0.》Chapter XI, with abridged contents, was published with the authorization of the publishing house.