The whole world is discussing ChatGPT, but you don’t understand?
Don't panic! There are many people who don’t understand it because it involves too many and complicated concepts. When you mention ChatGPT, you cannot avoid large language models, deep learning, neural networks, artificial intelligence... For non-professionals, it is really not that easy to understand this swarm of abstract concepts.
Is it possible to use some popular explanations so that even children as young as 7 years old can understand the basic principles and application prospects of ChatGPT? I think this thing is worth trying. After all, the popularity of ChatGPT has lasted throughout 2023 (even every update of it will be called an "epic update" by the media). Everyone should probably be more aware of this possibility. Something that will revolutionize our lives in the future, no matter how hard it is to comprehend.
In order to make it easier to understand, I present this article in the form of a dialogue, assuming that the person I am talking to is a 7-year-old child. It should be noted that: 1. The content of this conversation is purely fictitious and not a real conversation that took place; 2. There are many simplifications and analogies in the conversation in order to make it easier to understand, so it may not be accurate enough and may contain errors. , welcome corrections from professionals; 3. Part of the conversation content refers to ChatGPT’s answers to similar questions.
Okay, let’s start this conversation.
01 The first fruit of the big language model
Xiaobao: What is ChatGPT?
Tree Dragon: A chatbot.
Xiaobao: I’m glad you didn’t use a lot of unfamiliar words to describe it. So, what’s so special about this chatbot?
Tree Dragon: It is countless times more powerful than previous chatbots. Specifically, it understands your questions and gives great answers in context—answers that are creative, rather than just repeating certain lines. Some studies believe that ChatGPT’s mental level is close to that of a 9-year-old human child.
Xiaobao: Ha, then it is smarter than me. But it’s just a chatbot after all. I don’t understand why people are crazy about it?
Shulong: What really drives people crazy is the technology behind ChatGPT - the large language model. You can think of it as the trunk of a magical tree, and ChatGPT is just the fruit grown on one of its branches. The trunk of the big language model is a language expert who can communicate smoothly with humans in natural languages (English, Chinese and other natural human languages), and can also be competent in many language-related tasks.
Xiaobao: Well, I want to know what other branches will grow on this tree?
Shulong: Many people are studying this issue, and the current possible directions are as follows: assistant GPT, arranging your digital life like a secretary; learning GPT, helping you improve yourself like a knowledgeable teacher; programming GPT, speaking According to your needs, it will immediately help you create the software you want; customer service GPT, which replaces 99% of manual customer service problems; creation GPT, helps you write documents, make PPT, and even write stories and poems...
Xiaobao: If I didn’t know how it actually works, I would think everything you said is just bragging.
Tree Dragon: Be skeptical, it is valuable, it will make your understanding deeper.
02 Fill in the blanks expert
Xiaobao: So, how does ChatGPT work? Why is it so smart?
Shulong: You can understand it this way - it is just doing "fill-in-the-blank questions" and predicting what the next word that appears should be based on the known text.
Xiaobao: This is very different from humans.
Shulong: Yes, humans must first understand what the problem is, and then organize language and give answers based on the key points of the problem. And ChatGPT is just constantly doing this kind of calculation - based on the known words (or sentences), which words should appear next, what are the probabilities, and then selects the words it thinks are appropriate and organizes them as answers.
Xiaobao: What do you mean, ChatGPT actually doesn’t understand what humans are saying, nor does it understand what it itself is saying? Instead, are you doing "probability calculation questions" one after another?
Tree Dragon: Indeed. From the perspective of its working principle, we can say that ChatGPT does not really understand human language, but if you do not consider its working principle, but evaluate based on its performance, it not only understands human language, but is also a language master.
Xiaobao: I am beginning to be confused about the question "Is ChatGPT smart or not?"
Shulong: In fact, more than 70 years ago, a British man named "Alan Turing" was already confused by this problem. He came up with a method to test whether machines have the same intelligence level as humans. This method is called the "Turing test". Simply put, the Turing test is to let a machine talk to humans (in the form of text). If humans cannot tell that it is a machine, then it can be said that the machine has intelligence.
Xiaobao: ChatGPT can definitely pass this test.
Shulong: Yes, in July this year, "Nature" stated in an article that ChatGPT had broken the Turing test.
03 Use mathematics to solve language problems
Xiaobao: Now I know, ChatGPT is a chatbot that is good at filling in the blanks. I am a human child who can do fill-in-the-blank questions. I can do it because I have learned it. What about it?
Shulong: Just like your teacher will train you, the developers of ChatGPT will also train it. ChatGPT mainly improves itself by studying Internet web pages and e-books. So far, humans have created billions of public web pages and produced 5 million e-books, which are all "textbooks" for ChatGPT learning.
Xiaobao: Human beings will never be able to read so many things.
Shulong: Yes, such a huge amount of reading can turn a fool into a genius. Through training, ChatGPT has mastered the relationship between words, the grammatical rules hidden behind sentences, and the logical relationships of some languages. Finally, a "large language model" was built...
Xiaobao: Wait, you need to first talk about what a "model" is?
Tree Dragon: A model can be understood as a simplified description of a complex system, with the purpose of simulating and predicting the behavior of the system. For example, a map APP is a model of the real world. With a map APP, you can easily calculate the distance between two locations. Similarly, language models can understand and generate text, and image models can recognize objects in images. The essence of a language model is to do a series of complex mathematical calculations, but to the user, it is like a magic box. If you ask a question, it will give you an answer.
Xiaobao: Give me another example.
Shulong: Have you seen the TV series "The Three-Body Problem"? In "The Three-Body Problem", von Neumann used 3,000 soldiers to form a humanoid computer. In fact, he wanted to create a calculation-based model of the sun's movement to predict the constant epoch and chaos. epoch cycle, but failed. Talking back to ChatGPT, the large language model behind it successfully turns complex human language problems into mathematical problems, which is amazing.
"Three Body" Human Column Computer
Xiaobao: Do scientists want to use mathematics to solve all problems?
Tree Dragon: Mathematics is everywhere. Many scientific fields, such as physics, chemistry, biology, etc., rely on mathematics to build models, perform analysis, and make predictions. The fields of technical engineering and humanities are also inseparable from mathematics. Some people believe that mathematics is the universal language of the universe because it provides a precise, clear, and abstract expression that helps us understand and describe the laws of nature. But in the past, people did not believe that human intelligence (such as the ability to understand and use natural language) can also be calculated. And ChatGPT does this.
Xiaobao: What does this mean?
Tree Dragon: Maybe, human language is not as complex as we think; maybe, the human brain is just a model based on mathematical calculations.
04 Mimic the human brain
Xiaobao: Is ChatGPT considered a real artificial intelligence?
Shulong: Of course. Artificial intelligence (AI) simply means making machines possess intelligence similar to humans. It has a broad meaning. Speech recognition and face recognition are both considered artificial intelligence. Before the emergence of ChatGPT, the applications of artificial intelligence were relatively simple, such as for playing chess, or in vertical fields such as medical care and finance. You can think of these artificial intelligences as technical secondary school students who graduated from vocational schools. They have certain skills, but their abilities are not universal enough. ChatGPT is like a college student who has completed an undergraduate degree. He has a good foundation of abilities and can adapt to many types of jobs in the future.
Xiaobao: Why couldn’t ChatGPT be built before?
Shulong: ChatGPT stands on the shoulders of giants, and of course some timing and luck are added to it. There is a saying called "AI three elements", which refers to the three main components of artificial intelligence: data, algorithm and computing power. Among them, data is used to train models, and the development of the Internet has created massive data; training models requires a large amount of computing resources, that is, computing power. In recent years, with the widespread use of GPUs and cloud computing, the scale of computing power has continued to expand; At the algorithm level, in 2017, a scientific research paper titled "Attention is all your need" introduced a model architecture called Transformer, which inspired the OpenAI company, which used this architecture to create ChatGPT.
Xiaobao: I still can’t understand why a mathematical model can be as smart as the human brain?
Tree Dragon: Probably because it mimics the human brain.
Xiaobao: How does the human brain work?
Tree Dragon: There are 100 billion neurons (nerve cells) in the human brain. They are connected to each other and transmit information through electrochemical signals, forming a huge neural network. For example, when you see a fruit, the neurons will analyze its color, shape, taste and other information, and then combine it with previous judgment patterns to determine whether it is an apple or a peach. If you want to improve your brain's ability to recognize more fruits through training, then you need to bring many different types of fruits in front of you and tell you their names. Your brain will slowly establish a pattern for recognizing fruits. Therefore, learning is very important.
Xiaobao: You just completed a training session for me.
Tree Dragon: Yes. The situation is similar for large language models. ChatGPT is also a huge artificial neural network inside . Take the GPT-3 version as an example, it has approximately hundreds of millions of neurons and 175 billion weights (which can be understood as the strength of the connections between neurons). When training ChatGPT, these neurons work together to identify a large amount of input text data, learn the language knowledge, grammatical rules and common sense in it, and then continuously adjust the model to establish the model's response pattern to different texts.
Xiaobao: If you have to study billions of web pages and 5 million e-books, you will be exhausted?
Shulong: So large language models are very computationally intensive. As there are more and more large language models, the demand for GPUs (graphics processing units) is increasing. Companies that produce GPUs have made double their profits in the past two years. Why not use the CPU (central processing unit) used in traditional computers? Because training large language models requires a large number of complex parallel (simultaneous) calculations, which is the strength of GPUs.
Xiaobao: In comparison, the human brain is so powerful. It has 100 times the number of ChatGPT neurons. It can keep working as long as it eats a little food every day.
Shulong: Yes, the energy consumption is extremely low. Maybe the human brain is a supercomputer. In the 70,000-year history of mankind, genes have received various survival trainings. When we are born, we receive new training the day after tomorrow. Together, these trainings establish patterns in how we respond to events. Some studies believe that each person may make tens of thousands of decisions every day. Perhaps the essence of how our brains work is also mathematical calculations.
05 Be careful with AI!
Xiaobao: Okay, now I know what ChatGPT is. It is a master of human language and has a neural network similar to the human brain. After training with human knowledge, it has established a set of language patterns. If you input some text into it, it will calculate and output the corresponding "answer" ".
Shulong: The summary is very accurate.
Xiaobao: So, why should a 7-year-old child learn such boring things?
Shulong: Just like those born in the 1990s are called "Internet natives", those of you born in the 10s are becoming "artificial intelligence natives" and should know more about your era.
Xiaobao: But I think so far, it can do nothing but chat.
Shulong: Don’t be too impatient, many GPTs with different functions are on the way. In the future, you will have a secretary GPT to help you arrange your schedule, book tickets, find restaurants, and take taxis; you will also have several study GPTs to give you lectures, help you with your homework, and answer your questions; when you enter society, you will also There is no need to do basic copywriting or communication work, but focus on advanced problems that AI cannot solve; all traffic lights are managed by AI, and even our train, plane, and car drivers will become AI...
Xiaobao: Why do you feel something is wrong?
Shulong: My description reminds you of some science fiction movies, where AI controls everything, including humans. The emergence of ChatGPT seems to have opened the first page of the story. Many people have become wary.
"Rocky 2" artificial intelligence Miss Minute wants to become a human
Xiaobao: So why should we let AI take over traffic lights, airplanes, and cars?
Shulong: Because AI takes over is more convenient, more efficient, and safer (theoretically), especially AI that can communicate with humans in natural language. Just like humans stopped wandering after learning to farm, humans will not go back to the past after they start relying on AI; farming reduces the scope of human activities and degrades human bodies, and AI will also bring new problems. .
Xiaobao: According to the news, many people are worried about losing their jobs.
Shulong: Yes, it is inevitable that AI will cause some people to lose their jobs, but the essence of unemployment is a problem of wealth distribution. Some people have gained greater benefits through AI, and some people have lost their jobs because of AI. Wealth has been transferred from the latter to the former, requiring social readjustment. But it is very necessary for everyone to learn and use AI early.
Xiaobao: AI can also be used by bad people.
Tree Dragon: Very true. AI is just a tool that can efficiently complete the tasks assigned to it by humans. If AI is mastered by bad guys, it may be used for weapons, illegal profit-making, etc.
Xiaobao: Then we must firmly control AI in good hands!
Tree Dragon: Not entirely effective either. If AI achieves super intelligence, it may develop its own consciousness, and AI's own development will may pose a threat to humanity.
Xiaobao: We can write some basic rules for all AIs that will not harm humans, and they will be shipped out of the factory.
Shulong: What you mean is like the "Three Laws of Robots" - 1. It must not harm humans or ignore human beings being harmed; 2. It must obey human orders unless the order conflicts with 1; 3. Robots must protect themselves and not Under the premise of violating 1 and 2. But in fact, simple rules are not enough to solve all problems, just like we have created so many laws and regulations, but still cannot solve all real-life conflicts. Each AI will have its own goals, and humans will also have their own goals. Conflicts are inevitable.
Xiaobao: So what should we do?
Shulong: This is a major issue for the whole society, and no one has a ready answer. You just need to remember that the "AI threat theory" is not unfounded, and we should give this issue greater decision-making weight.
Xiaobao: That’s it. I would like to say that you successfully trained me on ChatGPT.
Shulong: I'm very pleased that this may optimize your brain's algorithm. However, ChatGPT is just the beginning. Regarding AI, we still have a long way to go.
This article comes from the WeChat public account "Shulong Talk" (ID: shulongtalk) , author: Shulong, 36 Krypton released with authorization.
