Joys and Sorrows of ChatGPT

At present, ChatGPT has not passed the evaluation of Turing test. In fact, so far, no artificial intelligence model has really passed the Turing test. In November last year, OpenAI, an American AI research company, released a chat robot called ChatGPT, which swept the world within a few weeks after its launch, and even triggered a new global AI competition. Elon Musk, CEO of social media Twitter, called ChatGPT "awesome" after using it, and even asserted that "AI is becoming more powerful, and is not far from threatening human beings". Aaron Benanaff, author of the book "Automation and the Future of Work" and professor of Syracuse University in the United States, said that 200 years ago, people had wave after wave of fear of automation technology. When people saw new technologies, they were afraid that they would completely replace human labor. At the beginning of February this year, ChatGPT launched the "whirlwind" mode, causing a new round of crazy technological upgrading, industrial restructuring, capital frenzy and giant competition. At first, the general public, who was in a mood of ups and downs as the wave of technology heat swept the world, felt no fear after a period of "immersion" experience, but still could not help feeling worried.

ChatGPT has only been released for two months, with more than 100 million monthly active users. You know, "100 million" is not a small goal. To reach this number, it took Facebook 4.5 years, WhatsApp 3.5 years, Instagram 2.5 years, and Google nearly a year. Zhang Shuwu, a researcher of the Institute of Automation of the Chinese Academy of Sciences and director of the Research Center for Digital Content Technology and Services, said in an interview with the reporter of Science and Technology Daily that people can regard ChatGPT as the most complex breakthrough in language recognition technology after voice, image and video recognition technology, and also a new leap in the development of perceptual intelligence technology based on pattern recognition and deep learning. Xiong Deyi, the head and professor of Natural Language Processing Laboratory of Tianjin University, said that from the perspective of AI development, this is an important milestone, and its significance is higher than the "alpha dog" seven years ago. He told the reporter that the "alpha dog" in the Go industry has refreshed the general public‘s view of AI technology, but it cannot be widely used and the general public is also difficult to contact it directly. ChatGPT is completely different. Compared with "alpha dog", ChatGPT can directly apply or support many scene applications and products. If the "alpha dog" is the "spring and snow" of technology application, ChatGPT may have the opportunity to become a "Xialipa". Indeed, after users have made unremitting efforts to explore the use of ChatGPT, people have found that it is not only a chat tool, but also can write code, love letters, essays, online consultation, writing scenario scripts, poems, novels, and even draw up travel routes and respond to brain turns. It can be described as "responding to requests", and its functions are too powerful to imagine. ChatGPT is rapidly penetrating all industries, and the general public can become users of ChatGPT, which is an important reason for its popularity. In fact, there have been many applications related to chat with robots before, and we are most familiar with the commonly used intelligent voice assistant. How is ChatGPT different from it? "The smart speakers also use natural language processing technology and conversational AI technology, but the language intelligence is relatively elementary. In many cases, we can only have a simple dialogue with them in a question-and-answer way." Xiong Deyi said, "The underlying technology of ChatGPT is not only different from the underlying technology of general smart speakers, but also much more powerful than them. Not only that, it improves the form of human feedback data collection of InstinctGPT, a similar technology. In the dialogue with people, it can support multiple rounds of dialogue based on historical content." Xiong Deyi believes that ChatGPT is the first step to achieve conversational AI. From the Internet level, it can be used as an information portal and combined with search engines; It can also be combined with office software to enhance office efficiency; From the perspective of human-computer interaction, it embodies an interaction mode based on natural language, which can be combined with the metauniverse and digital human. In addition, it can also act as a robot brain. "From the perspective of human information development, it has made a great breakthrough in the information interface between people and machines," Xiong Deyi said.

ChatGPT uses GPT-3 technology, that is, the third generation of generative pre-trained transformer 3. This is an autoregressive language model that uses deep learning to generate human-like text. According to the BBC‘s Science Focus magazine, the technology of GPT-3 seems simple, for example, it accepts user requests, listens to questions or prompts and quickly answers, but in fact, the implementation of this technology is much more complex. The model uses text databases from the Internet for training, including up to 570GB of data obtained from books, online texts, Wikipedia, articles and other texts on the Internet. To be more precise, it has conducted intensive training on the training data composed of 500 billion words. GPT-3.5 is the fine-tuned and optimized version of GPT-3, which is more powerful than the latter. ChatGPT is supported by the Large Language Model (LLM) of GPT-3.5 architecture. ChatGPT can also be described as an artificial intelligence model in the field of natural language processing (NLP). This means that it is a program that can code, model and generate human language. As a language model, ChatGPT needs to go through three stages of "refining" after self-supervised pre-training on large-scale text data. The first is the supervision fine-tuning stage. At this stage, the human AI trainer acts as both a user and an AI assistant, and generates dialogue samples in the form of self dialogue. ChatGPT conducts supervised training on these dialogue samples. Next, ChatGPT entered the second stage. The main purpose of this stage is to train a reward model based on the language model and score the machine-generated replies. In order to train the reward model, it is necessary to randomly select machine-generated text and sample multiple different versions. Human trainers will sort these different versions preferentially, and the sorting results will be used for the training of the reward model. Based on the well-trained reward model, the third stage uses reinforcement learning technology to further fine-tune ChatGPT. The difference of this technology is that it can strengthen learning technology based on human feedback introduced above, so that the generated text can match human intentions and values. Whether they understand technology or not, netizens all over the world are competing to open "brain holes" to test how "divine" ChatGPT is. According to CNN, it was asked to rewrite the popular song "Baby Got Back" in the 1990s in the style of "Canterbury Tales"; Someone wrote a letter asking ChatGPT to delete the bad account in the credit report; Others asked about its home decoration design scheme inspired by fairy tales.

After the boom, ChatGPT concept shares also followed the trend. According to media reports, Hanwang Science and Technology gained 7 limit boards in 9 days. Natural language processing research is the company‘s main business, and the technology behind ChatGPT is NLP. Xiong Deyi told reporters that NLP is a branch of AI and was first born in machine translation. Its history is actually longer than the name of "AI". It has been studied for more than 70 years. In short, the goal of this technology is to enable computers or robots to achieve the language ability of listening, speaking, reading, writing, and translating like humans. "ChatGPT is not a technology that can be achieved overnight. It is the product of qualitative change formed by the superposition of various technologies, and is the crystallization of NLP." Xiong Deyi explained that its underlying technologies, including Transformer, self-supervised learning, fine tuning, human feedback reinforcement learning (RLHF), AI alignment, and so on, have extensive research and application in natural language processing and artificial intelligence. "But ChatGPT skillfully combines these technologies. It effectively avoids the misaligned behavior of large models, takes advantage of the intelligent emergence ability brought by the scale of large models, enhances the instruction learning ability and user intention capture ability of large models, and unlocks the generalization ability of large models. These capabilities add up to form the qualitative change effect of ChatGPT," Xiong Deyi said. The Miami Herald reported that a user on Twitter took a simulated SAT exam with the help of ChatGPT and got 1020 points (out of 1600). According to the data of College Simply, this is only slightly below the average level. But when we ask ChatGPT "Can you help me pass the exam", it will kindly remind us not to use it to cheat, He said: "It is important to remember that the purpose of the test is to evaluate your understanding and mastery of the material. Cheating, or just relying on external sources like me, will destroy the value of the test, and will not promote learning or growth." This is because ChatGPT has been added to the pre-designed "moral" principle, that is, human feedback reinforcement learning mentioned above. OpenAI said: "This enables the tool to answer follow-up questions, admit mistakes, challenge incorrect premises and reject inappropriate requests." Xiong Deyi believes that this is because RLHF effectively improves the alignment ability of large models, that is, how to keep the output of AI models consistent with human common sense, cognition, needs and values. After the "training" of RLHF, ChatGPT has become highly anthropomorphic, and can learn and generate responses that sound very natural. RLHF also enables ChatGPT to understand instructions and make appropriate responses, thus making the dialogue more authentic. ChatGPT has therefore quickly become the "ceiling" of all intelligent chat robots.

With the development of science and technology, AI technology has exceeded the limits that people think cannot be reached again and again, from the IBM "Dark Blue" computer defeated human beings in chess for the first time in 1997, to the IBM AI robot "Watson" winning the intelligence competition "Jeopardy", and then to the "Alpha Dog" defeating human beings in the Go game in 2016, By 2019, Pluribus, known as the "AI god of gambling", will crush human players in Texas poker. Now, ChatGPT is within reach, pulling and bumping into the traditional mode and framework of people‘s participation in production and life, and continues to "storm" on the track beyond human beings. "ChatGPT triggered ten career crises", "20 occupations in the future may be replaced by AI" and other topics were frequently searched. Will it really compete with humans for jobs or even replace humans? The Miami Herald reported that according to a survey conducted by SortList on 500 ChatGPT users in six different countries, "migrant workers" have been considering what it means for their industry. According to the study, about 23% of employees working in the software and technology fields are worried about losing their jobs because of it, and more than 40% of employers said they are interested in "hiring" ChatGPT, which has a variety of marketing documents. According to an article published by CBS Fortune Watch, other professional fields may also change their work processes, including intermediate writing, advertising, media planning and legal documents. "ChatGPT and other AI technologies may change the nature of work, but they will not replace the need for human beings." When the question of "whether it will replace human work" was left to ChatGPT, it said that they were designed to assist and enhance human tasks rather than replace human beings. They serve as a complement to human resources, and improve human ability to solve problems, improve efficiency and effectiveness. Xiong Deyi believes that at present, AI plays more role in improving human efficiency and assisting human beings. At the same time, the public needs to keep pace with the times, be familiar with and master new technologies, just like learning to use computers and mobile phones. After all, these new technologies will certainly generate new employment opportunities.

The British philosopher Derek Parfait defined human beings as "creatures that can give reasons and respond" in his book "On Important Things". ChatGPT is an artifact that can give reasons and respond. How is it different from people? Are you self-conscious like people? Up to now, computers have become more mature, but they lack real understanding, which is cultivated by humans through evolution into autonomous individuals embedded in social practice networks. This is an experience that AI does not have. When asked about ChatGPT, It replied: "I am a machine learning model, and I don‘t have consciousness or self-consciousness like human beings. I can‘t feel emotions or experience the world like human beings. However, I can quickly process a large amount of data, and understand and respond to natural language input in a way similar to human communication." Zhang Jiaxing, a chair scientist of the Guangdong, Hong Kong and Macao Greater Bay Area Digital Economy Research Institute (IDEA), said in an interview with reporters, For practitioners in the field of artificial intelligence, there is no clear definition of self-consciousness. It is difficult to measure quantitatively such abilities as face recognition and speech recognition. "If we give a vague definition of ‘autonomous consciousness‘, I think first, the subject should have self-consciousness - it can be aware of what it is doing; second, it should emphasize autonomy - it can decide what to do and not to do spontaneously. However, at present, AI does not know what it is saying, nor can it decide or choose what it wants to say independently. From this perspective, there is still some distance from autonomous consciousness " Zhang Jiaxing said. Zhang Shuwu also believes that although ChatGPT has made a breakthrough, its answer is based on empirical judgment, is to make predictions according to requirements, and is also a "knowledge puzzle" refined on the basis of prior knowledge. It does not have real cognitive functions such as logical derivation and mathematical calculus, and its limitations are obvious. In 1950, brilliant computer scientist Alan Turing put forward a thought experiment he called "imitation game". The interviewer talked with two objects through a typewriter and knew that one was a person and the other was a machine. Turing suggested that if a machine can always make the interviewer believe that it is human, we can say that it has the ability to think. This is the famous "Turing Test". At present, ChatGPT has not passed the evaluation of Turing test. In fact, so far, no artificial intelligence model has really passed the Turing test. But according to the New York Times, OpenAI has built a system GPT-4 that is more powerful than ChatGPT. It can even generate images and text. It is rumored in the industry that it can pass the test. Whether it can be so, we will wait and see.