On January 8, 2024, the first China Public Relations Forum of the New Year, hosted by the China Public Relations Association, was held in Beijing. With the theme of "Challenges and Opportunities Brought by New Artificial Intelligence Technologies", Professor Sun Maosong, Foreign Academician of the European Academy of Humanities and Natural Sciences, Chief Scientist of the National 973 Program, and Executive Vice President of the Institute of Artificial Intelligence of Tsinghua University, was invited as a guest speaker. Guo Weimin, President of China Public Relations Association, Zhang Xiaoying, Vice President, Xia Jixuan, etc., attended the meeting, and Wang Daping, Executive Vice President and Secretary-General, presided over the forum.
In his speech, Sun Maosong pointed out that ChatGPT has brought broad application prospects and development opportunities to related fields, but it is also accompanied by new challenges. He pointed out that general artificial intelligence has begun to appear on the stage of history and will continue to expand various vertical applications of large models. ChatGPT may bring about a reshuffle of the industrial ecology and change the form of human culture more obviously. At the same time, the ethics and governance of science and technology caused by ChatGPT also deserve the attention of the whole society.
How does GPT make computers speak?
To answer what opportunities and challenges ChatGPT brings, we must first answer the technical principles of large language models.
It's about letting the machine speak for itself. Sun Maosong pointed out that the large language model takes a list of words and tries to follow the words that follow. This technology not only allows machines to learn to speak, but also allows machines to recognize real-world knowledge and logic in the process of learning.
Large language models built on top of the next word** have the ability to accurately understand and generate, but at the same time, they also have the weakness of being prone to "hallucinations" (the content generated by the model is inconsistent with real-world facts or user input). Sun Maosong said that this problem is accompanied by the emergence of technology and needs to be viewed dialectically.
In addition, Sun Maosong believes that the large language model represented by ChatGPT has created a new innovative model, so that human beings no longer need to think for themselves, but choose from a large number of candidates generated by the model, so that they can break through the limitations and inertia of human beings themselves, and discover more possibilities and creativity, which is not only applicable to the field of literature, but also to the field of natural science, such as new materials, genetic science, etc.
Large language models orSome of the work will be replaced
While the popularity and widespread application of large language models is also accompanied by hidden concerns.
Sun Maosong pointed out that although large language models are positioned to assist humans at work, they will still replace some people in the workplace – especially low-end workers – and have a greater impact on the text and image industry. He said that artist training is a pyramid structure, with artists at the top of the industry, and the downstream of the industry is composed of art and design and other occupations to form the job market. However, after the emergence of large language models, the downstream of the industry will be impacted.
In order to train top artists, the pyramid still needs to be there, but those below (the pyramid) don't need it? Where are these people going? This is something to think about. Sun Maosong further said that due to the strong understanding and generation ability of large language models, the closer the industry is, the more vulnerable it is.
For example, he said, "For example, many departments have cut off 50% of the translation department, and the rest of the people are put in various departments, and each department is assigned a translator, and the person is responsible for changing it after the machine is finished." Although the machine can't replace the top translation experts, it can replace the daily staff, and now many translation books are turned with chatgpt, and the quality is much better than the original. ”
In addition, large language models are also driving the development of digital humans and the metaverse. Sun Maosong pointed out that in recent years, the field of digital humans will have great development, and the application in the field of elderly companionship is also promising. But it also has potential negative effects, such as young people may seek emotional value from large language models or digital humans, and ignore real-world socialization, reducing the need for intimacy, which can exacerbate the problem of infertility.
At the same time, the issue of science and technology ethics and governance of large language models also deserves attention. The training of large language models involves ideological issues, and Sun Maosong took the example of asking "how did Indians disappear in the history of ChatGPT". Faced with this question, the answer given by ChatGPT shows that Indians are indigenous people who inhabit North America, and their disappearance is mainly due to the migration of European colonizers, and the consequent diseases, forced relocations, and conflicts. ChatGPT has no problem answering that. But in this process, the big ** is the biggest touch on the human heart, but it was lightly described by chatgpt, probably because the West does not focus on the big ** when discussing this issue, and the ideology is still reflected in the large language model. ”
There is still a gap between the general model at home and abroad
Another frequently debated question is how to balance the development of large models and governance? In his speech, Sun Maosong also shared his views on this issue.
He believes that in development and governance, the two must be balanced, "to guide the innovation ecology, pay attention to risk prevention, these two things must be balanced, both hands must be hard, otherwise one will kill the other." The industry is still evolving rapidly, he said, and after the advent of multimodal models at the end of last year, "it's a big unknown what will happen this year." ”
In addition, Sun Maosong pointed out that there is still a gap in general model technology at home and abroad, and the top priority is to make Chinese's large language model capability reach the level of GPT-4. In the field of artificial intelligence, the biggest gap between China and foreign countries is reflected in the original innovation, "For example, ChatGPT came out, we can keep up, but we can't make it (ChatGPT)." ”
This is related to human resources, Sun Maosong believes that China is not short of outstanding talents in the field, but lacks a very small number of top talents, "the obvious feature of artificial intelligence is that the most advanced things are very few particularly creative, particularly high level of people to make it, this we have an obvious gap, so we can generally only follow, there is no way to lead, this is the biggest gap between domestic artificial intelligence and foreign countries, this gap is not only in the field of artificial intelligence, in almost all fields of science and technology." ”
But he also said that although the current Chinese large language model cannot reach the level of GPT-4, it does not affect its application in the vertical field, "domestic application scenarios are special and complex, the United States does well, to China is not effective, domestic large models still have advantages." He concluded.
Written by: Nandu reporter Hu Gengshuo.