At the 2023 HUAWEI CLOUD AI Dean Summit held on December 23, Zhang Bo, academician of the Chinese Academy of Sciences and honorary dean of the Institute of Artificial Intelligence of Tsinghua University, delivered a speech to share the value of large models, the governance problems they face, and the future development direction.
Zhang Bo, academician of the Chinese Academy of Sciences and honorary dean of the Institute of Artificial Intelligence of Tsinghua University, delivered a speech
Large models take a step towards general-purpose AI
Large language models can learn and understand human language to carry out conversations, and can also interact according to the context of the chat, truly chat and communicate like humans, and even complete tasks such as writing news, emails, scripts, copywriting, translation, etc. In this regard, Academician Zhang Bo said that it is very close to the dialogue with real people and achieves the goal pursued by behaviorism, which is a milestone achievement of AI.
Zhang believes that compared with the knowledge-driven model of the first-generation AI and the data-driven model of the second-generation AI, it breaks through the three specifics of specific fields, uses specific models, and realizes specific tasks, and takes a step towards general AI (AGI).
The governance of AI needs to be based on the model and usage perspective
The emergence of large models has transformed traditional artificial intelligence into generative artificial intelligence, opening a new door for all artificial intelligence companies and bringing new opportunities for the development of the AI industry.
First, promote the AI technology revolution. Zhang explained that the rapid development of information science and technology is due to the fact that it has laid a solid theoretical foundation in the early stage of development, but there is no theory of AI so far, and the existing models and algorithms of AI are aimed at specific fields and specific tasks, and it is difficult to establish general theories. However, unlike the first and second generation of AI, large models break through three specifics, enabling explainable and robust AI theories, which will greatly promote the rapid development of AI technology.
Second, we need to promote the transformation of the AI industry. Due to the limitations of a single field and a single task, the application and industrialization of AI are always limited to a narrow field. However, at present, the large model has expanded from generating text to generating other modalities, and it is possible to use a general model to generate a diversity of human-level images, sounds, ** and **, etc. Large models can meet the needs of different fields and tasks through fine-tuning, and combined with knowledge bases and other tools, they can greatly expand the development space of the AI industry and applications, and may bring the AI industry into a new stage of development.
In addition, the development of large models has brought some challenges. First, large models are only versatile in a few tasks such as natural language processing and programming, and whether they can be versatile in fields such as decision-making and games still needs to be studied.
Second, the large model uses the self-supervised learning method of the next word to learn and produce the text, which is completely different from the principle of human text learning and language production. This makes it have three essential defects: first, the output quality is inconsistent and uncontrollable, and there is a possibility of making big mistakes;Second, it is greatly affected by prompt words (input words), and the output robustness is relatively poorThird, there is no self-knowledge, and it is difficult for them to discover and correct mistakes by themselves.
Due to its shortcomings, it is common for large models to generate content that does not meet moral, ethical, and political standards, and needs to be solved by AI alignment, which is a governance problem for models. At the same time, the correct content of the output of the large model can also be misused and abused, so the user needs to be governed. In the process of AI development, it is necessary to manage the relationship between development and governance to ensure the healthy development of AI.
The development of the third generation of artificial intelligence requires the joint efforts of industry, academia and research
Although GPT can only be presented in the form of human-computer dialogue, which has various limitations, it is also a step forward in the direction of general artificial intelligence. How to make AI go on the road to generalization?"The third generation of artificial intelligence must be developed. ”
Zhang believes that the tasks of developing the third generation of artificial intelligence include, first, establishing explainable and robust AI theories. Second, we need to develop secure, trusted, controllable, reliable, and scalable AI technologies. Third, we need to promote the innovative application of AI and the development of the industry.
To achieve this task, Zhang said that it is necessary to make full use of knowledge, data, algorithms and computing power, and through the three key technologies of text semantic vector representation, multi-head attention mechanism converter, and self-supervised learning, it is a key step to realize the processing of text from just treating text as data in the past to processing the content of text and the knowledge contained in the text.
Zhang Bo pointed out that the development of the third generation of artificial intelligence also requires the joint efforts of industry, academia and research to combine basic research, technological innovation and industrialization.