Shulou(Shulou.com)11/24 Report--

Last year, OpenAI started a wave of generative artificial intelligence, with the AI industry driven by big models running wild. In the process of the rapid development of the industry, how to make its behavior consistent with the true intention of human value and ethical principles has become the issue of "value alignment" in academic, policy, industry and other fields.

On October 26, the first phase of "Zhi Chuang Shanghai Federation", sponsored by the East China Base of Tencent Research Institute and Tencent YouTu Lab, was officially held in Shanghai Nuobu Center. The theme of the activity is "value alignment of large models", to explore innovative solutions to value alignment and legal and ethical governance, to ensure that AI technology meets social expectations, and to explore the balance between AI and human values. A number of scholars and experts from the forefront of the AI field share their views and research results.

Zhong Junhao, secretary general of the Shanghai artificial Intelligence Industry Association, pointed out in his speech that in the process of industry development, the "value alignment" of the large model, as an effective means to promote the coordinated development of various fields, has important practical significance and far-reaching influence, and will also provide new opportunities for global cooperation and development.

Liu Pengfei, an associate professor at Shanghai Jiaotong University, pointed out that most artificial intelligence promotion organizations, including OpenAI, are actively using various programs to enhance the perfect landing of value alignment, such as monitoring fine tuning, intensive learning based on human feedback, fact-based alignment based on retrieval, and trying to use tools to align models. These are all "AI value alignment" implementations that are highly expected by the industry at the present stage.

Man-machine cooperation becomes a reality, AI value alignment is indispensable.

How to avoid the abuse and production of harmful content of the AI model with increasing data scale and computing power has become one of the core issues in the artificial intelligence industry. For this reason, the "AI value alignment", which can effectively enhance the security and practicability of the AI model, will become the focus of the industry.

Zhang Qinkun, secretary general of Tencent Research Institute, pointed out that value alignment will become an important competitiveness of AI products. Because the realization of large model value alignment requires the combination of alignment techniques (human feedback reinforcement learning, AI supervision, etc.), training data intervention, model interpretability, confrontation testing and other technologies and governance measures. at the same time, we also need to focus on how to form a consensus on value alignment through more social participation in the future. And how to make the ability to supervise, understand and design the AI model develop synchronously with the complexity of the model itself, in order to better supervise the large model effectively.

"what we expect is that the large language model can be consistent with human values and real intentions, to ensure that AI technology can serve people reliably and safely, and to ensure that the social economy can grow healthily." Said Liu Pengfei, an associate professor at Shanghai Jiaotong University.

Professor Liu Pengfei said that with the growing practicality of generative artificial intelligence, in addition to the most basic performance optimization, the importance of value alignment has also increased rapidly with the rapid iteration of the large AI model. The large model without alignment will not only generate a lot of false or even harmful information, but also cause huge hidden dangers in the field of high-level applications.

Professor Yan Hongxiu of Shanghai Jiaotong University said that the development and progress of technology has always followed the trend of eliminating risks and striving for certainty. As an effective way to deal with risks, "trust" has always been a necessary choice for human beings. Without trust, the stagnation of technological development is almost inevitable. "AI value alignment" is an indispensable part of making artificial intelligence credible, controllable and reliable, and human-computer cooperation to reality.

"tamp the underlying ethical logic of value alignment with ethical structural performance, and construct the boundary of value alignment with the rational cognition of trust." Professor Yan Hongxiu concluded.

Let the value of AI be aligned and interpreted by legislation in many countries.

A number of guests said at the forum that a clear technological path and how to integrate technical theory with practical needs as soon as possible is undoubtedly the focus of the development of the topic of "AI value alignment" at this stage.

Teng Yan, a researcher from the Governance Research Center of Shanghai artificial Intelligence Laboratory, pointed out that in order to accelerate the development of "AI value alignment", in addition to "adjusting based on human feedback", it is one of the hot research directions in many countries to let artificial intelligence partially supervise or coordinate human technology supervision and provide some high-quality feedback data to improve value alignment ability. Including OpenAI, has devoted a lot of attention to this area.

"in addition, in the evaluation link of 'confirming whether the large model is aligned', in addition to using manual means to verify it, it has also gradually become a trend in the industry to evaluate through a dedicated AI model." Teng Yan believes that in terms of accuracy, the evaluation model still has room for improvement, but from the efficiency of execution and iteration, the use of AI large model to accelerate artificial intelligence value alignment, the prospect is very considerable.

In addition to the technical and ethical aspects, at the legal level, which is related to the operation of social reality, "AI value alignment" also brings new challenges. Zhu Yue, an assistant professor from Tongji University Law School, shares the latest industry views:

From the domestic "generative artificial intelligence information security service regulations" to the EU's "artificial Intelligence Act", and then to the deep autonomy experiments of American head enterprises, countries and regions that actively promote the development of AI large model technology around the world have put the relevant laws and regulations on the agenda. Detailed issues such as transparency, explainability and accountability, as well as privacy protection at the personal level, are all challenges in the legislative process today. "

Zhu Yue pointed out that although the current "AI value alignment" is still in the early stage of the iteration of basic technology, and it is too early to make a conclusion at the legal level, from the enthusiasm of governments and multinational enterprises, it is the general trend of the industry to interpret "AI value alignment" through legislation and constantly improve it through practice.

Focusing on the methodology and regulatory implementation of AI alignment, Professor Xiao Yanghua of the computer Science Department of Fudan University in Shanghai, Associate Professor Zhang Quanshi of Shanghai Jiaotong University, licensed Associate Professor of the School of Law of the University of International Business and Economics, Fang Liang, Senior Research Manager of Anyuan AI, Wang Binghao, expert of the overall risk Management Department of Ant Group, Li Ke, senior researcher of Tencent YouTu Lab, and Cao Jianfeng, senior researcher of Tencent Research Institute. Through the round-table discussion, this paper deeply exchanged views and analyzed and affirmed the significance of "AI value alignment" to the development of artificial intelligence industry from different angles.

Pay attention to AI value alignment Tech for Good and go further

AI value alignment is one of the most challenging basic research projects in the field of artificial intelligence at the current stage, which requires a wide range of disciplines and social participation, and its research value is related to whether the stronger AI in the next stage can be safely controlled to achieve the long-term goal of AI improvement. The ethical and moral significance of AI value alignment is also consistent with the Tech for Good concept consistently adhered to by the organizer Tencent Research Institute.

According to Xu Siyan, secretary general of the East China Base of Tencent Research Institute, as a think tank set up by Tencent, Tencent Research Institute has set up the East China Base in Shanghai as an exchange platform for regional research, combining various local industry-university-research institutions to carry out research in cutting-edge science and technology, artificial intelligence, future cities, science and technology, industrial Internet, and other fields by means of dialogue salon, subject research and white papers.

As a series of city brand activities built by the East China Base of Tencent Research Institute, the Zhichuang Shanghai Union Forum will regularly invite industry experts to discuss the changes in business models that may be brought about by the evolution of cutting-edge science and technology in the process of continuous development of scientific and technological innovation, as well as the relevant policy and legal issues involved.

As the first phase of the "large model value alignment" research activity, it was supported by Shanghai artificial Intelligence Industry Association, Shanghai key Laboratory of data Science, Shanghai artificial Intelligence Laboratory Governance Research Center, and the Digital economy and legal Innovation Research Center of the University of International Business and Economics.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.