【Editor's Note】
May 27, 2023 is the 118th anniversary of Fudan University. "The celebration of various activities is centered on the promotion of scientific research." Since 1954, on the eve of the school anniversary, then president Chen Wangdao put forward this idea, holding scientific reports during the school celebration has become one of the important academic traditions in Fudan.
Continuing the academic tradition, the string song is endless. From May 9th, more than 50 famous teachers from various disciplines of culture, social science, engineering and medicine will give academic speeches one after another.
The 20th session of the "2023 Xianghui University Celebration Series Academic Report", by Professor Qiu Xipeng of the School of Computer Science and Technology of Fudan University, Li Hao, researcher of the Institute of Artificial Intelligence Innovation and Industry of Fudan University, and Lin Feng, Vice President of DingTalk and Head of DingTalk Intelligent Hardware Ecology, jointly presented a report entitled "The era of general artificial intelligence is coming?" Driving Scientific Breakthroughs and Driving Social Change".
As the ultimate goal in the field of artificial intelligence - general artificial intelligence is gradually approaching reality, are we ready for change?
Qiu Xipeng Photo by Cun Fei
Updated MOSS model with hundreds of billions of parameters: called "Mars"
If you could define artificial intelligence, what would it be? In Qiu Xipeng's view, artificial intelligence is to make machines have human intelligence.
Starting from the McCulloch-Pitts model, Qiu Xipeng reviewed the development process of artificial intelligence in the report, with the advent of the era of deep learning, the rapid development of artificial neural networks, and the establishment of large-scale language models, so that the era of general artificial intelligence has come to us.
As a revolutionary data model, Qiu Xipeng believes that ChatGPT makes people see the promise of general artificial intelligence. Its breakthrough lies in aligning with real human needs, "especially in 'helpfulness', which can help people work more efficiently and answer questions about life." Under the relaxed standards, its ability to solve objective questions in the national college entrance examination paper is comparable to that of college entrance examination students with about 500 points, showing amazing understanding and reasoning ability.
In February this year, Fudan people's own large-scale model MOSS came out, as the first conversational large-scale language model officially released in China, Qiu Xipeng shared the team's journey from the pedestal research in 2020 to the final release. As a large-scale conversational language model, MOSS has the ability to emerge such as thought chain and scenario learning, and can execute human instructions and iterate according to feedback in dialogue with humans. During the model training process, the team paid special attention to the harmlessness and morality of MOSS, "For example, if you ask it how it destroys humans or criticizes children who are at the bottom of the grade, it will refuse to answer." ”
In Qiu Xipeng's view, the use of tools or plug-ins will be a very important core capability of the future large model, and in April this year, the team released the first large model with plug-in version capabilities in China. The first version of the MOSS model named "Moon" has about 16 billion parameters, and there are certain deficiencies in the ability to many factual and reasoning problems, so the updated version of the MOSS model named "Mars" and hundreds of billions of model parameters is about to be launched, "the effect will be very good", Qiu Xipeng said.
At present, various types of robots, Copilot, AutoGPT are gradually integrated into people's lives, based on large models, configuring perception multimodal capabilities, strengthening action behavior output, controlling the external world through programming, and then self-learning through interaction with the external environment, this general artificial intelligence prototype will be the next step of MOSS.
Large-scale language models will accelerate the realization of general artificial intelligence, and how to use general artificial intelligence to drive social change and promote the development of science and technology has become people's next thinking. So, what if the era of general artificial intelligence really arrives? Qiu Xipeng's answer is: higher medical care, better education, and greater scientific progress will bring a better life to everyone in society. But at the same time, we also need to do a good job of safety guidelines for AI models to ensure that they are beneficial to humans.
Live at the event
Digital productivity platform exploration in the era of general artificial intelligence
Lin Feng joined the Natural Language Processing Laboratory of Fudan University as early as 2004 to learn natural language processing from his supervisor Mr. Wu Lide. After graduation, he continued to innovate and explore the landing of AI in the industry.
Since its official release in 2015, DingTalk has 600 million users and 23 million organizations today, as a participant in application construction, Lin Feng introduced DingTalk's exploration in collaborative office and application development in the report, and shared the practical cases of enterprises on DingTalk in breaking information silos, realizing knowledge integration, and promoting efficient communication inside and outside the organization.
The era of general artificial intelligence is approaching, what does the industry think about it? In Lin Feng's view, enterprise digitalization is an opportunity for enterprise innovation in the intelligent era, and DingTalk will be fully upgraded to an intelligent collaborative office platform and an intelligent application open platform after accessing Alibaba's "Tongyi Qianwen" model.
"In the future, every industry will have a relationship with AI, and every company can have its own big model. Mastering AI tools to complete work efficiently will become a must-have for future professionals. ”
Lin Feng
Empowering Earth science with artificial intelligence
When AI meets remote sensing technology and meets the meteorological field, what kind of spark of wisdom will burst out?
Li Hao introduced that the Institute of Artificial Intelligence and Industry of Fudan University, based on AI for Science, uses artificial intelligence to empower the exploration of earth science.
Focusing on multi-source data aggregation, AI algorithm blessing, industry capability output, the use of AI analysis technology to achieve intelligent interpretation and analysis of remote sensing data, extract land surface coverage status information and dynamic change information, Li Hao introduced, the Institute uses AI+ remote sensing technology, has made comprehensive capacity output in many industries such as natural resource supervision, water conservancy river protection, ecological environment monitoring, agricultural yield estimation and emergency disaster prevention and mitigation.
Li Hao
Daily weather forecasting is mostly numerical weather forecasting, which solves the atmospheric motion equation system by numerical methods, so as to predict the atmospheric state at the future moment from the known initial moment of atmospheric state. Li Hao's team found that the use of large model technology in short-term and medium-term weather forecasting, using only artificial intelligence methods in the public test set can also achieve or exceed the results of numerical weather prediction. "In the future, how to reduce the cost of cloud computing and make more models calculated on the end is also a research direction for us."