Guangzhou, March 3 (ZXS) -- Writing poems, paintings, and conducting medical diagnoses... Recently, the conversational large-scale language model technology represented by ChatGPT and MOSS has attracted widespread attention. Qiu Xipeng, professor of Fudan University and professor of the AI Model Algorithm Research Center of Pazhou Laboratory (Huangpu), said at the award ceremony of the first Guangdong-Hong Kong-Macao Greater Bay Area (Huangpu) International Algorithm Research Competition held in Guangzhou on the 25th that the emergence of large-scale language models will accelerate the realization of general artificial intelligence.
Not long ago, Qiu Xipeng's team released China's first large-scale language model MOSS. Qiu Xipeng said that compared with ChatGPT, the scale of MOSS parameters at this stage is relatively small, and the cost of research and development is relatively low, which can be used in vertical industries with a small amount of knowledge and only about 10 billion parameters. However, due to its small capacity and limited knowledge capacity, it still needs to be improved in terms of answering questions and logical reasoning ability, and then the knowledge in this field can be expanded by expanding the scale of parameters or accessing the knowledge base of the vertical field to expand the knowledge and enhance the reasoning ability.
"At the current level of ChatGPT, it is not technically difficult for us to catch up, but because it is also dynamically developing itself, it requires us to make more efforts to catch up." Qiu Xipeng said that the advantages of China's high-level talents and massive application scenarios can be leveraged to improve the underlying algorithm capabilities and catch up with application-driven innovation, and promote the development of Chinese large-scale language models.
In Qiu Xipeng's view, the establishment of large-scale language model technology requires multiple exploration and research, and after exploring its technical route, a large number of engineers are needed to implement it. "We need to improve the innovation ecology or mechanism and cultivate an atmosphere that encourages scientific research and innovation." He said.
Overseas leading research institutions have recently begun to turn to general artificial intelligence based on large models. Qiu Xipeng explained that the ultimate goal of achieving the development of artificial intelligence is to achieve general artificial intelligence, which is the same as the artificial intelligence image in science fiction movies, which has the same ability as people, especially the general learning ability to complete various tasks. "I am optimistic about this, because according to the capabilities currently demonstrated, we are not far from the arrival of the era of general artificial intelligence, and the next step is to consider connecting large models to the real world, and 'align' large models with the real world through embodied learning, cross-modal learning, etc."
"China's AI market is relatively active and rich in scenarios, and it is necessary to establish Chinese large-scale language models." Zhu Jun, a professor at Tsinghua University and a professor at the AI Model Algorithm Research Center of Pazhou Laboratory (Huangpu), said that in the field of knowledge annotation, China has a high-end knowledge talent dividend, and in the process of dynamic catch-up, it is necessary to form a high-quality team of engineering optimization and scientific research exploration to carry out collaborative research.
Zhu Jun said that Chinese mainland has established many artificial intelligence computing centers. "In the long run, we need to master the underlying algorithmic capabilities. But for now, new model algorithm innovation can be done first. He said.
It should be pointed out that the road of artificial intelligence needs to go far, and it is necessary to establish red line constraints. "Because AI models iterate very quickly, ethical issues are important." Qiu Xipeng suggested that artificial intelligence products should be restricted by laws and regulations and some algorithms should be eliminated through the model side to remove harmful information. (End)