News analysis： ChatGPT, change and risk
Xinhua News Agency, Beijing, February 13th News analysis: ChatGPT, change and risk
Xinhua News Agency reporter Peng Qian Huang Kun
ChatGPT, the big language model of American artificial intelligence company OpenAI, has reached 100 million monthly active users in January about two months after its launch, making it the fastest growing consumer application in history. Relevant experts predict that ChatGPT is not only a breakthrough of a new generation of chat robots, but also will bring great changes to the information industry. However, the risks brought by it, such as academic fraud, technology abuse and public opinion security, cannot be ignored.
The prototype of a new generation operating system platform
Write imaginative poems in multiple languages, write operational programs, quickly generate abstracts of papers, automatically make data tables, correct grammar and expression errors in articles, and write a news summary of the week … … ChatGPT can not only understand many human problems and instructions, but also smoothly carry out many rounds of dialogue, and also show its ability to solve a variety of general problems in more and more fields.
ChatGPT also easily passed some professional-level tests that are difficult for human beings: it recently passed the Google Code L3 (entry-level) engineer test; Passed the final exam of MBA in wharton school and the postgraduate exam of four courses in the University of Minnesota with B and C+ respectively; Passed the qualification examination of American medical practitioners … … The industry described its birth as an "iPhone moment" in the era of artificial intelligence, which means that artificial intelligence has ushered in a revolutionary turning point.
"The success of ChatGPT should not only be regarded as a breakthrough of a new generation of chat robots, but also pay attention to the revolution it has brought to artificial intelligence and even the entire information industry." Huang Tiejun, dean of Beijing Zhiyuan Artificial Intelligence Research Institute, said in an exclusive interview that the past decade in the field of artificial intelligence was a decade of deep learning, but there was no outbreak of mobile Internet and cloud computing level in the industry as a whole. "The emergence of ChatGPT is epoch-making, and the big model +ChatGPT has formed the prototype of a new generation of operating system platform."
Huang Tiejun said that ChatGPT adopted the route of "big data+big computing power+strong algorithm = big model" in the technical path, and explored a new paradigm in the direction of "basic big model+instruction fine-tuning", in which the basic big model is similar to the brain, and instruction fine-tuning is interactive training, and the combination of the two realizes language intelligence approaching human beings. ChatGPT applies the training mode of "reinforcement learning based on human feedback", and uses human preferences as a reward signal to train the model, which makes the model more and more consistent with human cognitive understanding mode.
"This kind of AI can help human beings to make real creation, especially to help human beings improve their creative efficiency, such as improving the efficiency of obtaining information or putting forward novel ideas, and then people can solve their authenticity problems. The improvement of creative efficiency will have great benefits and various influences, which can change the world information pattern. " Chen Xiaoping, director of the Robotics Laboratory of China University of Science and Technology, told reporters.
Initiate a new round of artificial intelligence technology competition
The advent of ChatGPT is triggering a new round of science and technology competition in the field of artificial intelligence. In the early morning of February 8th, Beijing time, Microsoft launched the latest version of bing search engine and Edge browser supported by ChatGPT, and announced that it would "reshape search". All products such as Microsoft’s Office and Azure cloud services will be integrated with ChatGPT across the board.
In the early morning of February 7th, Beijing time, Google also released Bard, the next generation dialogue AI system based on Google LaMDA model. On the same day, Baidu officially announced that the large-scale model project "ERNIE Bot" under development is scheduled to complete internal testing in March and then open to the public. China enterprises such as Alibaba and JD.COM also indicated that they are developing or planning to develop similar products.
The global competition in the field of artificial intelligence big model has become increasingly fierce. Huang Tiejun believes that ChatGPT is expected to evolve into a new generation of operating system platform and ecology in the future. This change is similar to the transformation of mobile Internet from personal computer to mobile phone, and most of the computing load will be taken over by the new generation information infrastructure with big model as the core. This new paradigm will affect all levels from application to infrastructure, causing great changes in the whole industrial structure. The ecological dispute between the big model and its software and hardware support system will become the focus of the information industry in the next decade.
It is worth noting that ChatGPT sometimes "talks nonsense in a serious way", and there are many problems such as factual errors, blind spots in knowledge and common sense deviations. It also faces the general risks of artificial intelligence such as compliance of training data sources, bias in data use, generation of false information and copyright disputes. Many world-renowned academic journals have updated their editorial guidelines, including that any large-scale language model tool will not be accepted as a signed author of research papers.
"The signed author of an academic paper must meet at least two conditions. One is to make ‘ Substantive contribution ’ The second is to be able to bear relevant responsibilities. At present, these two conditions ChatGPT (and other AI systems) are not satisfied. " Chen Xiaoping said.
ChatGPT also has the potential to be used in public opinion information warfare. The research team of McGill University in Canada used the previous generation model GPT-2 of ChatGPT to read about 5,000 articles about the COVID-19 epidemic broadcast by the Canadian Broadcasting Corporation, and then asked it to generate "counterfactual news" about the crisis. Even OpenAI warns users who use ChatGPT that it "may occasionally generate incorrect information" and "generate harmful instructions or biased content".
"In order to solve these problems, we need to control the application boundary of ChatGPT while developing technology, establish the management regulations for the content generated by artificial intelligence, and evade the use of artificial intelligence to generate and disseminate false and bad content. At the same time, strengthen the development of governance tools and identify the content generated by artificial intelligence through technical means. This is an important prerequisite for content detection and work confirmation. " Tang Jiayu, vice president of Beijing Ruilai Smart Technology Co., Ltd. said.