On February 20, China’s C9 League member Fudan University released MOSS, the first large-scale conversational language model from China. It shares its name with the computer that controls the space station in the popular Chinese sci-fi film The Wandering Earth 2. This announcement comes amid the current popularity of ChatGPT, a chatbot developed by OpenAI.
MOSS can perform tasks such as generating dialogue, generating programming code, answering questions, and more. It was made available to the public by Professor Qiu Xipeng’s team via https://moss.fastnlp.top/. If a user does not have an invitation code to test out MOSS, they can join the “waiting list” after entering a mobile phone number. However, on the evening of February 20, the high volume of visitors to the page caused the MOSS server to overload, resulting in significantly slowed response times.
“Although there is lots of room for improvement, the advent of MOSS proves that Chinese research teams have the ability to overcome technical challenges in developing ChatGPT-like products,” said Qiu. The project has received strong support from Shanghai Artificial Intelligence Laboratory.
The basic steps of MOSS development are the same as ChatGPT and comprise of two stages: training in natural language processing and training in understanding human intention. In the first training stage for ChatGPT, OpenAI collected over hundreds of thousands of human queries and asked professional annotators from all walks of life to write replies. This data was then entered into its database to help ChatGPT gradually understand various human prompts. The Fudan University team adopted a different technical approach. By allowing MOSS to interact with humans and other AI models, its rate of learning and modeling was significantly enriched. Therefore, its natural language training was completed within a shorter timeframe than ChatGPT.
Qiu explained, “The difference between MOSS and ChatGPT is mainly in the pre-training stage of the natural language model. The number of parameters of MOSS is much fewer than ChatGPT. There is still much room for improvement in task completion and knowledge reserve.” In the beta testing stage, data and user feedback will be collected with the user’s permission to improve MOSS’s conversational capabilities. In addition to dialogue, MOSS will have more capabilities in the future, such as drawing, speaking, music composition, teaching, and even assisting scientists to conduct more efficient scientific research.
Currently, a major shortcoming of MOSS is its understanding of the Chinese language. This is mainly due to disarranged content on Chinese web pages used for training, such as intermittently appearing advertisements. Therefore, the Natural Language Processing Laboratory of Fudan University is focusing on screening of Chinese text content for the next stage of model training.
However, there are still many issues to be resolved before these kinds of AI chatbots are officially ready for prime time. Recently, Fudan University convened with executives and scholars from more than a dozen artificial intelligence companies to analyze the development, potential use cases, and security risks of ChatGPT.