政府新闻
复旦MOSS团队：打造具有中国特色中文大型语言模型 研究成果将会开源 2023-03-02
MOSS, a ChatGPT-like AI bot developed by Fudan University, has "achieved fundamental functions" as a generative language model, though it is currently outnumbered by ChatGPT in parameters and interactive data with users.
On Thursday, a week after its soft launch got a lot of attention and caused its server to crash, the team talked more about MOSS's flaws and plans for the future.
According to Qiu Xipeng, the head of the developers, MOSS has a sufficient number of parameters to be serve enterprises after some data fine-tuning.
The model is outnumbered 10:1 by ChatGPT, which has 175 billion parameters.
He compared MOSS to a "smart kid" who, while not yet capable of writing poetry or solving problems, has demonstrated promise as an artificial general intelligence (AGI). Qiu asked his 6-year-old daughter to talk to MOSS, and they ended up having a pleasant and lengthy conversation.
MOSS does more than just talk. It is capable of most tasks in natural language processing, including machine translation, information extraction, and error correction.
"It assists humans in a variety of ways, but in the form of conversations," Qiu added.
Qiu Xipeng (left), head of MOSS's developers and a professor at the School of Computer Science at Fudan University
MOSS has a lower knowledge reserve of facts than ChatGPT due to a lack of parameters and interactive data with humans. It is more prone to errors when answering fact-based questions as a result of this.
However, a "smart kid" learns. MOSS's capabilities will be enhanced when it gains access to external knowledge bases or learns to search for or use tools in specific domains, "just like humans look up dictionaries or use search engines."
Qiu believes that MOSS's ability to comprehend, learn, and think, rather than its factual knowledge reserve, merits more public attention.
"If you look at its logic, I believe MOSS's performance is actually quite good."
The developers said that they intend to release the results of their work as open source at the end of March.
On February 20, shortly after MOSS's beta test went live, the website (https://moss.fastnlp.top/) was flooded with enthusiastic users. MOSS servers currently have a maximum capacity of tens of thousands of users.
"We discovered a lot of problems during the beta test, and the quality of feedback varies," said Sun Tianxiang, a leading developer.
The current version includes both good and bad answers (especially Chinese ones).
"In the next version, we hope to raise the lower limit."
Source: Shanghai Daily
