MOSS is nowhere near the level of ChatGPT yet, but that is where the Chinese researchers want the Chinese chatbot to be one day. Releasing it as open source may help.
ChatGPT is currently the benchmark in the AI chatbot space, although it is far from perfect, and its integration with Microsoft’s Bing search engine has been met with both excitement and irritation.
The creators of the Chinese AI chatbot compare MOSS to its US counterpart. Like ChatGPT, MOSS is capable of answering questions, generating or summarizing code or text. In addition, MOSS would detect false claims and reject inappropriate requests.
High demand on Chinese ChatGPT causes service outage
Unlike ChatGPT, whose code is kept under wraps, Chinese scientists at Fudan University in Shanghai plan to develop MOSS in an open-source environment starting in late March.
“Large-scale language models are almost monopolized by foreign countries. They only develop APIs or do not open them to us,” Qiu said to China Daily. “We hope to share MOSS and model codes and development experience with everyone, and hope that China can be at the forefront of the world in terms of large-scale language models.”
In an effort to improve the language model through human feedback, researchers opened access to MOSS on February 20. As with ChatGPT, the technical infrastructure was unable to cope with the demand and went offline.
MOSS speaks better English than Chinese
Screenshots on the official website give an impression of the performance. Due to the training data used (“significantly fewer parameters than ChatGPT”), MOSS does not yet know any languages other than English. Next on the researchers’ to-do list would be to improve the chatbot in Chinese.
Although MOSS has acquired some capabilities of ChatGPT, we know that many limitations are remained and MOSS still lags far behind ChatGPT due to the lack of high-quality data, computing resources, and the model capacity.
Tianxiang Sun & Xipeng Qiu
Stable Diffusion has shown how open-sourcing AI models can drive development for image generation – Midjourney and DALL-E 2 cannot be trained with custom data and customized for individual use cases. However, the two commercial alternatives deliver better results out of the box.