Chinese internet company Alibaba launches Tongyi Qianwen, a ChatGPT-like chatbot for Chinese businesses and consumers. Shortly after, the Chinese government unveils strict rules for generative AI.
Alibaba's Tongyi Qianwen is to be used primarily for business applications, driving Alibaba's cloud applications. Initially, Chinese business customers will have access.
"We are at a technological watershed moment driven by generative AI and cloud computing, and businesses across all sectors have started to embrace intelligence transformation to stay ahead of the game," said Daniel Zhang, Chairman and CEO of Alibaba Group and CEO of Alibaba Cloud Intelligence.
From business meetings to bedtime stories
Thanks to Tongyi Qianwen, the business application "DingTalk" can automatically convert meetings into text, summarize meeting notes, write emails or proposals, or plan marketing campaigns. As with ChatGPT and the like, simple text commands called "prompts" are all that is needed.
Alibaba points to the ability to use DingTalk to take an idea written on a piece of paper and develop it into an application. OpenAI gave a similar demonstration during its presentation of GPT-4's code capabilities.
With the help of Tongyi Qianwen, the Siri-like application "Tmall Genie" will be able to have more lively conversations and offer more functions. For example, it can generate children's stories, write down healthy recipes, give travel tips, or automatically select the right music for sports.
Tongyi Qianwen is based on Tongyi, Alibaba's proprietary framework for pre-trained models, which combines various AI models, including models that can convert text into images and short videos. Like GPT-4, Tongyi Qianwen will be upgraded in the future to include multimodal capabilities such as text-to-picture and image understanding.
In addition to Alibaba, surveillance company Sensetime and internet company Baidu have also unveiled their first large Chinese AI language models.
Chinese chatbots must serve the Chinese system
Shortly after Tongyi Qianwen's reveal, China's cyberspace regulator unveiled its rules for large generative AI models. One such rule is that users must identify themselves so that their activities can be tracked.
In addition, providers must take responsibility for the training data and the political and social orientation of the models. Each model undergoes a security review by the government before it is released.
The government also prohibits the generation of content that
- undermines state power,
- calls for the overthrow of the socialist system,
- contributes to the division of the country,
- or undermines national unity.
While Microsoft and OpenAI conduct public beta tests and accept that their chatbots may be overly emotional or politically biased, Chinese tech companies must tread much more carefully, walking a fine line between censorship and innovation.
Baidu's chatbot Ernie, for example, avoids critical topics and questions about Chinese President Xi Jinping. In a first public presentation, Ernie lagged behind Western competitors such as ChatGPT, which may have something to do with the strict regulations.