The popularity of artificial intelligence (AI) chatbot ChatGPT, launched by US firm OpenAI last November, has opened the floodgates to similar services in China, with many firms pledging to catch up or even outperform the American company in the world of Chinese language AI.
Despite China’s lack of access to the most advanced chips – the crucial requirement for training AI models – Chinese tech firms have enthusiastically jumped into generative AI development, but some experts have warned that the industry is anything from 18 months to three years behind ChatGPT.
Voice and language recognition developer iFlyTek is the latest to publicly boast about its potential to outperform ChatGPT. Company chairman Liu Qingfeng said at a product launch on Saturday that iFlytek’s own large language model (LLM) application, SparkDesk, would be able to surpass ChatGPT in the Chinese language world by October.
As ChatGPT is not available in China and the central government is applying its strict censorship rules to generative AI content, the closed market has provided hope for local tech firms to stake their claim domestically.
Chinese search engine giant Baidu was the first local tech firm to announce a ChatGPT challenger in mid-March, although it got off to a rocky start after failing to provide a live preview. On Tuesday, the Beijing-based company said more than 300 partners from different industries have signed up to use its Ernie Bot. In light of the Baidu debacle, Alibaba Group Holding launched its Tongyi Qianwen LLM in early April.
Alibaba owns the South China Morning Post.
Meanwhile, the debate over whether China can catch up continues. Some generated images from Ernie Bot, which suggested that the system was simply translating Chinese into English for prompts, fanned discussion that Baidu had borrowed the core of ChatGPT, an allegation the company denied. OpenAI’s generative pre-trained transformer (GPT) 3.5 model is the core technology behind the US start-up’s chatbot.
According to a leaked memo dated early April that cited an employee at Alibaba’s research unit Damo Academy, Alibaba’s Tongyi LLM is about 18 months behind OpenAI’s GPT 3.5 model. Alibaba has neither confirmed nor denied the authenticity of the memo, which has circulated online.
The memo also stated that the Damo Academy had assembled a 100-strong team to work on iterating and improving its LLM, which was initiated in January 2020 before being launched as Tongyi in September last year. Alibaba did not respond to a request for comment.
ChatGPT has the capability to generate human-like responses to questions, and reached 100 million users in the first two months after its launch.
Baidu founder and CEO Robin Li Yanhong told local media 36Kr in an interview late March that Ernie Bot was only about two months behind ChatGPT.
Wang Xiaochuang, founder and ex-CEO of China’s second-largest search engine Sogou, has publicly questioned Li’s assessment, saying Ernie Bot lags ChatGPT “definitely by more than two months”. Wang, who says his AI start-up Baichuan Zhineng will be launching the best Chinese LLM by year-end, has suggested that mainstream Chinese LLMs are about three years behind ChatGPT.
Chinese tech companies were caught off guard by ChatGPT. Domestic AI champions SenseTime, Cloudwalk, Megvii and Yitu Technology have in recent years focused on applications such as computer vision and facial recognition, not development of LLMs.
As OpenAI’s Chinese rivals attempt to catch up to its GPT 3.5 model, the San Francisco-based company in March launched an improved GPT 4 version, which was hailed as the most advanced LLM to date due to its improved accuracy and enhanced ability to understand text, images and video prompts.
Chinese AI experts have noted that some of the obstacles to China’s progress in LLMs include the complexity of the Chinese language, the government’s censorship of sensitive topics, and lack of computational power. South China Morning Post