Chinese tech giant Baidu cancelled its high-profile livestreamed launch event for its ChatGPT alternative, Ernie Bot. The media and public event had been scheduled for Monday afternoon but the company will now hold closed-door meetings with partners instead.
The chatbot is built on the massive ERNIE large language model created by the search engine giant and was unveiled last month by CEO Robin Li where he walked journalists through pre-recorded demos. This was due to be the first live demo of the capabilities in public.
Li’s previous demonstration of Ernie’s capabilities did not go down well with investors, with the company’s share price dropping 10% as a result of the limited range of functions displayed and the lack of opportunity to interact with the chatbot. The company saw the value of its stock fall again this morning.
Baidu says it has had more than 120,000 companies express an interest in utilising the power of natural language search through Ernie Bot and so rather than a large-scale public announcement, would instead focus on working with them in closed-door meetings.
There is a strong demand in China for ChatGPT-style chatbots and AI in general, with government officials also calling for more integration of the technology into the economy and society. When Ernie Bot was first demonstrated, shares in the company fell, but rebounded the next day due. Shares once again fell on the back of the cancelled livestream today.
Ernie bot is multi-modal, but unlike GPT-4 from OpenAI which is only able to take images and text input, it can also produce different media. According to Baidu it can generate text, images, audio and video from a given text prompt and produce voice output in several dialects. Its image generation capability is “totally self-developed”, according to Baidu as the Ernie-ViLG model used to power the chatbot has been trained to convert text to a range of formats.
Ernie Bot will have real-time information access
Ernie also has access to real-time information similar to Bard from Google and Microsoft’s Bing AI search, which uses the technology behind GPT-4. Reports suggest that despite having realtime access Ernie Bot has restrictions on it that prevent answers to certain political questions.
It isn’t the only chatbot in development in China. Alibaba is working on a chatbot built on top of its M6 foundation model which has been optimised to work with natural language tasks, text classification and sentiment analysis but requires less computational power than similar models. Tencent is building a chatbot on its own Hunyuan model, which was designed for Chinese-to-English and English-to-Chinese language translations. It was trained on a parallel corpus of English and Chinese data to improve accuracy and fluency.
Outside of the big tech companies, Chinese researchers are also actively pursuing large language models. Tsinghua University opened-sourced its GLM-130B project, which is a Chinese and English-language model trained with high accuracy and the Beijing Academy of Artificial Intelligence built the sparse language model WuDao which has a trillion parameters – significantly more than then few hundred billion in GPT-4.
While the Beijing Academy approach is different to other language models, utilising a mixture of expert architecture, it can understand and generate human-like text, translate languages and generate images so could be used for chatbot applications.
Speaking to Tech Monitor last month, venture capitalist and creator of Tech Buzz China, Rui Ma, said there was discussion in the country over finding the best opportunity and right opportunity for chatbot-like technology. It has the potential to transform search, copywriting and even automate white-collar roles at a large scale which could solve some of China’s ageing population challenges, she explained.