Baidu’s Ernie Bot arrived with a broad pitch: answer questions, write copy, and produce multimedia output, including Chinese speech and video. At its March 16 launch in Beijing, cofounder and CEO Robin Li presented the system as a response to demand for a ChatGPT alternative—and as technology with potential applications across many businesses.
A launch built around Chinese-language capabilities
Li demonstrated Ernie Bot answering math questions, discussing Chinese literature and writing marketing copy. He also emphasized the model’s performance on subjects rooted in Chinese culture, such as historical facts and traditional poetry. The name Ernie stands for “Enhanced Representation from kNowledge IntEgration;” its Chinese name is 文心一言, or Wenxin Yiyan.
The presentation’s most striking examples involved more than text. In a recorded interaction, Ernie Bot generated an illustration of a futuristic city transportation system, read a response aloud in a Chinese dialect, and edited and subtitled a video using the same text. Li said the model could generate audio in varieties including Sichuanese, Cantonese, Hokkien, and the Dongbei dialect.
Those capabilities made for a different comparison with ChatGPT and GPT-4. The source article notes that Ernie Bot offered more output options, while GPT-4 could analyze an image but generated text. Yet demonstrations did not establish how reliably Ernie Bot could perform its multimedia tasks: a Chinese publication later failed to reproduce video generation, and reviewers in a livestream could not test it successfully.
Access and performance remain open questions
At launch, Ernie Bot was available only to a very select group of Chinese creators. Businesses could apply for API access, but Baidu had not said whether consumers would be able to use it. The company also had not explained when the chatbot would be integrated into products such as its search engine or self-driving cars.
Early answers raised questions about quality. In a livestream, Chinese technology publication X.Pin asked Ernie Bot and GPT-4 some of the same questions in Chinese. Ernie Bot responded coherently to most of them, but made more mistakes. It struggled with Chinese history trivia, keeping track of the context of questions, and generating code for a mini game.
The launch itself offered limited evidence for judging the system. The event used five pre-recorded sessions rather than a live demonstration. Recorded material can be filtered and edited, and the video-generation feature could not be reproduced in later testing. Baidu’s release therefore showcased what the company said the model could do, while leaving its consistency difficult to assess.
Business adoption was central to Baidu’s pitch
Baidu presented Ernie Bot as a tool for companies, not just a chatbot for individuals. Li said, “Ernie Bot won’t just impact search engines and internet companies. It will impact every single company,” and said it would shorten the distance between companies and their customers.
According to Baidu, 650 companies had signed up before the launch, and more than 30,000 others had applied for API access afterward. The companies interested in using the chatbot reportedly included Lenovo, Trip.com, and several Chinese automotive companies. The source describes no details of those possible partnerships, so their shape and scope remained unclear.
The business focus also helps explain the event’s emphasis on bringing several capabilities together. Baidu already had AI tools for image creation and video editing. Observers noted that some launch demonstrations could be achieved with existing tools; the change was to make capabilities available through a more accessible interface. In that sense, Ernie Bot was presented as a way for business users to adopt Baidu’s existing AI work more easily.
Competition raised the pressure to release
Baidu had planned the mid-March release for months, but OpenAI’s GPT-4 launch on Tuesday became an immediate point of comparison. Li acknowledged the high expectations, saying, “People are expecting to benchmark Ernie Bot against ChatGPT, or even GPT-4. That’s a very high bar.” He also conceded that the model was imperfect, then explained the timing: “Because the market demands it.”
Other Chinese chatbots had not yet offered a clear alternative. MOSS, developed by researchers at Fudan University in Shanghai, attracted enough demand for its server to break down within a day of its late February launch; it had yet to return. The Chinese startup MiniMax released Inspo earlier that month, but it had been suspected of repackaging OpenAI’s GPT-3.5 model.
Baidu had previous AI work to draw on, including Ernie 3.0, released in 2019, and the text-to-image model Ernie-ViLG, released last year. Baidu said Ernie Bot was trained on “trillions of web pages, tens of billions of search and image data, hundreds of billions of daily voice data, and a knowledge graph of 550 billion facts.” The company also used its Kunlun AI computing chip to train and operate Ernie models.
Still, the launch left important comparisons unresolved. Baidu did not disclose the latest Ernie’s number of parameters, and the article says both companies guarded technical details. With limited access, recorded demonstrations and mixed early results, the next evidence would come from how the system performed as businesses began to use it.