Chinese tech giant Baidu releases its answer to ChatGPT

On Thursday, Robin Li, Baidu’s cofounder and CEO, took the stage in Beijing to showcase the company’s new large language model, Ernie Bot. Accompanied by art created by Baidu’s image-making AI, he showed examples of what the chatbot can do, including solve math questions, write marketing copy, answer questions about Chinese literature, and generate multimedia responses. 

Baidu had planned for this mid-March product release for months. But it was intercepted by the unexpected release on Tuesday of OpenAI’s GPT-4, which clearly became a reference point for everyone watching Baidu’s activities, including the CEO himself. “People are expecting to benchmark Ernie Bot against ChatGPT, or even GPT-4. That’s a very high bar,” Li said at the beginning of his presentation.

As expected, Ernie Bot (the name stands for “Enhanced Representation from kNowledge IntEgration;” its Chinese name is 文心一言, or Wenxin Yiyan) performs particularly well on tasks specific to Chinese culture, like explaining a historical fact or writing a traditional poem. (Li says as a Chinese company, Baidu “has to perform better than any pre-trained LLMs” in terms of understanding Chinese.) 

But the highlight of the product release was Ernie Bot’s multimodal output feature, which ChatGPT and GPT-4 do not offer (OpenAI has bragged about GPT-4’s ability to analyze a photo of the contents of a refrigerator and come up with recipe suggestions, but the model generates only text). Li showed a recorded interaction with the bot where it generated an illustration of a futuristic city transportation system, used Chinese dialect to read out a text answer, and edited and subtitled a video based on the same text. However, in later testing after the launch, a Chinese publication failed to reproduce the video generation. 

The Chinese public has been hungry for a ChatGPT alternative; both OpenAI and the Chinese government have barred individuals in China from using the American chatbot.

But so far, Ernie Bot has been made available only to an extremely select pool of Chinese creators. Companies can apply for API access. But Baidu has not said whether the technology will be available for consumers. It’s also unclear when the bot will be integrated into Baidu’s other products, like its search engine or self-driving cars, as the company promised.

Compared with the rollouts of ChatGPT and GPT-4, Ernie Bot’s release felt rushed. The presentation did not feature any live demo but instead used five pre-recorded sessions. Li also repeatedly said that Ernie is still imperfect and will improve once it reaches more users. Baidu’s stock price slipped by 6.4% on Thursday, and social media is full of disappointed reactions.

Li seemed prepared for such a response. “People have been asking me for a while: Why are you releasing [Ernie Bot] so soon? Are you ready for it?” he said during his presentation. “From what I personally saw when conducting internal tests on Ernie Bot, it’s not perfect. But why do we want to release it today? Because the market demands it.” 

The race to be the first

While a few ChatGPT-style bots have already been released by Chinese companies or researchers, none of them has shown satisfying results. MOSS, an English-language chatbot developed by Fudan University researchers in Shanghai, was met with such high demand that its server broke down within a day of launch in late February. It has yet to return. MiniMax, a Chinese startup, released a chatbot called Inspo earlier this month, but it has been suspected of merely repackaging the GPT-3.5 model developed by OpenAI.

Many people expected that Baidu would be the first Chinese company to go head to head with ChatGPT. Back in 2019, Baidu released a GPT-3 equivalent—Ernie 3.0. It also released a decently powerful text-to-image model called Ernie-ViLG last year. 

The company has a few advantages that enable it to stand out among its Chinese peers. It has designed its own AI computing chip, Kunlun, that was used in training and operating the Ernie models and could shield the company from the ever-growing US-China tension around semiconductors. Also, having made a search engine, an online encyclopedia, a discussion forum, and a media publishing platform since 2000, Baidu can access Chinese language training material from a variety of proprietary resources. According to Baidu’s press release, Ernie Bot is 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.”

At the launch, Li compared Baidu to big tech firms in the West. “I can say Baidu is the first one among international tech giants to release [a ChatGPT alternative developed internally]. Microsoft just uses OpenAI access. Google, Meta, Amazon—none of these has released a product of the same kind and at the same level,” he said.

The inevitable comparison to GPT-4

With the fresh release of GPT-4, it’s no surprise that people are looking to compare the two. But it’s difficult to do so. Both companies are guarded about the technical details of their chatbots.

Like OpenAI, Baidu also decided to not reveal how many parameters there are in the latest version of Ernie. The number of parameters in a model is usually seen as an indicator of how powerful it is. Figures are available for their last-generation products: OpenAI’s GPT-3, released in June 2020, had 175 billion parameters, and Baidu’s Ernie 3.0 Titan, released in December 2021, had 260 billion parameters. 

Although Ernie Bot can’t analyze images like GPT-4, it does offer more output options. In the presentation, the chatbot read out the text answer in Sichuanese, a popular dialect spoken in southwestern China. Li also said the model can generate audio in other varieties of Chinese, like Cantonese, Hokkien, and the Dongbei dialect.

The quality of the answers it provides might be another matter. In a livestream after the launch, X.Pin, a Chinese tech publication, asked both Ernie Bot and GPT-4 some of the same questions in Chinese. While the Baidu technology could answer most questions coherently, it made more mistakes. It had trouble correctly answering trivia questions about Chinese history, remembering the context in which questions were posed, and generating code to make a mini game. The reviewers were also unable to test out the video generation ability. Ernie Bot refused to do so, saying it needed some time to edit and process the data.

Rushing it out for business partners

Earlier this week, the Wall Street Journal reported that to get Ernie Bot ready for the big launch, Baidu asked employees to work through public holidays, hired additional contractors to review the bot’s answers, and pooled resources like Nvidia’s A100 computing chips from other AI teams at the company. 

Since then, there have been other hints that the chatbot was not ready for wide deployment. Baidu had previously said that Ernie would be integrated into many of the company’s products, including self-driving vehicles and the flagship search engine. But the product release featured none of those applications or explanations of how such integration would work.

Many observers were disappointed that the release event used only pre-recorded videos of interactions with the chatbot, which can be easily filtered and edited. It was also pointed out that many of the multimodal functions showcased on Thursday can already be achieved with Baidu’s current AI tools, like the image-making AI from 2022 or a video editing tool it released in 2020, so the innovation is more about integrating them into one more accessible interface.

While Baidu has developed different kinds of AI models for years, Ernie Bot looks more like a way to package the company’s existing capabilities for business users to adopt more easily. 

And it’s clear that enterprise clients, instead of the general public, were the main target of this launch event. “Ernie Bot won’t just impact search engines and internet companies. It will impact every single company,” Li said during his presentation. “It will shorten the distance between every company and their customers.”

According to Baidu, 650 companies had signed up before Ernie Bot’s launch to use the technology, and more than 30,000 others have applied for the API access since the launch event. Previous news reporting suggests the companies interested in using the chatbot include the computer maker Lenovo, the travel portal Trip.com, and several Chinese automotive companies. While there’s currently no indication of what these partnerships may look like, we’ll likely find out more as Baidu rolls out the API in the coming months.