China enters the race to create a ChatGPT competitor – 05/04/2023 – Tech

The launch of ChatGPT, with a staggering 175 billion parameters, has impressed many people – including Robin Li, CEO and co-founder of Baidu.

Sensing a potential disruptive threat to the Chinese artificial intelligence industry, Li saw the need for his company to launch its own large language model (LLM) capable of competing with ChatGPT.

Li personally led the development. Baidu’s vice president of technology Wang Haifeng was tasked with executing the project. Leveraging more than a decade of artificial intelligence know-how, Baidu managed to become the first of the Chinese tech giants to launch a ChatGPT competitor.

On March 16, Li unveiled the Ernie Bot at the company’s headquarters in Beijing. Li praised the “chatbot’s” capabilities in terms of literary and business writing, its math skills, understanding of the Chinese language and multimodal generation — the ability to produce different types of content, such as images, based on text instructions.

However, it did not disclose the parameters of the model that underpins the system, nor did it offer the public a live demonstration. At the end of April, Ernie was still only available to test users.

The Alibaba group is generally regarded as one of the few companies in China that can compete with Baidu in LLM research. Three weeks after Ernie’s launch, Alibaba launched Tongyi Qianwen for invited business customers only. The artificial intelligence language model is capable of performing multiple rounds of dialogue, writing emails and code. And solve simple math problems.

On April 11, Alibaba Group Holding Chairman and Chief Executive Daniel Zhang announced that Tongyi Qianwen will be incorporated into all Alibaba products and services in the “near future”. The company’s Slack-like DingTalk app and its smart home app Tmall Genie became the first products to be integrated into the new AI model.

Armed with deep AI capabilities, Baidu and Alibaba are trying to take advantage of their pioneering position. During the Ernie launch, Li said that current artificial intelligence technology can be divided into four layers: the chip layer, the framework layer, the model layer and the application layer.

He added that Baidu has carried out self-developed innovations across the four layers – from its Kunlun artificial intelligence chip to the PaddlePaddle deep learning framework, to Ernie’s large pre-trained model and to search, cloud computing and self-driving vehicle applications.

Likewise, Alibaba has its own advancements in each layer: the Hanguang 800 artificial intelligence chip, the Easy Parallel Library, which is a deep learning framework for forming distributed models in the framework layer, the M6 ​​multimodal, and applications such as DingTalk and Alibaba Cloud.

By comparison, Tencent doesn’t seem to be in much of a hurry. With its AI development efforts spread across different departments, including WeChat and the AI ​​Lab, there hasn’t been a coordinated effort to create a unified product.

The company announced at the end of February that it had created a team to develop a tool similar to ChatGPT, called HunyuanAide, which will incorporate the Hunyuan artificial intelligence training model created by Tencent. Tencent Holdings chairman Martin Lau sees chatbots as one of the many AI products of the future and sees them as a long-term development opportunity rather than an area of ​​competition that needs to be addressed right away.

Meanwhile, ByteDance chose a different path. Instead of creating its own system, the company opted to offer its services to those developing LLMs through its cloud computing subsidiary, Volcano Engine. The company’s partners in the area of ​​large language models include artificial intelligence technology companies MiniMax, Zhipu AI and Kunlun Tech. Currently, there are dozens of large language model makers in China, and more than 70% of them use Volcano Engine’s cloud computing services, Tan Dai, an executive at the company, told Caixin.

While generative artificial intelligence is closely associated with ByteDance’s core business—short video and content delivery—the company actually has the least expertise and least developed technology in this area compared to its larger peers. Today, the company has two teams working on LLMs, one focused on language, led by Yang Hongxia, who previously oversaw the development of Alibaba’s M6 model, and another focused on audio and video, which could potentially be integrated into Douyin and TikTok. . However, Vice President Yang Zhenyuan said there is no release schedule yet.

In fact, companies like Baidu, Alibaba, Tencent, and Huawei Technologies had been creating LLMs for several years, but were primarily using them for research or enterprise-grade applications. They had not considered chatbots for public use. “They are too expensive,” said a Baidu executive.

The executive explained that, at Baidu, business development usually begins with identifying the commercial application, before product development, as in the case of artificial intelligence imagers and artificial intelligence compositing. However, in the case of ChatGPT-type products, it is up to users to identify what they want to use the tool for. And now, to stay ahead, all search services like Baidu will need to have a ‘chatbot’ plug-in, otherwise they will become useless, the person said.

In China today, ChatGPT is seen as the winning ticket to the next internet revolution. And companies that fail to secure one of those tickets fear being marginalized in the future. Since the ChatGPT frenzy, at least 17 companies have announced plans to step up LLM development.

Alibaba’s Zhang puts the point more bluntly. He believes it is worthwhile to restructure all industries, applications, software and services based on AI-generated content technology and LLM-supported technology.

While Chinese tech giants are diving headlong into the big language models race, their products are still far from being “easy to use”. In the case of Ernie’s image generation function, some people complained that the question was effectively translated into English before an answer was generated, and thus the system did not understand Chinese idioms and the names of certain local dishes. Many answers given by Tongyi Qianwen also contained factual errors and outdated information. Of course, these issues aren’t just limited to the aforementioned products – similar complaints have surfaced regarding ChatGPT.

Baidu’s Li acknowledged that ERNIE falls short of the latest version of ChatGPT. But she is confident it would only take “a month or two” to narrow the gap.

Not everyone shares his optimism. Yin Qi, CEO and one of the founders of Megvii Technology, believes that the next 12 months will be a crucial window for companies to introduce technological advances that can challenge the performance of ChatGPT-4. In his opinion, a maximum of two of them will find success.

Translated by Paulo Migliacci