Baidu, China’s top search engine provider and robotaxi developer, is apparently working on its own counterpart to ChatGPT. The news, first reported by Bloomberg and The Wall Street Journal, sent Baidu’s stock price rising on Monday to reach its highest point since September.
A spokesperson for Baidu declined to comment on the reports. But it wouldn’t be surprising that Baidu, which bills itself as the pioneer in China’s artificial intelligence field, is stepping up to build the Chinese equivalent of today’s most powerful chatbot. The question is how big a difference the tool can make, and where its limitations lie.
A driving force shaping China’s tech development over the past few years is the rise of digital sovereignty, which refers to a country’s ability to control its own “digital destiny” and can include autonomy in critical software and hardware in the AI supply chain. Episodes of U.S. export bans on China have pushed Beijing to further call for tech independence in areas ranging from semiconductors to basic research on AI.
As OpenAI’s ChatGPT shows the potential to disrupt sectors from education and news to the service industry, China’s tech leaders and policymakers are likely pondering how AI can also be used to drive productivity at home. China naturally wants its homegrown ChatGPTs, not just to secure control over how data flows through such tools but also to create AI products that better understand local culture and politics.
Slated to debut in March, Baidu’s conversational robot will first be integrated into the firm’s search engine, according to The Wall Street Journal. That suggests the chatbot will mostly generate results in Chinese. Nonetheless, the deep learning model is trained on both Chinese and English data sources, including information gleaned outside the Great Firewall, the country’s elaborate internet censorship infrastructure.
That’s where things get interesting. Like all other channels of information in China, the Baidu chatbot will no doubt be subject to local regulations and censorship rules. As we wrote earlier, the firm’s text-to-image application, ERNIE-VilG, already rejects politically sensitive prompts. But conversational AI handles much more complex inquiries than image generators — how will Baidu walk the line between censorship confinement and leaving enough freedom and creativity to its bot?
Also important to machine learning performance is the undergirding algorithms. According to The Wall Street Journal, Baidu adapted a “core breakthrough” that Google developed in 2017 and open-sourced, an algorithm that has also powered ChatGPT. Most likely, though, there are other key pieces of proprietary algorithms that Baidu has acquired or developed to form the backbone of its chatbot.
Hardware plays another important role in training large-scale neural networks. U.S. chip sanctions against China are posing a threat to China’s AI industry as companies lose access to advanced semiconductors that power supercomputers and large data centers.
Baidu, however, believes the chip ban has a “limited” effect on its AI business, as we reported. In the near term, the company “already stocked enough [chips] in hand.” As for the future, Baidu is counting on its Kunlun AI chip developed in-house to drive high-performance computing. Alternatively, it could work on increasing the efficiency of its algorithms to take some work off the hardware.