The half-hearted presentation of an alternative to ChatGPT in China underscores the difficulties faced by local firms in the artificial intelligence race in the context of technological rivalry with the United States and censorship.
Chinese internet giant Baidu unveiled “Ernie Bot” in Beijing last week, its answer to American chatbot ChatGPT, whose skills are being passionately followed in China despite being locked down.
ChatGPT, launched in November by the Californian start-up OpenAI, allows you to formulate detailed answers on a wide variety of topics, write essays or create audiovisual content in just a few seconds.
ChatGPT is not accessible in China without bypass software (VPN) and foreign phone number. Despite this, it is popular as it is the subject of many media articles and discussions on social networks.
Baidu was one of the first Chinese groups to position itself to offer an equivalent to ChatGPT.
However, Ernie Bot’s press presentation last Thursday was limited to pre-recorded footage and a simple algebraic equation to solve.
– Sensitive content –
No live interaction with the chatbot was offered, in a country where digital companies are constantly pressured by the authorities to remove any content deemed sensitive or politically incorrect.
Baidu shares fell nearly 10% in the stock market before rebounding the next day, buoyed by positive comments from a handful of testers of its robot, including analysts at Citigroup Bank.
No general public launch date has been announced for Ernie Bot either, which works in Mandarin and is aimed exclusively at the Chinese market.
The headache for developers in China is to offer a powerful conversational robot that does not deviate from the very strict permitted framework in terms of content.
For example, if you ask him if Chinese President Xi Jinping is a good leader, a conversational robot designed by Beijing’s renowned Tsinghua University will step in and prompt you to “enter a new request.”
“Content regulation and censorship in China” are clear obstacles, says Lauren Hurcombe, technology specialist at law firm DLA Piper.
As a result, Chinese firms have “much less data” than their Western competitors to feed and train their systems, Ms Hurcombe tells AFP.
– fleas –
Besides Mandarin, Ernie Bot understands some Chinese dialects but doesn’t speak English well, the Baidu chief admitted Thursday.
Another challenge for Chinese companies: access to American technologies, such as chips that can run conversational robots’ algorithms, at a time when Washington is tightening restrictions in the name of national security.
China aims to become the world leader in artificial intelligence by 2030, which will revolutionize a wide range of sectors, including automotive and medicine.
But it will “no longer have access” to the most efficient chips running complex artificial intelligence, warns Lauren Hurcombe, who marvels at the ability to find such efficient Chinese alternatives.
“In 2016, China was already building world-leading supercomputers using its own chips,” said Steven Miller, professor emeritus of information systems at Singapore Management University (SMU).
“If that was the case seven years ago, it certainly has more capacity today to design high-end chips,” Miller told AFP.
Baidu is prepared and is already making its own chips.
– billions of parameters –
All Chinese tech giants say they are preparing a conversational robot, like internet and video games heavyweight Tencent or e-commerce champions Alibaba and JD.com.
Equal enthusiasm in the US, where computer giant Google on Tuesday launched public access to its bard with the aim of improving the quality of its answers.
But in the artificial intelligence race, the United States had twice as many startups as China and three times more private investment in the space in 2021, according to the latest available data from Stanford University, which is investigating these questions.
Two years ago, the Academy of Artificial Intelligence, a state institution founded in 2018 to stimulate innovation, had already developed a precursor to ChatGPT in Beijing.
It was dubbed Wu Dao and described by its creators as “the world’s largest artificial intelligence language model”.
At that time, it had 1.750 billion parameters, which is 100 times more than the previous version of ChatGPT (GPT-3). But Wu Dao never really broke through.