ChatGPT: Smart chatbot or rogue machine power?

ChatGPT adoption by commercial companies has triggered a frantic competitive scramble. Tech leaders in Silicon Valley and China are galloping full speed ahead. Large language models finesse regenerative machine intelligence. Image, text, and voice replication has intellectuals worried about losing control of the technology. Is the new tech a threat to society or added value? Zhang Tianyuan reports from Hong Kong.

The ChatGPT artificial intelligence chatbot, from the San Francisco-based research lab that developed Whisper AI, showed netizens its awesome power to respond to prompts when it was launched on Nov 30. Within five days, it had attracted 1 million users. Paid versions with greater sophistication are now available. It is the digital tsunami sweeping the world with potentially a greater impact for good or bad than anything before.

Big-tech switches

Silicon Valley tech giants Google, Microsoft, Amazon, Meta and others have announced large layoffs and redirected mega-investments into regenerative AI development by acquiring small, agile innovators and supercharging their own teams.

Chinese tech giants Baidu Inc, Tencent and Alibaba, recovering from intense central authority scrutiny in 2020, are also racing to cash in on Chinese equivalents to ChatGPT. They hope to make their products work in both Chinese and English. As vast language databases are vital to train AI chatbots, using the large body of knowledge embedded in English-language websites, along with Chinese databases, is a double advantage. The Chinese mainland market access to the American products is blocked. That buys time for the local tech giants to raise their game.

Search-engine giant Baidu Inc launched its Ernie Bot in March. Tencent announced plans without a release date. SenseTime, backed by Alibaba, showed a suite of AI products — SenseNova and SenseChat — to assist computer coding, taking layman-level questions in English or Chinese, to translate into a workable product. Alibaba Cloud unveiled chatbot Tongyi Qianwen for beta testing by selected corporate clients.

The 2021 Global AI Innovation Index Report, published by the Institute of Scientific and Technical Information of China in March, revealed that China has 1,598 top AI scholars, surpassing the US figure of 1,483. Stanford Institute for Human-Centered Artificial Intelligence confirmed that China remained the top country for the number of AI journal publications between 2010 and 2021. In 2021, Chinese companies claimed to have created 21 large language models, up from the two produced a year earlier.

Vast database needed

The United States leads the AI revolution partly due to ChatGPT's earlier debut and also because it trains complex AI human-like responses with the much-larger English-language database. Internet World Stats records 26 percent of internet users worldwide use English, while 19.4 percent use Chinese.

The size and quality of the data determine the effectiveness of chatbots. Large language models like ChatGPT are trained on a mass of statistics from public resources, data vendors, and customer information. These AI models understand complex language patterns and generate high-quality responses.

Calvin Tang Siu-fung, a member of China Retold and Beta Gamma Sigma, a business honor society, said, "The English language encompasses a larger information corpus compared to Chinese." Apart from the limited data volume of the Chinese information database for training AI models, Tang noted that Chinese character have multiple meanings, which requires higher machine information processing compared with English text.

The "parameters" of language models, reflecting the size of the training database, reveals the gap. The ChatGPT-4 version registers 100 trillion parameters, more than 10 times the parameters in one of the largest Chinese-language models. According to Tang, the Chinese language models "need better Chinese interpretation".

Keith Li King-wah, chairman of the Hong Kong Wireless Technology Industry Association, said that cross-lingual models trained on English language databases cannot provide satisfactory Chinese responses, which is where Chinese internet companies can sharpen their competitive edge. "We expect the mainland's large language models to offer more precise Chinese answers," he said.

What about regulations?

The US government has taken a hands-off approach to regulating ChatGPT type programs, although concerns about personal data privacy are belatedly pushing Congress to act on that. The European Union took a tough stance on data privacy in its 2018 General Data Protection legislation. It is debating a wide regulation of regenerative AI applications, proposing to classify as "unacceptable" those applications to be banned outright, or "high risk" to be regulated, and tagging those to be left unregulated. The sentiment in Europe is anti-big business and pro-people.

China is urgently regulating regenerative AI development. Under the rules of the Cyberspace Administration of China, tech companies will be responsible for the "legitimacy of the source of pre-training data" and to ensure content reflects the "core value of socialism" The regulator is anxious to clamp down on fake news and uphold ethical and moral standards.

In May, mainland authorities arrested a ChatGPT user who wrote fake stories about a train crash that killed several people, and posted it on several social media sites, from where he generated revenue from clicks.

Chris Shum Chiu-fai, co-founder and chief financial officer of startup Asiabots Ltd, said the imposition of regulations on advanced AI models is a double-edged sword. "Regulations can enhance personal privacy protection and uphold social ethics, or they may potentially hinder technological development," said Shum, adding that striking a balance between fostering innovation and ensuring effective data protection is essential when formulating related laws.

Niche strategy

Guo Yike, a data scientist and provost of the Hong Kong University of Science and Technology, advised Chinese technology firms during a livestream on April 15 to shift from following ChatGPT, to improve on its limitations, and to embrace innovation to excel in the global AI race.

"The pace of iteration is accelerating rapidly, making it increasingly challenging to keep up. It is imperative for us to learn to stand on the shoulders of giants," he said.

Guo recommended that Chinese tech firms explore the development of smaller and more-efficient AI products. "The human brain uses a mere 20 watts of power when processing complex tasks, in contrast to the millions of kilowatt-hours consumed by ChatGPT each month," he added.

Moreover, he said, incorporating judgment of falsehood and value orientation into ChatGPT-like AI products could be a valuable innovation direction. "Bestowing ChatGPT with confidence, belief, and causality will elevate it beyond just a run-of-the-mill chatbot," Guo said.

"Currently, ChatGPT remains rudimentary in human-machine communication, lacking the ability to engage in exchanges to refresh its knowledge base." According to Guo, this limitation can yield inaccurate responses and a lack of critical thinking, leading to spreading misinformation or reinforcing bias.

Brad Smith, president of Microsoft, stated that Chinese research institutions and companies, including the Beijing Academy of Artificial Intelligence, which developed China's first large language model, Wudao, are poised to become major rivals to ChatGPT in the AI sector.

Microsoft is investing another $10 billion in OpenAI this year, following its investments in 2019 and 2021. Although ChatGPT is not directly available in Hong Kong, Microsoft Hong Kong has provided businesses in the city with access to its Azure OpenAI services. These include its enterprise-grade AI model and a preview version of GPT-4, available from April 1.

According to the China Academy of Information and Communications Technology, investment and financing in China's AI industry soared by 40 percent year-on-year, reaching $20.12 billion in 2021, representing 28 percent of the global total.

Chip restrictions

The intensifying trade sanctions imposed by the US complicate AI development. In October, the US announced export restrictions on technology to produce chips that are 14 nanometers or better.

The 14-nanometer limit refers to the size of each transistor on a chip. The smaller the size of the transistors, the more of them can be packed onto a chip, which boosts its speed and efficiency. ChatGPT type learning requires huge processing power, so the chip capacity becomes a limitation.

US semiconductor designer Nvidia Corp stated in a US Securities and Exchange Commission filing that the US government imposed a license requirement that prohibited the export of the A100 and H100 chips to the mainland and the Hong Kong Special Administrative Region.

China Retold and Beta Gamma Sigma's Tang said China has market advantages with its population of 1.4 billion, which is "highly attractive for overseas semiconductor providers. Abandoning the Chinese market entirely would not be beneficial to them".

Integrate English & Chinese AI

Hong Kong, as a gateway linking the Chinese mainland to foreign nations, holds great potential to become the country's test platform for both English- and Chinese-language models. It can bring together mainland and US AI models, to fuel a digital technology boom for the city.

Asiabots' Shum said ChatGPT can complement the language model developed in-house, to bolster the competitiveness of the firm's natural language processing technology for voicebots and chatbots in retail, fintech, medical, and other fields. "With the release of ChatGPT, we can update our AI products to field a broader range of questions beyond customized content," Shum said.

Roland Leung Tai-yin, co-founder and managing director of Datality Lab Ltd, is researching how to incorporate ChatGPT into its AI speech training platform. The Cyberport-incubated startup created Moodie.ai to enhance public speaking skills. It records the user's speech and evaluates, integrating body language, voice, speech content, and facial expressions.

"Integrating ChatGPT into Moodie.ai enables us to capture the speaker's logical thinking, ultimately improving the precision of the scoring system," Leung said. He believes advanced AI products from both the mainland and the US present vast business opportunities for Hong Kong enterprises.

The international financial center plans to establish an AI supercomputing facility per its 2023-24 budget. The city is expected to develop and train advanced language models. The HKSAR government also invests through its innovation and technology initiatives to promote AI-chatbot development within its digital ecosystem.

What's next

1. Hong Kong is well-positioned to test large language databases for joint English and Chinese artificial intelligence applications.

2. Identify niche opportunities to exploit AI in specific functions for small and midsize enterprises.

3. Provide extensive training to employees in AI skills to optimize applications.

4. Promote cross-training between Hong Kong and Shenzhen AI tech and academic experts.