Hong Kong is blessed with the services of two of the world’s eminent technology and innovation creators, taking the city to the forefront in the global AI race with milestones in the making.
HK carves out a niche in AI
“Imagine talking to a machine that not only speaks like a human, but also has an unmistakable ‘Hong Kong character’. That would be beautiful,” says Guo Yike, the visionary behind HKGAI V1 — Hong Kong’s first homegrown artificial intelligence large language model.
The model emerged as a red-hot headliner following its public debut in February amid the global AI revolution.
The next major milestone for HKGAI V1? Guo says it’s for the model to obtain a good pass in a culture-specific Turing Test.
The Turing Test — named after British computing pioneer Alan Turing — is being commonly used to describe any evaluation to assess a machine’s ability to emulate human behavior.
READ MORE: CE says deeper ties with mainland boosted HK’s competitiveness
For a Hong Kong LLM to get through a “localized Turing Test”, its responses to prompts must be virtually indistinguishable from those of local residents, he says.
Data is key to realizing this vision. Training an LLM involves, in the simplest terms, sampling, analyzing and annotating relevant data before feeding it into the LLM for pattern learning. “What we need is refined data highly reflective of what it means to be a Hong Kong person — something that’s distinguishable from other communities,” says Guo.
Citing styling recommendations, he explains that when asking HKGAI V1 for advice on fashion, ideal outputs should embody “Hong Kong chic”. “The same goes for topics like food, housing, marriage, healthcare and education. When asked for opinions, the model should be able to give answers that reflect the common sense and values of the public within the city,” he says.
Local media reports constitute one source of such data, as well as the steady flow of social media content, says Guo, while noting there may be “difficulties” in acquiring certain data despite the ongoing data-sharing collaboration the HKGAI V1’s research and development team has with some Hong Kong-based newsrooms.
“There’s no easy shortcut for this. Asserting the legitimacy of your data collection and gaining owners’ full permission — these are all, rightly, factors to consider before accessing real-world data.”
Last year, the Office of the Privacy Commissioner for Personal Data — the city’s privacy watchdog — issued Hong Kong’s first generative AI-specific personal data protection guidelines for businesses which, among other things, urge developers to minimize the amount of personal data gathered for model training.
To proactively skirt privacy-related issues, Guo says his team is generating synthetic data — datasets created artificially, unconnected to real people — specifically to address HKGAI V1’s training needs.
He says HKGAI V1 remains strictly a not-for-profit, public-serving initiative led by the Hong Kong Special Administrative Region government, a position that potentially lends its appeal to all segments of society in contributing high-quality data.
Balanced participation
The model has been piloted in about 70 HKSAR government departments since mid-2024, with the trial’s scope continually expanding.
According to Guo, the number of staff involved in the trial run has increased from roughly 1,000 in February to around 20,000 as of early June.
Upgrades based on the feedback have been introduced monthly, with the latest iteration focused on enhancing the model’s abilities to digest and recommend academic papers, generate graphic and table content, and support more personalized material creation.
Guo says he appreciates the government taking the lead in Hong Kong’s AI development, adding that such an approach is “the best, most cost-effective, most goal-oriented, and fastest” path to fulfilling the city’s technology-driven ambitions.
“In this case, ‘government-led’ means much more than the government just injecting funding into R&D centers,” he says. “The cleverest part of all these is that the government took the initiative to be the first mover as a trial user.”
The government has extensive, yet well-defined application scenarios, covering various realms — from legal affairs, finance and environmental protection to education and other fields — that minimize the risk of aimlessly fumbling in the dark, Guo explains.
If R&D centers were to operate independently, their emphasis on academic achievements may downplay practical improvements to the user experience.
Governmental oversight also ensures that the feedback received for upgrades is not heavily swayed by market interests and profit-seeking motives that would present a challenge if leading corporations were to dominate the development landscape, says Guo.
However, the tech guru recognizes the team’s expanding needs for computing capacity for the model’s further evolution. “The demand is essentially distilled into two things — the amount of computing power available, and how much money is needed to purchase it.”
In his view, financial input from the private sector could significantly expedite the model’s ultimate opening up to public use.
A HK$200 million ($25.6 million) donation from the Ng Teng Fong Charitable Foundation and property developer Sino Group in March to the Hong Kong Generative AI Research and Development Center — the institution behind the HKGAI V1’s R&D team — marked the first of such private sector support, primarily aimed at providing computing power and manpower for an HKGAI V1-powered mobile application to residents.
In the long run, Guo is sanguine about Hong Kong-based entrepreneurs capitalizing on commercial opportunities to develop innovative, industry-specific applications after the HKGAI LLM evolves into a full-fledged foundational model for such initiatives.
‘A hundred flowers in bloom’
Guo, an avid art buff, cites the first project he spearheaded in Hong Kong in 2022 — an innovative art-tech trio comprising an AI choir, media artist and dancers — as a compelling demonstration of AI’s creative potential and the exciting prospects of AI-empowered arts, culture and creative industries.
“One of the vertical applications of HKGAI V1 now being tested for the transcription of minutes of meetings was developed from the voice recognition technology originally devised for the AI choir,” he says.
Hong Kong’s frontier industry insiders should have confidence in the city’s chosen path toward establishing itself as a world-leading AI center.
He argues that the notion of a singular “best development route” is a fallacy that fuels an untenable discussion of regional and global AI rivalries. Instead, AI prosperity is a scene with “a hundred flowers in bloom”, watered by diversity, openness and collective progress.
Guo considers his team, along with the developers at DeepSeek — a Hangzhou-based AI startup whose data learning model laid the groundwork for HKGAI V1 — as epitomizing the synergistic nature of AI development.
“Technical communication between us has been frequently initiated,” he says, adding that, on a broader level, exchanges between the HKSAR and the Chinese mainland in the AI realm have been gaining sustained traction long before HKGAI V1 entered the picture, primarily through cross-border talent interflow.
Guo, the provost of the Hong Kong University of Science and Technology, takes pride in the critical involvement of the school’s doctoral graduates in DeepSeek’s development of low-cost, open-source LLMs, an achievement that jolted the global tech community earlier this year.
In recent years, an average of 90 percent of HKUST’s PhD students are from the mainland, with many playing mainstay roles in leading businesses and research institutions across the country after graduation.
When asked whether his relatively recent arrival in Hong Kong poses cultural or language barriers to his ambition to install his brainchild with “a genuine Hong Kong character”, Guo’s answer is emphatic: “None whatsoever.”
“For us technicians and scientists, it’s always about technical advancement. What I do is to swim with the irreversible tide of our times while being self-driven by a common desire shared among peers to nudge higher.”
Shooting for the moon — pioneer’s dream tool
Make no mistake. It’s a mission renowned artificial intelligence trailblazer Yang Hongxia has in mind for the futuristic tool — creating a generative-AI language model for different professions.
The innovator, who is among the leaders behind the world’s first 10-trillion-parameter pretraining model, as well as a seasoned observer of the AI race in Silicon Valley and major Chinese internet firms, says that while tools like the GPT-4 perform well across general applications, large language models are far from perfect.
Her goal is to create a platform for Hong Kong that experts across various industries can partake in for AI development to break the data barriers in cutting-edge fields, such as healthcare and energy.
When American internet titan Google unveiled a new chatbot called Bard (now renamed as Gemini) in early 2023, the interactive agent falsely claimed that the James Webb Space Telescope had captured the first ever image of an exoplanet; AI experts call such occurrences “hallucinations”.
Such “hallucination problems”, linked to the absence of pertinent data references during model training, are common in professional domains, Yang tells China Daily.
The platform she is developing, with a decentralized architecture, would enable global experts to collaborate in developing domain-specific models without having to share raw data, significantly lowering the barriers of computational power and expertise.
Yang believes the Hong Kong Special Administrative Region, leveraging the advantages of “one country, two systems”, an internationalized research ecosystem, and the synergistic capabilities of the Guangdong-Hong Kong-Macao Greater Bay Area’s industrial chain, could be the starting point for democratizing the technological revolution.
With nearly 15 years’ experience in the industry, Yang joined the Hong Kong Polytechnic University last year, bringing with her the platform’s initial concept to break data barriers. “Having spent such a long time in this sector, I’ve learned that problems, especially for generative AI, are best solved by getting more people involved, not by restricting access to a very few people.”
As executive director, she set up the PolyU Academy for Artificial Intelligence in April, aiming to pilot the idea within the university and create a channel that would bring together experts from diverse disciplines on the campus.
“Imagine collaborating beyond physical limits, all experts at PolyU can embed their domain knowledge into AI training — from providing data to monitoring model development and testing,” she says.
The academy comprises 10 research centers focusing on urban energy, business transformation, intelligent manufacturing, robotics, medical reasoning, grid upgrades, new material manufacturing, education, construction engineering management, and neuromorphic computing.
According to Yang, each field will generate its own specialized models, with plans to expand into additional disciplines, such as medicine and genetics. “The academy is just the starting point.”
She says domain experts should get involved in the model-building process, highlighting its potential for groundbreaking advancements, particularly in healthcare.
The art of fusion
While general-purpose AI models like GPT-4 excel in broad applications, she says they often falter in specialized fields due to “hallucinations” — inaccurate or irrelevant outputs. “This stems from a lack of domain-specific training, not model capability.”
By leveraging a model fusion approach, experts can collaborate across different institutions, while maintaining strict data privacy protection. Such methodology enables the integration of specialized model parameters — whether they are from different or related fields — to drive transformative advances in cutting-edge technologies.
In healthcare, for example, Yang says her team can collaborate with top hospitals like Peking Union Medical College Hospital on liver cancer models and Zhejiang Cancer Hospital on lung cancer models. Rather than simply aggregating data from these two hospitals, she believes that by integrating them at the parameter level, these models can be fused together like building blocks to construct a comprehensive cancer foundation pattern.
In theory, directly combining data isn’t feasible due to constraints, such as privacy protection and data security concerns, she explains.
Her team has developed InfiR series models, and its 1B version’s reasoning capabilities have caught up with, and even surpassed international mainstream advanced AI language models, such as Llama3.2 and Qwen1.5B, created by Meta and Alibaba respectively, for tasks like text generation, translation and conversational applications.
In fusion models, the team has created InfiFusion models that can achieve state-of-the-art results by fusing Mistral, Qwen and Phi-4 with only 160 graphics-processing-unit hours.
The fusion at the model level not only ensures better data privacy. Decentralization also reduces the need for a large number of GPUs — a specialized chip for high-performance computing resources, ensuring their access is not restricted to large companies.
This innovative fusion technique needs just a small amount of GPU resources, making it applicable to small and medium-sized enterprises, as well as terminal devices with limited computing power, says Yang.
Google, for instance, has far surpassed OpenAI in terms of computational power and talent, yet OpenAI, which released ChatGPT, achieved that success with just a few hundred employees. Such a platform, the decentralization model and OpenAI’s experience, according to Yang, would be compatible with Hong Kong’s situation — a place with limited space and high labor costs, which render GPU usage quite costly for AI development.
By leveraging the concept of decentralization, platform-based infrastructure can be utilized at different locations, like supercomputing centers in Hong Kong or in other cities of the Greater Bay Area, or at sites like Pengcheng Laboratory — a research facility in Shenzhen specializing in network information, AI and cybersecurity. Such a setup would provide supercomputing resources for fundamental research and development within the AI domain, facilitating ongoing pretraining of domain-specific models.
According to Yang, the approach doesn’t require concentrating GPUs and datasets on one location, thus overcoming limitations related to cross-border data and physical aspects. Only a small-scale GPU card is needed to meet the training requirements of specific domain models, effectively dismantling the computational bottleneck.
Hong Kong’s strengths
“Hong Kong has favorable conditions to serve as a starting point as it’s really a special city,” says Yang, pointing to its abundant talent pool, superior geographical location, and a globally oriented research environment as vital for realizing her vision.
“No other city matches Hong Kong’s talent density — eight globally-ranked universities continuously supplying foundational talent,” she says, adding that the SAR also has strong appeal for top-tier professionals.
Citing her own experience, Yang says that less than a year after joining PolyU, she secured funding through RAISe+, a HK$10 billion ($1.27 billion) initiative launched by the HKSAR government in October 2023 to accelerate research commercialization.
Other programs among Hong Kong’s robust funding mechanisms to support scientific innovation include the Research Grants Council’s Theme-based Research Scheme launched in 2010, which grants up to HK$75 million per project to meet the city’s long-term strategic needs.
The funding programs demonstrate Hong Kong’s concentrated investment in strategic technologies and this level of support is truly irresistible, says Yang.
She also highlights Hong Kong’s strategic advantages in the Greater Bay Area, particularly its complementary industrial synergies with neighboring cities like Shenzhen which is home to the nation’s leading robotics industry and large-scale computing centers. She sees Hong Kong as a bridge between Eastern and Western cultures, facilitating global AI research and possessing a rich pool AI of talent.
“I want to promote the decentralized generative AI platform infrastructure with the city’s top talent and, probably, worldwide.”
To further cultivate Hong Kong as a hub for AI talent, she would like to see AI education introduced early in basic education, comparing its importance to that of mathematics.
Instilling a profound understanding of AI in students at secondary school, or even earlier, she believes, would ignite groundbreaking advances across various industries. “Eventually, we hope to establish an open ecosystem of generative AI to provide a platform for sharing models and solutions among experts in various fields in Hong Kong and even globally to collectively drive the development of AI.”
Yang is drawn to such supportive policies, and is driven in pursuing her personal goals.
ALSO READ: Chan: HK aims to be global innovation hub with AI at core
She has been a researcher at IBM’s Watson Global Research Center, chief data scientist at Yahoo, head of Alibaba DAMO Academy’s large-scale model M6, and a technical expert in large language model development at ByteDance, witnessing the rise of deep learning and the transformative impact of transformer models.
But, for her, it all began in the mathematics department of Nankai University in Tianjin, where she met Professor Shiing-shen Chern — a world-renowned Chinese American mathematician — as a freshman.
“It was exhilarating attending lectures by a legendary figure, who remained active in research well into his 90s,” she recalls.
While pursuing a doctorate at Duke University under renowned statistician David Dunson, a recipient of the COPSS Presidents’ Award, the highest honor in statistics, Yang specialized in Bayesian statistics, a cutting-edge field in AI at the time.
As a female scientist, she attributed the underrepresentation of women in AI largely to family responsibilities, but felt that women could excel in AI and other fields as they possess natural advantages, such as diligence and the ability to pay attention to detail.
Gender, she says, shouldn’t be restricted to STEM (science, technology, engineering and mathematics) or AI research. Everyone should persist in pursuing their passions, Yang adds.
“I firmly believe generative AI and distributed AI are the most critical frontiers for the future. Success hinges on execution. This is where I’m channeling all my energy.”
