Opening Remarks at GenA.I. Symposium
Speeches
31 Oct 2025
Opening Remarks at GenA.I. Symposium
Carmen Chu, Executive Director (Banking Supervision), Hong Kong Monetary Authority
- Good morning, everyone. A very warm welcome to the GenA.I. Symposium.
- Today, we gather to celebrate the successful completion of the first cohort of our GenA.I. Sandbox. And we welcome the next wave of innovators in the second cohort.
- The concept of an “academy” is a fitting metaphor. The GenA.I. Sandbox is more than a testing environment; it’s an academy for the future of banking; a place of learning, and more importantly, of tangible progress.
- A year ago, we welcomed our founding class. They were pioneers, venturing into a fast-evolving curriculum, tasked with exploring GenA.I.’s potential in banking.
- After months of intensive trials at Cyberport’s A.I. Supercomputing Centre, the efforts have paid off. Today, it is my honour to announce that our first cohort has officially graduated.
- Their achievements are not just on paper, they are here today. I encourage you to visit all “Graduation Showcases” at the booths. Let’s engage with the teams. They are now our esteemed alumni, armed with practical insights that will benefit the entire industry.
- Beyond confirming GenA.I.’s potential in risk management, anti-fraud, and customer engagement, the first cohort transformed ideas into practice—delivering insights that redefine responsible implementation.
- First, we discovered the power of a “Specialised Major” over “General Studies”. A GenA.I. model, when strategically fine-tuned with domain-specific knowledge, excels at specialised tasks. For example, one team built a model for fraud investigation and reporting. By fine-tuning a smaller model on historical investigation reports, it achieved accuracy comparable to a much larger generic model, while using far less computing power.
- Second, we validated the value of “Targeted Tutorial Sessions”. Full model fine-tuning is like retaking an entire course—exhaustive and slow. Instead, methods like Low-Rank Adaptation, or LoRA, a focused, intensive tutorial—perfect for fast, efficient learning.
- For example, a team applied this “tutorial” concept to specialise a model in days instead of months, teaching it to generate personalised and relevant banking product recommendations to customers. The model mentored the new subject without forgetting its prior knowledge—offering a faster, more cost-effective path to expertise.
- Third, together we developed an “Anti-Hallucination Curriculum”. We know that we can’t just tell a model “don’t make things up”. We must create a structured learning environment—like schools do. This means:
- Providing a “Required Reading List”—using Retrieval-Augmented Generation, or RAG, to ground responses in factual sources.
- Giving clear “Exam Instructions”—via prompt engineering that demands citations.
- Offering “Feedback”—through fine-tuning to correct recurring errors.
- In just a while, one of our first cohort graduates will share how they have combined these techniques to transform their credit assessment model from a creative writer into a precise, compliant assistant.
- Building on this good foundation, we now proudly welcome our second cohort. The HKMA’s vision on this is clear and firm. Through the GenA.I. Sandbox and beyond, our banking system is lifting efficiency while staying resilient. We will advance further—creating purpose-fit solutions for increasingly complex challenges.
- Our next steps are to pioneer the “A.I. vs. A.I.” paradigms for security and quality assurance. Together we can push forward scalable, governed A.I. to build true resilience into our banking and financial systems.
- The number of technology partners has tripled for this cohort; it’s a testament to the vibrant ecosystem we are building. It’s not a solo endeavour anymore; it’s a collaborative consortium, and it’s bringing together experts from banking, technology, and academia to advance responsible A.I. adoption.
- The first cohort now becomes mentors. This cycle of learning, graduating, and giving back is key to securing Hong Kong’s leadership in fintech.
- Our aspiration doesn’t stop at the Sandbox. Inspired by the Sandbox, the HKMA initiated Project Noor together with the Bank for International Settlements Innovation Hub Hong Kong Centre and other regulators. “Noor”, the Arabic word for “light”, means that we seek to lighten up the “black box” of A.I. model, by developing an A.I. model auditing prototype. Through this collaboration, it will empower both banks and regulators to better oversee model risks—especially in explainability, robustness, and fairness.
- Today’s agenda is packed. Besides the exhibition, industry leaders will discuss GenA.I. from application, business, and hardware perspectives. And thanks to the Insurance Authority and the Office of the Privacy Commissioner for Personal Data, we’ve arranged a first-ever joint regulator panel.
- In a display of our own innovation, this event is powered by GenA.I. We are excited to demonstrate an interactive chatbot—developed with Amazon Web Services—to answer questions using insights from the Sandbox.
- The entire event features live transcription and translation. And our fireside chat with A.I. avatars will explore the future of human-A.I. collaboration—complete with a GenA.I.-powered photobooth. Our thanks to Votee.ai, BeingAI, BytePlus, and Vizper for making this possible.
- In preparing for today’s event, I used GenA.I. to generate a visual summary of my remarks. As you can see, it captures the key themes—progression, collaboration, and innovation—in one powerful image. It’s a compelling example of how technology can synthesise and communicate complex ideas.
- Finally, let me extend sincere gratitude to our co-host Cyberport, as well as our partners in the Hong Kong Association of Banks, the Hong Kong Institute of Bankers, and the FinTech Association of Hong Kong, for their support.
- Let us continue to learn, to share, and to advance—together—as one ecosystem. Thank you.
