Opening Remarks at FiNETech5 on “Launching the Second Cohort: The Next Epoch of GenA.I. Sandbox”
Speeches
28 Apr 2025
Opening Remarks at FiNETech5 on “Launching the Second Cohort: The Next Epoch of GenA.I. Sandbox”
Carmen Chu, Executive Director (Banking Supervision), Hong Kong Monetary Authority
- Good afternoon everyone. Welcome to the fifth FiNETech, featuring the next epoch of GenA.I. Sandbox.
- Last August, in partnership with Cyberport, we launched the GenA.I. Sandbox with a bold mission: to champion responsible innovation and accelerate the adoption of this transformative technology. Today, as we gather, our first cohort of trailblazers is already deep in the trenches of technical trials, exploring use cases that will redefine what’s possible in banking.
- The response to the first cohort was very impressive. We received over 40 proposals, and the evaluation committee has done an incredible job of prioritising ideas from a pool of exceptional quality. This enthusiasm speaks volumes; it confirms what we’ve long believed: A.I. is not just a tool, but a paradigm shift for banking.
- In just around three months into the GenA.I. Sandbox, we have offered over 20 supervisory feedback and technology deep-dive sessions with participating banks and their technology partners. Let me take this opportunity to share some early observations and insights from practical perspectives.
- First, data readiness is the cornerstone of success. High-quality data is essential for model training and testing, and in some cases, data collection and annotation can take significant time.
- Especially when scaling up GenA.I. use cases involving unstructured data or data in new formats, banks may benefit from establishing a repeatable pipeline to process unstructured data effectively. For example, banks’ fraud and scam investigatory work involves multiple data sources like financial proofs, application forms, and transaction histories. Some banks are usefully experimenting with GenA.I. to synthesise and process data to expedite data preparation.
- Second, a holistic approach to evaluating model performance is essential. While public benchmarks can provide a reference, banks may need to develop their own metrics for assessing the performance of the solution and creating a feedback loop.
- For instance, in customer-facing use cases such as GenA.I. chatbots, ongoing evaluation and post-deployment monitoring are essential to ensure consistent and professional responses to customers. Unlike conventional UAT (user acceptance test), which only has “pass” or “fail” criteria for each particular use case, banks will have to devise multidimensional metrics such as accuracy, completeness, and reliability to rate GenA.I. solution’s outputs over time.
- Third, fine-tuning strategies will require rethinking. Some banks have initially sought to fine-tune models for generating outputs, such as drafting funding proposals, in a specific structure, by injecting a large volume of past samples. However, with the recent releases of LLMs (large language models) with powerful reasoning and “mixture of experts” capabilities, whether extensive fine-tuning is still needed does worth revisiting critically.
- Initial testing reveals that state-of-the-art LLMs possess domain knowledge specific to banks’ risk management practices, which is in addition to the models’ broad spectrum of market understanding. Banks can explore tweaking their conventional fine-tuning strategies with alternative approaches, such as RAG (retrieval-augmented generation) or advanced prompt engineering, to achieve similar outcomes in a more time- and cost-efficient manner.
- While we are continuing with the first cohort, many of you may be wondering: “What’s next?”. Today, we are pleased to launch the second cohort of the GenA.I. Sandbox, a milestone that reflects both our early successes and the industry’s accelerating momentum. But innovation is never static. Just as we challenge participants to iterate and refine their ideas, we have applied the same ethos to the Sandbox itself.
- Building on the experience from the first cohort, we are doubling down on collaboration. At the heart of the GenA.I. Sandbox lies a powerful truth: breakthroughs happen when banking expertise and cutting-edge technology converge. During the first cohort, we witnessed remarkable partnerships blossom, with banks and technology firms co-creating solutions that blend domain knowledge with A.I.’s limitless potential.
- To supercharge this synergy, we will further introduce the GenA.I. Sandbox Collaboratory, a dynamic fusion of collaboration and laboratory. It is an early-stage launchpad for innovative ideas. For those asking, “Where do I start?” or “How do I find the right partner?”, the Collaboratory is your answer. Through a series of workshops, participants will brainstorm with technology experts to identify problem statements, design hypotheses, and prototype solutions. The Collaboratory is about building trust, alignment, and momentum, a jump-start for the GenA.I. Sandbox.
- Notably, one of the top problem statements that is concerning the banking industry is the growing threat of deepfake scams. In the first cohort, some participants have begun exploring more advanced anti-deepfake solutions. To further amplify our efforts in the light of the evolving threat, we recognise the need to deepen our collaboration.
- In response, a dedicated Collaboratory workshop focused on combatting deepfake attacks using A.I. will be held in the coming weeks. This workshop will bring together experts from digital banks and tech firms as well as key stakeholders to develop innovative and practical solutions. These efforts will complement our recently introduced “E-Banking Security ABC”, which is, “Authenticate in-app, Bye to unused functions, Cancel suspicious payments”, essentially adding “Deepfake detection” as the “D”.
- Combating deepfake threats with A.I. is a prime example of what we call “A.I. vs. A.I.”. But we don’t just stop here. In the GenA.I. adoption journey, as we push boundaries, we must also safeguard them. Therefore, the second cohort will place a strong emphasis on utilising A.I. to mitigate the risks and challenges posed by the adoption of A.I. itself. Through the GenA.I. Sandbox, banks are encouraged to examine the possibilities of integrating A.I. into the second and third lines of defence for risk management as they come up with innovative A.I. use cases.
- The ability to leverage the vast potential of “A.I. vs. A.I.” will be a foundation of a more agile and responsible A.I. adoption journey. With innovative control tactics such as automated monitoring, dynamic cybersecurity testing and adaptive guardrails, the Sandbox use cases will not only ensure the viability of novel operations, but also their resilience.
- Today’s agenda is a microcosm of what’s to come. We will hear first-hand from pioneers of the first cohort, gain insights from A.I. risk management experts, and connect with potential collaborators in our networking session.
- We recognise that innovation thrives in ecosystems, not silos. To continue fuelling this collective energy in GenA.I. adoption, we will continue to work closely with the first cohort participants to leverage the sandbox as a platform to uncover additional insights, which will be shared in an industry A.I. forum in Q4 this year.
- As we embody the essence of FiNETech, let me leave you with words from a famous quote: “Coming together is the beginning; keeping together is progress; working together is success.”
- Without further ado, let’s start FiNETech5. I also look forward to welcoming many of you into the second cohort and the Collaboratory. Thank you.
