Taishin Bank, a commercial bank in Taiwan, is rolling out a series of internally developed large-language models for both internal and external uses, says Steve Sun, chief information officer.
The bank has partnered with OneDegree Global to provide the cybersecurity around the bank’s LLM.
Sun (pictured, left) says Taishin Bank’s goal is to become a regional bank, not just one in Taiwan (it has a branch in Hong Kong). One aspect of that big-picture objective is to leverage generative artificial intelligence.
Taishin experimented with OpenAI’s ChatGPT, but found it didn’t meet the bank’s needs. ChatGPT lacked the specific knowledge and the Chinese language chops to be effective, Sun says. Therefore in late 2023, the bank decided to develop its own LLM.
“Our LLM may not be as smart as ChatGPT,” Sun told DigFin, “but it knows all about Taishin Bank.”
Building the Brain
The bank developed its internal LLM in three ways. First it partnered with a Taiwan-based tech company, AI Lab, to feed its model with information about the finance industry. Second, it exposed the LLM to Taishin’s own brand and product information. Third, it trained it to operate with traditional Chinese characters.
The result is called ‘Taishin Brain’. The bank intends to deploy the first use cases by the end of the year.
One application is to enhance its existing chatbot services for customer service. Another is to help bank employees query for information internally. Sun says this should be especially useful for junior staff.
Looking ahead, he says Taishin Brain will become useful for marketing and communications. It will work on top of Stable Diffusion, a genAI image generator, to devise versions of corporate logos and marketing materials.
The bank is also keen to use Taishin Brain to generate reports and summaries on behalf of corporate clients. The bank is now experimenting with four open-source models, combining them with Taishin Brain to figure out which is the best for its needs.
Verifying the LLM
The bank isn’t doing this alone: it is relying on OneDegree Global to test the models for reliability and safety.
OneDegree Global is a Singapore-based cybersecurity vendor that is part of Hogn Kong-based insurtech OneDegree Insurance. Stanley Chou, chief information officer at the Singapore business (pictured, right), says the firm is providing AI verification to Taishin Bank.
“This ensures the bank’s LLM operates in a trustworthy manner,” Chou said. GenAI outputs can be notoriously wrong or misleading, so it’s OneDegree Global’s job to interrogate the bank’s various LLM developments to make sure they work properly and securely.
The OneDegree method is a version of alignment, a typical mix of tactics that tech companies use to ensure LLM outputs match the human intent.
This involves first generating questionnaires – robots questioning robots – to identify vulnerabilities or inappropriate responses. The second layer are libraries, one cataloguing potential risks (such as a compliance breach), and another for academic research on AI techniques.
All these ingredients get combined to produce hundreds of test cases, which OneDegree updates quarterly.
OneDegree’s heritage is insurance. DigFin asked whether its interest in cybersecurity is to gather the data it needs to offer cyber coverage. Chou says this vendor work helps it build expertise in AI-related risks. But the ultimate opportunity will lie with primary carriers.
“First we work with banks to learn,” Chou said, “then we can help the insurance understand how to quantify the risk.”
