Alibaba’s fintech arm, Ant Group, has unveiled a large language model and related applications for the financial services industry that it thinks can offer advice to both professionals and consumers.
The LLM and applications were announced at a conference in Shanghai.
“Wealth managers can deploy the LLM to evaluate financial products, analyze markets and for investor education. Insurance service professionals could also use the LLM to explain insurance products, design family insurance plans and verify insurance claims,” promised a company blog post on Monday.
According to the author, Ant Group’s LLM distiguishes itself by specializing.
“General-purpose LLMs struggle to make sense of industry jargon and lack the domain expertise that helps financial professionals do their jobs,” wrote Alibaba managing editor, Alison Tudor-Ackroyd.
The LLM is currently being tested on Ant’s wealth management and insurance platforms, and the plan is to make it widely available on the fintech’s digital finance platform in China.
Ant Group began developing the finance-focused LLM at the end of 2022. It claims to have trained the model on “hundreds of billions of token datasets containing Chinese financial documents and over 1,000 billion tokens from general corpus datasets,” while drawing on over 600,00 instructions from more than 300 financial industry cases. It’s been fine-tuned using Ant’s general-purpose LLM.
One of Ant Group’s related apps is a financial assistant application for customers, Zhixiaobao 2.0, which awaits regulatory approval after six months of testing. Zhixiaozhu 1.0, a business assistant app targeted at industry professionals, is still in testing phase.
Ant Group claims Zhixiaobao 2.0 can provide market analysis, portfolio diagnosis, asset allocation suggestions and investor education. And suggests its advice is 95 percent accurate in recognizing financial intention, and its market analysis and reasoning is on par with the average professional. It’s noteworthy, however, that those figures were calculated using the group’s in-house evaluation tool.
Meanwhile, Zhixiaozhu 1.0 can purportedly help finance professionals with investment analysis, information extraction, content creation, business opportunity insights and using financial tools.
Chatbots powered by LLMs are known to be often inaccurate. And newer versions of LLMs appear not to be improving much: academics at Stanford and the University of California found OpenAI’s ChatGPT appeared to have worsened at generating some code and performing other tasks despite jumoing from version 3.5 to 4.0.
Native Chinese language LLMs have experienced their own struggles as the Middle Kingdom raced to deliver chatbots while remaining compliant with local requirements to avoid certain political issues.
But while flaws related to content can be funny, “What could possibly go wrong?” is perhaps too obvious a question given the predictable downside of financial loss. ®