On December 20, at the "Alpha Summit" co-hosted by Wall Street News and CEIBS, Wang Lei, General Manager of the AI Native Business Unit at Ant Digital, delivered a keynote speech titled "Exploring the Deep Waters of Financial AI: Practices and Insights on Vertical Model Implementation."
Wang noted that with the emergence of open-source foundational models like DeepSeek and Qwen, industry-specific AI models no longer rely on costly pre-training but instead adopt "post-training" approaches. This shift has reduced the iteration cycle for financial vertical models from months to just two weeks, while computing power requirements dropped from "tens of thousands of GPUs" to "hundreds of GPUs," slashing training costs by a hundredfold and significantly lowering barriers to industrial adoption.
He emphasized that deploying AI in regulated industries like finance demands rigorous attention to precision, expertise, and compliance. Wang highlighted that large language models (LLMs) cannot entirely avoid "hallucinations," which may even increase as reasoning capabilities improve. Thus, establishing systems and methodologies to mitigate hallucinations is the top priority for vertical industry applications, ensuring safety remains non-negotiable.
Wang outlined that the core of implementing LLMs in finance lies in building "trusted agents," which require three foundational pillars: 1. A "financial LLM" enhanced with industry data as the "brain"; 2. A "financial knowledge base" incorporating real-time and proprietary data as "experience"; 3. A "financial toolset" connecting to business systems as the "hands." Only by integrating these can AI function like a professional employee.
Wang further stated that LLMs represent not just a technological revolution but also a strategic reshaping of business operations. He urged companies to move beyond neutrality and rethink workflows entirely through LLMs to unlock transformative value.
Key Insights from the Speech: - The rise of models like DeepSeek has shifted industry focus from foundational model R&D to practical applications. - LLMs' breakthrough in natural language understanding has dramatically lowered human-machine interaction barriers. - Industries no longer need massive computing power or data retraining; instead, vertical-domain data augmentation offers a viable path. - LLMs are a "double-edged sword," requiring robust hallucination-suppression frameworks for vertical applications. - Model architecture defines capability ceilings, while safety sets the floor. - Evaluation is the starting point for LLM deployment, with agent development being an iterative, ongoing process. - Financial vertical models now require only 1% of the computing power (hundreds vs. tens of thousands of GPUs) compared to pre-training. - Trustworthy AI in finance hinges on precision, expertise, and compliance. - Strategic reinvention, not incremental neutrality, should guide LLM adoption in business workflows.
Historical Context: Wang traced AI's evolution over the past decade, from AlexNet's 2012 ImageNet victory (pioneering convolutional neural networks for image recognition) to Alipay's QR code payments and AlphaGo's 2016 triumph. He highlighted how decision models enabled AI-driven financial services like Yu'ebao and Huabei, paving the way for today's trillion-parameter LLMs that revolutionized natural language understanding.
Ant Group's AI Roadmap: As part of Alibaba ecosystem, Ant Digital positions 2024 as the "Year of AI Agents" for industrial ToB applications. Open-source models like DeepSeek have democratized access, enabling vertical model fine-tuning with domain-specific data—reducing iteration cycles from 3-6 months to weeks.
Financial AI Challenges & Solutions: Banking CIOs cite six pain points: limited computing power, poor data quality, rapid model iterations, knowledge gaps, lack of implementation frameworks, and talent shortages. Ant's "trusted agent" approach addresses these via: 1. Two-stage training blending general and financial data (100x cheaper than pre-training). 2. Built-in "safety guardrails" embedding regulatory knowledge. 3. Continuous evaluation and tool-based corrections (e.g., API calls for accurate calculations).
Conclusion: Wang closed by stressing that LLMs demand reimagining both technology and business strategy, urging organizations to break conventional paradigms. The speech underscored Ant Digital's vision for responsible, high-impact financial AI.
Comments