李礼辉：构建可信任的数字金融 | 金融与科技

Core Viewpoint - Trustworthy digital finance should possess characteristics such as model reliability, strong interpretability, and high security, while also clarifying the legal status, behavioral boundaries, and responsibilities of financial intelligent agents [2][12]. Group 1: Breakthroughs in AI Models - China's DeepSeek-V3 has received high praise in global AI model rankings, being compared favorably to GPT-4o, with training costs significantly lower at under $6 million compared to GPT-4o's $100 million [4]. - Innovations in algorithms, such as MLA multi-head potential attention mechanisms and MoE mixed expert architecture, are crucial for the future of AI development in China, particularly for financial institutions [4][5]. Group 2: Challenges in AI Technology - Security risks remain prominent, including unauthorized access to models, data theft, and malicious attacks that can compromise model integrity and stability [8]. - The phenomenon of "model hallucination" persists, with various models including Grok-3 and GPT-4 exhibiting certain levels of hallucination rates [9]. - Issues such as model bias, algorithmic resonance, and privacy breaches continue to pose challenges, complicating the interpretability of AI models [10]. Group 3: Digital Finance Innovation - The evolution of digital finance must balance security and efficiency, transitioning from mere usability to leading-edge capabilities [12][13]. - Trustworthiness in digital finance innovation is essential, requiring proactive measures to prevent AI pitfalls and ensure model reliability and interpretability [13]. Group 4: Pathways to Building Trustworthy Digital Finance - High reliability is critical, necessitating the implementation of advanced security measures, including firewalls and zero-trust architectures, to protect against malicious attacks [15]. - Interpretability is a key requirement, enabling the transformation of model behavior into understandable rules and utilizing visualization tools to clarify model processes [15]. - Legal frameworks must be established to define the status and responsibilities of financial intelligent agents, ensuring they operate within clear boundaries [16]. - Economic efficiency can be achieved by pre-training industry-level financial models and customizing enterprise-level applications, fostering collaboration between tech firms and financial institutions [16].