蚂蚁数科王磊：垂直大模型训练成本呈百倍级下降，金融AI落地需构建“可信智能体”三大基石 | Alpha峰会

Core Insights - The emergence of open-source foundational models like DeepSeek and Qwen has shifted the focus of the industry from expensive pre-training to a "post-training" model, significantly reducing the iteration cycle for financial vertical models from months to weeks and lowering computational requirements from "ten thousand cards" to "hundred cards," resulting in a hundredfold decrease in training costs [1][7][15]. Group 1: AI Implementation in Finance - The application of AI in serious industries like finance requires a focus on rigor, professionalism, and compliance [3][8][17]. - A "trustworthy intelligent agent" in finance relies on three pillars: a financial model as the brain, a financial knowledge base for experience, and a financial toolset for execution [3][20][21]. - The introduction of large models has revolutionized natural language understanding, significantly lowering the barriers for human-computer interaction [4][14]. Group 2: Challenges and Solutions - The financial industry faces six major pain points in implementing large models: limited computational power, insufficient and low-quality data, rapid model iteration, lack of knowledge accumulation, absence of application methodologies, and talent shortages [16]. - To address these challenges, a robust system to suppress "hallucinations" in large models is essential, as these hallucinations can increase with enhanced reasoning capabilities [3][5][17]. Group 3: Training Methodology and Future Outlook - The training of financial models should adopt a two-phase approach, balancing general and financial data to enhance capabilities without compromising general knowledge [23]. - Continuous evaluation and iteration of intelligent agents are necessary, treating their development as an ongoing process rather than a one-time software delivery [6][23]. - The application of large models in industries is not just a technological transformation but also a strategic business reshaping, necessitating a departure from traditional workflows [9][10][24].