《科创板日报》记者了解到，字节跳动豆包大模型视觉基础研究团队负责人冯佳时已于近期离职，下一 步动向尚未披露。冯佳时自2019年加入字节以来，长期专注于计算机视觉与机器学习方向的基础研究， 是字节内部典型的学术型技术人物之一。 公开资料显示，冯佳时本科毕业于中国科学技术大学，硕士就读于中科院自动化所，博士毕业于新加坡 国立大学（NUS）。 字节跳动豆包大模型视觉基础研究团队负责人冯佳时已于近期离职，下一步动向尚未披露。 在学术界，他曾任NUS电子与计算机工程系助理教授、机器学习与视觉实验室负责人，发表深度学习、 物体识别、生成模型、机器学习理论等领域论文400余篇。其成果屡获国际顶会认可，曾获MIT科技评 论"35岁以下创新者（亚洲）"、ACM MM最佳学生论文奖、ICCV TASK-CV最佳论文奖、CVPR 2021最 佳论文提名，并担任CVPR、ICML、ICLR、NeurIPS等顶会领域主席。 此前，已有多家AI企业出现"核心人员离职——快速再创业"的连锁现象，新团队获得资本青睐，延续了 对前沿研究方向的探索。 类似的人才流动并非中国独有。海外大模型企业也在经历高管与科学家"再分配"。今年上半年， Ope ...

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].