摘要
DeepSeek引领了生成式人工智能领域的技术变革,并具有混合专家模型的架构创新、知识蒸馏模式的技术升级以及开源模式下机器学习的资源富集这三大创新优势,但这三大创新优势同时也引发了法律监管风险。在DeepSeek的混合专家模型中,应该基于比例原则来优化门控网络的分配路径,消除过拟合现象,完善专家模型与门控网络的衔接路径并优化算法监管。在DeepSeek的知识蒸馏模式中,教师模型需要保持算法稳定性,基于可解释原则开展算法备案,同时根据学生模型的算法特征来避免数据泄露并合理处理数据。DeepSeek在开源后具有市场竞争优势,平台应该积极履行看门人义务来强化监管,在敏捷治理框架下治理DeepSeek生成的虚假信息,根据开源情况动态调整数据保护范围,对不同阶段的训练用数据、处理中数据以及新生成数据提供类型化保护,并基于控制理论来优化DeepSeek的机器学习模型,实现系统化监管。
DeepSeek has led the technological transformation in the field of generative artificial intelligence and has three innovative advantages:the architectural innovation of the Mixture of Experts model,the technological upgrade of the Knowledge Distillation mode,and the resource enrichment of machine learning under the open-source model.However,these three innovative advantages have also triggered legal supervision risks.In DeepSeek’s Mixture of Experts model,the allocation path of the gating network should be optimized based on the proportionality principle,the overfitting phenomenon should be eliminated,the connection path between the expert model and the gating network should be improved,and the algorithm supervision should be optimized.In DeepSeek’s Knowledge Distillation mode,the teacher model needs to maintain algorithm stability,carry out algorithm filing based on the interpretability principle,and avoid data leakage and reasonably process data according to the algorithm characteristics of the student model.After being open-sourced,DeepSeek has a market competitive advantage.The platform should actively fulfill its gatekeeper obligations to strengthen supervision,manage the false information generated by DeepSeek under the agile governance framework,dynamically adjust the scope of data protection according to the open-source situation,provide typed protection for training data,data in processing,and newly generated data at different stages,and optimize DeepSeek’s machine learning model based on control theory to achieve systematic supervision.
出处
《上海政法学院学报(法治论丛)》
2025年第4期112-133,共22页
Journal of Shanghai University of Political Science & Law(The Rule of Law Forum)
基金
2021年度国家社会科学基金重大项目“数字经济的刑事安全风险防范体系建构研究”的阶段性研究成果,项目编号:21&ZD209。
