Insights from the 2024 Inclusion·Bund Conference: From Data for AI to AI for Data

On September 5, 2024, the Inclusion·Bund Conference, co‑hosted by Ant Group, Shanghai Jiao Tong University and Fudan University, held the forum “From DATA for AI to AI for DATA,” where academia, industry and research representatives discussed the transformation of data value in the AI era.

Professor Zheng Weimin, a member of the Chinese Academy of Engineering and professor at Tsinghua University, highlighted that every stage of a large‑model lifecycle is tied to storage systems: massive multimodal small files during data acquisition, frequent random reads in preprocessing, intensive checkpoint I/O during training, and critical model‑parameter loading and intermediate‑result saving during inference, all driving new storage‑technology solutions.

Yan Shuicheng, chief scientist of Kunlun Wanwei & Tian Gong Intelligence and a Singapore Academy of Engineering fellow, noted that as model architectures continue to evolve, synthetic data will become essential; rather than simply augmenting existing data, high‑quality synthetic data should emerge from dialogues, discussions and evaluations among large models themselves.

Yang Chuanhui, CTO of the domestic distributed database OceanBase, introduced a unified system that supports SQL + AI and vector databases, leveraging AI techniques to enhance database development and management tools.

Chen Wenguang, director of Ant Technology Research Institute, argued that AI alignment must start from the underlying system layer, encompassing hardware architecture, programming languages and compilation systems, and presented the FABS (Fused AI, Big Data and Science) computing model.

Luo Ji, vice‑president of Ant Group’s Platform Technology Business Unit, described Ant’s next‑generation intelligent data system, which shifts focus from cost and efficiency to data‑value centricity. The system integrates a multimodal storage‑compute engine, a unified data lake, vector‑database capabilities, full‑modal cache acceleration for large‑model training, and comprehensive data governance, metadata management, security, and quality assurance mechanisms.