圖靈獎得主：AI Agent 最終全是資料庫問題

🔑 增強重點摘要

• •Stonebraker advocates for 'Data-Centric AI' where the focus shifts from model parameter scaling to the rigorous management of data quality, lineage, and schema integrity within enterprise environments.
• •The critique of LLM SQL generation stems from the 'semantic gap' between natural language intent and the complex, non-normalized schemas found in legacy enterprise data warehouses, which lack the clean metadata LLMs require for high-accuracy translation.
• •Stonebraker's perspective aligns with his ongoing work at companies like Tamr and DBOS, which prioritize deterministic data processing and transactional integrity over the probabilistic nature of current generative AI architectures.

🛠️ 技術深入

• •Stonebraker argues that AI agents acting as autonomous database interfaces require ACID (Atomicity, Consistency, Isolation, Durability) compliance to prevent 'hallucinated' updates that could corrupt transactional state.
• •The 0% accuracy figure cited for LLMs on production data warehouses refers to the inability of models to handle 'schema drift' and complex join logic that is not explicitly documented in the training corpus.
• •Proposed architectural shift: Moving away from 'black-box' LLM-to-SQL generation toward 'Neuro-Symbolic' systems that use LLMs for intent parsing but rely on deterministic, rule-based SQL generators for execution.

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線