一部影片看完 Stanford AI 系統課程，從 LLM 到 Agentic Workflow

Key Takeaways

• Large Language Models (LLMs) have four major limitations in business applications: lack of domain knowledge, outdated information, difficulty in controlling output, and degraded performance with long contexts. Relying solely on more powerful base models cannot solve these issues.
• The three core tools for enhancing a single LLM are Prompt Engineering, Fine-Tuning, and Retrieval-Augmented Generation (RAG), with RAG being the most cost-effective and practical approach.
• The most important technique in Prompt Engineering is Prompt Chaining, which breaks complex tasks into multiple independent prompts. Each step can be tested and debugged independently, improving observability and output quality.
• Fine-Tuning should be avoided when possible because it is costly, prone to overfitting, and has poor timeliness; it can often be replaced by Prompt Engineering. It is only worth considering in high-precision fields like law and science.
• RAG retrieves relevant document chunks from a vector database and feeds them to the LLM along with the prompt, effectively addressing issues of information timeliness, hallucinations, and inaccurate long-context retrieval.
• Agentic Workflow combines prompts, external tools, and components into a structured workflow, transforming AI from a passive question-answer system into an active task-executing system.
• Building an agentic system requires a shift in engineering mindset: data from structured to free text, logic from deterministic to fuzzy, architecture from precise control to goal-oriented, and testing from deterministic to iterative exploration.
• In practice, follow the principle: 'Use deterministic solutions for deterministic problems, and add guardrails for fuzzy parts.' For example, in an AI scoring system, design an appeal mechanism that allows human intervention to correct errors.