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

Key Takeaways

• AI development should focus on augmenting existing LLMs (the 'vertical axis') rather than solely relying on developing new base models, as this is more accessible and practical for most.
• LLMs have inherent limitations including lack of domain knowledge, outdated information, difficulty in control, and performance degradation with long contexts, necessitating augmentation techniques.
• Prompt engineering is a fundamental skill, not a standalone profession, requiring an understanding of AI's capabilities and limitations ('jagged frontier') to avoid errors and use appropriate interaction models (Centaurs vs. Cyborgs).
• Prompt chaining, breaking complex tasks into sequential prompts with intermediate outputs, enhances LLM performance and provides crucial observability into the AI's decision-making process.
• Fine-tuning LLMs is generally discouraged due to high data requirements, risk of overfitting, poor timeliness, and often being less effective than prompt engineering, except in niche high-precision domains.
• RAG (Retrieval-Augmented Generation) is essential for providing LLMs with up-to-date, domain-specific information by retrieving relevant data chunks and incorporating them into prompts, overcoming context window limitations and hallucination risks.
• Agentic workflows transform LLMs from simple response generators into functional systems by integrating RAG, tool calls, memory, and multi-step decision-making, requiring a shift in engineering mindset towards managing goals and boundaries.
• Building agentic systems requires a new approach, treating data as unstructured text, embracing the probabilistic nature of LLMs, thinking like a manager, and adopting iterative, exploratory testing methodologies.