cs146s

https://themodernsoftware.dev/

Week1 Introduction to Coding LLMs and AI Development

• LLMs (large language models) are autoregressive models for next-token prediction
• Training Process
  • Self-supervised pretraining
  • Supervised finetuning（指令遵循）
  • Preferencing tuning（RL)
• Prompting
  • Zero-shot prompting
  • K-short prompting
    • in-context learning
    • Ideal for tasks that don’t have too many reasoning steps
  • Chain-of-Thought Prompting
    • Show reasoning steps for a given task（Multi-shot CoT, Zero-Shot CoT）
  • Self-consistency Prompting
    • 多次推理，汇总最常见结果（aka bagging）
• Tool Use
  • 与外部系统进行交互，减少幻觉
• Retrieval Agumented Generation（RAG）
• Reflexion（反思）
  • 类似：Now critique your answer. Was it correct? If not, explain why and try again.

some Terminology：

• System Prompt：
  • First message provided to LLM (usually not seen by end user)
  • Provides persona, rules about LLM output, style
• User Prompt：
  • The actual ask or instruction from a human
• Assistan：
  • LLM generates

最佳实践：

• Use our prompt improver to optimize your prompts - Claude Docs
• Clear prompting
• role prompting
• Prompts should be formatted with structure
• Be explicit about what you want (languages, tech stacks, libraries, constraints)
• Decompose tasks（分解任务）