人人都能懂的大語言模型：從基礎到實用應用（2026 年版）
本課程是針對初學者設計的實用性入門課程，介紹大型語言模型（LLM）如 ChatGPT 與 Gemini。無論您來自何種背景，此課程將以高階方式說明 LLM 的運作原理、其能與不能做到的事，以及如何在學習、工作和日常生活中有效運用。透過實際操作示範與導引練習，您將學習提示技巧、如何批判性評估輸出內容、如何處理幻覺與偏見，並安全且負責任地使用常見工具（例如文件處理、摘要生成、翻譯、資料任務）。課程結束時，您將能夠建立屬於自己的「LLM 工作流程」，用於實際任務——如撰寫文章、研究、規劃與提升效率，無需具備進階程式設計技能。
課程總覽
📚 內容摘要
本課程是針對初學者設計的實用性入門課程，介紹大型語言模型（LLM）如 ChatGPT 與 Gemini。無論您來自何種背景，此課程將以高階方式解釋 LLM 的運作原理、其能與不能之事，以及如何在學習、工作與日常生活中有效運用它們。透過實際示範與引導式練習，您將學習提示技術、如何批判性評估輸出結果、處理幻覺與偏見，並安全且負責地使用常見工具（例如文件、摘要、翻譯、資料任務）。完成課程後，您將能夠建立個人化的「LLM 工作流程」，用於實際任務——撰寫、研究、規劃與生產力提升——而無需具備進階程式設計技能。
從基礎數學邏輯到分散式代理協調：塑造大模型時代頂尖系統架構師。
🎯 學習目標
- 認知層面：理解機器學習的數學基石（線性代數、微積分、機率），並追溯神經架構從感知機到 LSTM 的歷史脈絡。
- 技能層面：使用 Unix shell 命令導航遠端伺服器，並利用自動微分引擎實現基本的計算圖。
- 情感層面：重視「理論根基」勝過「過早抽象化」，尤其在除錯複雜系統（如梯度爆炸）時尤為重要。
- 已生成
- 認知層面：說明後訓練管道的機制，包括監督微調（SFT）與強化學習（RL）框架（如 GRPO）之間的差異。
- 技能層面：設計多階段訓練流程——從冷啟動到最終對齊——並運用參數效率型微調（PEFT）技術，如 LoRA。
- 情感層面：重視從將 AI 視為「神奇黑箱」轉變為一個由機械層次與刻意內部推理構成的工程系統。
- 認知層面：比較線性整合架構與循環式、圖形化協調，區分垂直整合（MCP）與水平整合（A2A）協議。
- 技能層面：使用圖論原則定義特殊節點與條件邊界，並利用 FastMCP 實現 MCP 伺服器，連接代理至外部資料。
- 情感層面：重視「循環執行」與狀態管理在模擬複雜人類認知流程中的重要性。
課程 共 7 课时 · 预计 21.0h
課程
Lesson
This lesson challenges the "API fallacy" by emphasizing that true LLM mastery requires moving beyond high-level wrappers to understand the underlying mathematical foundations of linear algebra, calculus, and tensor mechanics. Students learn that grounding their practice in these core principles is essential for effective debugging, hardware optimization, and transitioning from cloud-based models to localized architectures.
This lesson explores the engineering mechanics behind Transformers, focusing on how Scaled Dot-Product Attention and the QKV framework enable models to process and predict text. Students will learn how to implement these concepts using matrix operations and stability techniques like positional encoding and layer normalization to ensure efficient, stable model training.
This lesson explores how raw base models are transformed into reliable assistants through a post-training pipeline involving supervised fine-tuning, reinforcement learning, and efficient adaptation techniques like LoRA. Students will learn how to build reasoning capabilities and optimize model performance on limited hardware by focusing on targeted parameter updates rather than full-model retraining.
This lesson explores the transition of prompt engineering into a formal discipline, emphasizing the use of Retrieval-Augmented Generation (RAG) and multi-provider orchestration to ensure system resilience and accuracy. Students learn to move beyond basic heuristics by implementing semantic chunking and architectural safeguards to mitigate hallucinations and prevent single-point-of-failure vulnerabilities in production environments.
This lesson explores the strategic shift toward local LLM deployment to address enterprise needs for data privacy, cost management, and operational resiliency. It also clarifies the critical distinction between OSI-compliant Open Source models and Open Weights models, emphasizing the importance of verifying licensing and usage restrictions for compliance.
This lesson explores the shift from linear AI chains to autonomous agentic workflows, which utilize graph-based architectures to enable cyclic execution, self-correction, and complex decision-making. Students will learn how to leverage state management, the Model Context Protocol (MCP), and multi-agent communication to build intelligent systems capable of reasoning and tool use.
This lesson guides students in transitioning from passive API users to expert architects by mastering autonomous system design, including the Model Context Protocol (MCP) and Agent-to-Agent (A2A) communication. It emphasizes building empirical engineering intuition through local pipelines, rigorous performance evaluation, and an understanding of advanced post-training alignment techniques like GRPO.