深度学习先驱 Jeremy Howard 探讨 AI 编程中的控制幻觉、交互式学习的重要性，以及他 20 年来致力于阻止不人道工作方式的使命。
Deep learning pioneer Jeremy Howard discusses the illusion of control in AI coding, the importance of interactive learning, and his 20-year mission to stop inhumane work practices.
要点 · TL;DR
大语言模型缺乏真正理解，无法外推到训练数据之外。 LLMs lack true understanding and cannot extrapolate beyond training data.
软件工程不同于编码；大模型在创新设计上失败。 Software engineering is distinct from coding; LLMs fail at novel design.
过度依赖 AI 编码会侵蚀人类能力和组织知识。 Over-reliance on AI coding erodes human competence and organizational knowledge.
核心观点 · Key points
在通用语料上预训练然后微调能创建强大模型，但它们缺乏真正的理解。 Pre-training on a general corpus then fine-tuning creates powerful models, but they lack true understanding.
LLM 擅长组合性创造，但无法外推到训练分布之外。 LLMs are good at compositional creativity but cannot extrapolate outside training distribution.
软件工程不同于编码；LLM 在新颖设计和架构上失败。 Software engineering is distinct from coding; LLMs fail at novel design and architecture.
像笔记本这样的交互式、有状态环境更适合人机协作和学习。 Interactive, stateful environments like notebooks are superior for human-AI collaboration and learning.
过度依赖 AI 编码会逐渐侵蚀人类能力和组织知识。 Over-reliance on AI coding erodes human competence and organizational knowledge over time.
反共识 · Contrarian takes
AI 编码工具制造类似老虎机的控制幻觉，而非真正的生产力。 AI coding tools create an illusion of control similar to slot machines, not genuine productivity.
当前 AI 生成的代码常常不可维护，且无人完全理解。 Current AI-generated code is often unmaintainable and no one understands it fully.
AI 的最大风险不是超级智能，而是权力集中在少数人手中。 The biggest risk of AI is not superintelligence but centralization of power in few hands.
笔记本实际上对 git 友好，使用得当能实现更好的软件工程。 Notebooks are actually git-friendly and enable better software engineering when used properly.
传统软件工程使用死文件是不人道的，且比交互方法效率低。 Traditional software engineering with dead files is inhumane and less effective than interactive methods.
本期章节 · Chapters（共 18）
开场：使命与理解本质Opening: Jeremy Howard's mission and the nature of understanding
ULMFiT 微调突破Fine-tuning breakthrough with ULMFiT
预文本任务与迁移学习Pretext tasks and transfer learning from vision to NLP
优化与分布假设Optimization and the distributional hypothesis
LLM 的浅层知识与创造力Superficial knowledge and creativity of LLMs
创造力与约束Creativity and Constraints
氛围编码与生产力Vibe Coding and Productivity
无银弹与软件工程No Silver Bullet and Software Engineering
知识：视角化与具身化Knowledge as Perspectival and Embodied
AI：学习工具 vs 拐杖AI as learning tool vs. crutch
自动化与组织能力Automation and Organizational Competence
AI 辅助调试体验AI-assisted debugging experience
软件工程 vs AI 编程Software Engineering vs. AI Coding
人机协作与交互环境Human-AI Collaboration and Interactive Environments
通过交互工具优化心智模型Refining mental models through interactive tools
连接数据科学与软件工程Bridging data science and software engineering
AI 存在风险辩论AI existential risk debate
AI 权力的集中与分散Centralization vs. Decentralization of AI Power