Anthropic Claude Code 负责人 Boris Cherny 讨论 AI 如何改变软件开发，让每个人都能编程，并用构建者取代软件工程师。
Boris Cherny, head of Claude Code at Anthropic, discusses how AI is transforming software development, making coding accessible to everyone and replacing software engineers with builders.
要点 · TL;DR
自 11 月起 100%代码由 AI 编写，编码问题已基本解决。 100% of code is AI-written since Nov; coding is virtually solved.
使用 Claude Code 后，每位工程师的拉取请求生产力提升了 200%。 Productivity per engineer up 200% in pull requests with Claude Code.
软件工程师头衔将被‘构建者’取代，每个人都会编码。 Software engineer title will be replaced by 'builder'; everyone will code.
核心观点 · Key points
自去年 11 月以来，我 100%的代码由 Claude Code 编写，我没有手动编辑过一行。 100% of my code is written by Claude Code since November; I haven't edited a line by hand.
采用 Claude Code 后，每位工程师的拉取请求生产力提高了 200%。 Productivity per engineer has increased 200% in pull requests since adopting Claude Code.
编码问题几乎已解决；一两年内学习编码将不再重要。 Coding is virtually solved; in a year or two learning to code won't matter.
“软件工程师”头衔将被“构建者”取代；每个人都将成为产品经理并编写代码。 The title 'software engineer' will be replaced by 'builder'; everyone will be a product manager and code.
为 6 个月后的模型构建，而非今天的；押注更通用的模型。 Build for the model 6 months from now, not today; bet on more general models.
潜在需求是关键：观察用户如何滥用你的产品以发现下一个功能。 Latent demand is key: observe how users abuse your product to find the next feature.
反共识 · Contrarian takes
使用最强大的模型通常比便宜模型更省钱，因为消耗的 token 更少。 Using the most capable model is often cheaper than cheaper models due to fewer tokens.
团队资金不足会迫使工程师更多地使用 AI，从而带来更好的结果。 Underfunding teams forces engineers to use AI more, leading to better outcomes.
不要用严格的工作流程限制模型；给它工具，让它自己解决。 Don't box the model in with strict workflows; give it tools and let it figure it out.
印刷术是 AI 对编码和读写能力影响的最佳历史类比。 The printing press is the best historical analog for AI's impact on coding and literacy.
安全通过早期发布来研究；内部评估不足以反映真实世界行为。 Safety is studied by releasing early; internal evals aren't enough for real-world behavior.
计划模式（暂不写代码）用于 80%的任务；它显著提高了准确性。 Plan mode (no code yet) is used for 80% of tasks; it improves accuracy significantly.
本期章节 · Chapters（共 37）
引言与 Boris 背景Introduction and Boris's background
Claude Code 对软件工程的影响Impact of Claude Code on software engineering
反思 Claude Code 的影响与增长Reflections on Claude Code's Impact and Growth
Claude Code 起源与用户反馈Origin of Claude Code and user feedback
通过实验创新Innovation through experimentation
现状：100% AI 编写代码Current state: 100% AI-written code
下一前沿：AI 生成想法Next frontier: AI generating ideas
编码已解决，拓展至通用任务Coding is solved, branching into general tasks
Claude 修复内存泄漏Claude fixing a memory leak
原则：让 Claude 做与资金不足Principles: Let Claude do it and underfunding
你怀念编码吗？Do you miss coding?
编程之美与其目的The beauty of programming vs. its purpose
AI 作为编码助手AI as a coding assistant
对工作的影响与杰文斯悖论Impact on jobs and Jevons paradox
在 AI 时代蓬勃发展的建议Advice for thriving in the AI era
成为通才的建议Advice to be a generalist
Metaview 广告Metaview ad
AI 工具提升工作乐趣调查Survey on job enjoyment with AI tools
潜在需求原则Latent Demand Principle
Claude Code 的快速发展Claude Code's Rapid Development
用 Claude Code 构建Building with Claude Code
三层安全Three Layers of Safety
机制可解释性Mechanistic Interpretability
早期发布以学习安全Early Release for Safety Learning
安全与开源Safety and Open Source
对智能体的焦虑Anxiety About Agents
编码的演变Evolution of Coding
构建 AI 产品的建议Advice for Building AI Products
构建 AI 产品的原则Principles for building AI products
使用 Claude Code 的技巧Tips for Using Claude Code
对 Codex 与竞争的看法Thoughts on Codex and Competition
后 AGI 计划与味噌制作Post-AGI Plans and Miso Making
闪电轮：书籍推荐Lightning Round: Book Recommendations
Chrome 扩展与 Cowerk 使用Chrome extension and Cowerk usage