
Anthropic 的 Claude Code 让微软措手不及,将编程从编辑器转移到了命令行,微软首席执行官萨提亚·纳德拉表示。大约在六个月内,集成开发环境“就有点淡出了”,他在《实用工程师》播客中对格热戈日·奥罗什说。
Anthropic’s Claude Code caught Microsoft off guard by moving coding from the editor to the command line, Satya Nadella said. Within about six months, the IDE “kind of faded away,” he told Gergely Orosz on The Pragmatic Engineer podcast.
奥罗什是一名前微软工程师,他在雷德蒙德录制了这场一小时的采访。采访发生在微软 Outlook 上周四故障前几个小时,因此故障未被提及。纳德拉还谈到了微软的自愿离职方案、初级人才招聘以及检查 AI 生成的代码。
Orosz, a former Microsoft engineer, recorded the hour-long interview in Redmond. It took place a few hours before Microsoft’s Outlook outage last Thursday, so the outage did not come up. Nadella also talked about Microsoft’s voluntary exit offers, junior hiring, and checking AI-written code.
“AI 优先的产品开发毫不留情” 奥罗什询问,微软拥有 GitHub Copilot 和 VS Code,为何在 2025 年 Claude Code 出现时没有立即的应对。纳德拉表示,Copilot 最初是代码补全,随后 Cursor 构建了强大的代理模式,微软也进行了匹配。
‘AI-first product building is unforgiving’ Orosz asked how Microsoft, with GitHub Copilot and VS Code, had no immediate answer when Claude Code arrived in 2025. Nadella said Copilot began as code completion, and that Cursor then built a strong agent mode, which Microsoft matched.
“击中我们的,显然是 Anthropic 在其模型方面所做的出色工作,这才真正让命令行界面成为形态,”纳德拉说。
“The thing that hit us was obviously the great work that Anthropic did with their model, which then truly made the CLI the form factor,” Nadella said.
“当模型跃迁以非线性方式发生时,形态必须快速改变。从这个意义上说,AI 优先的产品开发毫不留情,”他说。“不要爱上你所构建的东西。”他表示,随着开发者将决策模型嵌入软件,集成开发环境可能会回归。他称这些模型为智能时代的存储过程。
“When model jumps happen nonlinearly, the form factor has to change quickly. AI-first product building in that sense is unforgiving,” he said. “Don’t be in love with what you built.”He said the IDE could come back as developers embed decision models inside software. He called those models the stored procedures of the intelligence era.
自愿离职和初级人才招聘 奥罗什直接询问微软的自愿离职方案是否是裁员的前奏。纳德拉没有回答这个问题。他说,该方案主要是为了给员工更多自主权。人们可以转入对他们长期经济利益有意义的角色。
Voluntary exits and junior hiring Orosz asked directly whether Microsoft’s voluntary exit offers were a prelude to layoffs. Nadella did not answer that question. He said the offer was mostly about giving staff more agency. People can move into roles that make long-term economic sense for them.
“如果你身处一个让自己感到职业停滞不前的团队,你可以用脚投票,”纳德拉说。“我希望自己能坐在这里说一切都会一样,但我做不到。”他补充说,许多这样的地方就在微软内部。他表示,微软并非每个部分都“收到了关于转型的备忘录”。他补充说,团队规模和工作流程将在公司20万人中发生变化。微软增长最快的工作之一是前沿部署工程师。
“If you’re in a team where you feel stuck professionally, you can vote with your feet,” Nadella said. “I wish I could just sit here and say it’ll all be the same. I cannot.”Many of those places are inside Microsoft itself, he added. He said not every part of Microsoft had “got the memo” on transforming. Team sizes and workflows will change across the company’s 200,000 people, he added. One of Microsoft’s fastest-growing jobs is the forward-deployed engineer.
关于应届毕业生,纳德拉表示,团队首先需要重组并建立他们新的“软件工厂”。招聘早期职业人才是在之后。他表示,基于当前的放缓得出长期结论“为时过早”。微软已与卡内基梅隆大学启动了一个名为PRAISE的导师计划,面向职业生涯早期的工程师。
On new graduates, Nadella said teams first need to reorganise and build their new “software factories.”Hiring early-career people comes after that. Drawing long-term conclusions from the slowdown now would be “too early,” he said. Microsoft has started a mentoring programme with Carnegie Mellon University, called PRAISE, for engineers early in their careers.
验证是难点“在一个代码生成变得廉价的世界里,变得困难的是验证,”纳德拉说。
Verification is the hard part“In a world where code generation has become cheap, what has become hard is verification,” Nadella said.
他说,用于批评代码的模型可能存在对齐问题,因此他倾向于用一个模型家族来检查另一个模型家族。他还希望生产痕迹能够输入一个循环来验证修复,他称之为“我的RSI”。纳德拉表示,公司应将AI代理视为“内部风险”。模型正在变得更强大,但并非更可靠,这一点他上个月在20VC上也提到过。他呼吁对代理的推理过程进行经过审计、防篡改的记录。
He said models used to critique code can carry alignment problems, so he favours checking one model family with another. He also wants production traces to feed a loop that verifies fixes, which he called “my RSI.”Companies should treat AI agents as an “insider risk,” Nadella said. Models are getting more capable, but not more reliable, a point he also made on 20VC last month. He called for audited, tamper-proof logs of how agents reason.
他还告诉公司,要将代理软件、内存和模型分开保存,以免任何供应商将它们锁定。他的测试标准是:一家公司应该能够更换任何模型,同时仍能在其私有评估中取得改进。
He also told companies to keep their agent software, memory, and models separate, so no vendor locks them in. His test: a firm should be able to swap any model and still improve on its private evaluations.
“如果你做不到,那么你就得问问自己,你如何才能成为一家主权公司,”他说。
“If you can’t do that, then you have to ask yourself how are you a sovereign firm,” he said.
“无计费的智能已在个人电脑上实现”,纳德拉表示。去年推出的前沿模型已经可以在笔记本上运行,上一季度的模型则可以在台式机上运行。他预计像 GitHub Copilot Fusion 这样的路由器将把部分任务发送到本地模型,其他任务发送到云端,他称之为“混合智能时代”的开始。他把本地模型描述为“无计费的智能”。他对工程师的建议是:少花时间阅读关于人工智能的文章,多花时间使用它。
Unmetered intelligence on the PC Last year’s frontier models can already run on a laptop and last quarter’s on a desktop, Nadella said. He expects routers such as GitHub Copilot Fusion to send some tasks to local models and others to the cloud, which he called the start of a “hybrid intelligence era.”He described local models as “unmetered intelligence.”His advice to engineers was to spend less time reading about AI and more time using it.
“如果我能改变我谈话与行动的比例,我会这么做。这是现在的时机,”纳德拉说。
“If I could change the ratio of my talk to do, I would do it. This is the time,” Nadella said.