"It Just Really Didn't Work for the First 6 Months"

"前六个月几乎不能用"

Boris Cherny didn't set out to change how the world writes code. In late 2024, he joined Anthropic Labs — a tiny incubator team inside Anthropic, just a handful of people. The mission: find the product overhang. Find the thing the models could already do that no product had captured yet.

Boris Cherny 并没有打算改变世界的编码方式。2024 年底，他加入了 Anthropic Labs——Anthropic 内部一个小小的孵化团队，只有几个人。任务很简单：找到产品悬垂——模型已经能做、但没有产品去捕捉的事情。

Coding was the obvious target. The state of the art was tab-complete — press tab, get one line. But the team asked: what if you didn't need tab-complete at all? What if the agent just wrote everything?

编码是最明显的目标。当时最先进的技术是 Tab 补全——按 Tab，出一行代码。但这个团队问：如果根本不需要 Tab 补全呢？如果 AI 直接写出所有代码呢？

So Boris built Claude Code. And for six months, it barely worked. He used it for maybe 10% of his own code. Even after launch, it wasn't a hit. No exponential growth. The team was building for a model that didn't exist yet — and they knew it.

于是 Boris 做出了 Claude Code。前六个月，它几乎不能用。他只在自己的代码中用上了大约 10%。即便发布后也没有爆发。没有指数增长。团队在为还不存在的模型做产品——他们心里很清楚。

Then Opus 4 landed in May 2025. That's when the curve bent. Opus 4.5. 4.6. 4.7. Each model release bent it again.

然后 Opus 4 在 2025 年 5 月发布了。曲线开始陡峭。Opus 4.5。4.6。4.7。每个模型版本都把曲线再推高一次。

"For Me, It's 100%"

"对我来说，是100%"

Boris asked the room a question: who writes 100% of their code by hand? A few hands. Who writes 100% with an agent? A few more. Most were somewhere in between.

Boris 问了现场一个问题：谁 100% 手写代码？举手的没几个。谁 100% 用 AI 写？也没几个。大多数人介于两者之间。

For Boris, it's not a spectrum anymore. It's 100%. The Claude Code codebase — TypeScript and React — is written entirely by the model. He reviews a few dozen PRs on a typical day. On one day last week, he pushed 150.

对 Boris 来说，已经不存在什么光谱了。就是 100%。Claude Code 的代码库——TypeScript 加 React——全由模型写。他平时一天审几十个 PR。上周有一天，他提交了 150 个。

He's careful to qualify: this isn't true everywhere. Massive legacy codebases. Niche languages. Enterprise compliance. For those, the answer is the same one that's been true since Claude Code launched: wait for the next model.

他也小心地补充：并非处处如此。大型遗留代码库、小众语言、企业合规环境——这些地方还没到。但对它们来说，答案跟 Claude Code 刚发布时一样：等下一个模型。

"Most of My Work I Do from My Phone"

"大部分工作我在手机上完成"

Six months ago, Boris shared his coding setup on Twitter. It went viral — he hadn't realized it would be surprising. To him, it was just the way he coded.

六个月前，Boris 在推特上分享了自己的编码配置，意外走红——他没觉得这有什么特别的。对他而言，这就是他写代码的方式。

Today, most of his work happens on an iPhone. The Claude app has a code tab. Inside: 5 to 10 sessions with dozens or hundreds of agents each. At night, a few thousand agents run deeper work while he sleeps.

如今，他大部分工作在 iPhone 上完成。Claude 应用有个代码标签页，里面有 5 到 10 个会话，每个会话有几十到几百个代理。夜里，几千个代理在他睡觉时跑着更深入的工作。

The tool he's most excited about: loop. It's deceptively simple — use cron to schedule a repeat job. Every minute, every 5 minutes, whatever makes sense. Boris runs dozens:

他最喜欢的工具叫 loop（循环）。简单得不像话——用 cron 调一个定时任务，每分钟、每 5 分钟都行。Boris 跑着几十个循环：

• A loop babysitting PRs — auto-fixing CI, rebasing branches
• 一个循环看管 PR——自动修 CI，自动 rebase
• A loop keeping CI healthy — hunting flaky tests
• 一个循环保持 CI 健康——抓 flaky 测试
• A loop scraping Twitter feedback and clustering it every 30 minutes
• 一个循环每 30 分钟抓取 Twitter 反馈并归类

Anthropic also launched routines — same idea, server-side. Close your laptop and the agents keep running.

Anthropic 还推出了 routines（例程）——同样的想法，但在服务器端运行。合上笔记本，代理也不会停。

"Every Single Person on Our Team Writes Code"

"我们团队每一个人都在写代码"

Boris's prediction for what teams will look like: cross-disciplinary generalists. Not in the engineering sense — "they do iOS and web and server." Generalists across entire disciplines.

Boris 对团队未来的预测：跨学科的通才。不是工程师意义上的通才——"他会 iOS、Web 和服务器"。而是跨整个学科的通才。

The Claude Code team already looks like this. The engineering manager codes. The product manager codes. The designers code. The data scientist codes. The finance person codes. The user researcher codes. Every single person.

Claude Code 团队已经是这样了。工程经理写代码。产品经理写代码。设计师写代码。数据科学家写代码。财务同事写代码。用户研究员写代码。每一个人。

They're specialists in something — finance, design, research — but coding is now the universal layer underneath. The bottleneck shifted. Used to be: can this person code? Now: does this person understand the problem deeply enough to tell the model what to build?

他们各有专长——财务、设计、研究——但编码已经变成了底层的通用技能。瓶颈变了。过去是"这人会写代码吗"；现在是"这人理解了问题、能告诉模型要造什么吗"。

"Switching Costs Are Going Away. Network Effects Aren't."

"转换成本在消失，网络效应不会"

Drawing on Hamilton Helmer's Seven Powers — the canonical framework for business moats — Boris walked through which survive AI and which collapse.

借 Hamilton Helmer 的《七大力量》——商业护城河的经典框架——Boris 分析了 AI 下哪些能活、哪些不行。

Getting less important:

越来越不重要：

• Switching costs — you can use the model to port from one platform to another. Lock-in weakens when migration becomes a prompt.
• 转换成本——你可以让模型帮你把东西从一个平台搬到另一个。当迁移变成一句提示词，锁定就没那么牢固了。
• Process power — Claude 4.7 can hill-climb anything. Give it a target and tell it to iterate until it's done. Companies whose advantage is workflow are vulnerable.
• 流程优势——Claude 4.7 能够爬山式优化一切。给它目标，让它迭代到完成。靠流程立足的公司面临威胁。

Still mattering:

仍然重要：

• Network effects — AI doesn't change the fact that everyone being on the same platform is valuable.
• 网络效应——大家都在同一个平台上的价值，AI 改变不了。
• Scale economies — bigger still means cheaper per unit.
• 规模经济——规模大意味着单位成本低，这个逻辑没变。
• Cornered resources — if you own the thing, you still own the thing.
• 独占资源——你拥有的东西，AI 也夺不走。

Then the number that should keep incumbents awake: the volume of disruptive startups in the next decade will increase roughly 10x.

然后是一句让所有大公司睡不着的话：未来十年颠覆性初创公司的数量将增加大约 10 倍。

"Software Will Be a Skill Like Sending a Text Message"

"写软件会像发短信一样普通"

Boris reads two things: sci-fi and tech history. From the latter, he draws his clearest parallel:

Boris 只读两种东西：科幻和科技史。从后者那里，他找到了最清晰的类比：

Before the printing press, about 10% of Europeans could read and write. Literacy was a profession — scribes employed by kings and lords who couldn't read themselves.

在印刷术出现之前，大约 10% 的欧洲人识文断字。识字是一门职业——抄写员受雇于那些不会读书写字的国王和贵族。

After Gutenberg: in 50 years, more literature was published than in the thousand years before. The cost of a book dropped roughly 100x. Over centuries, literacy climbed to ~70%. Today, reading and writing aren't professions — they're baseline skills.

古腾堡之后：50 年内，出版的著作比之前一千年还多。一本书的成本下降了大约 100 倍。几个世纪后，识字率提升到约 70%。如今，读写不是职业，是基础技能。

Software will follow the same arc — much faster than 50 years. Domain expertise becomes the differentiator. Coding becomes the commodity.

软件将遵循同样的轨迹——只是比 50 年快得多。领域知识成为差异化的来源，编程变成通用商品。

"To the Model, It's Just Tokens"

"对模型来说，只是 token"

Asked what products to build today that get more interesting as models improve, Boris offered clear signals:

被问到现在该做什么产品、等模型变强后就会爆发，Boris 给出了明确信号：

• Claude Design — already good, about to get much better. Visual creation follows the same arc as code generation.
• Claude Design——已经不错了，即将大幅提升。视觉创作将走和代码生成一样的路线。
• Loops and batch processing — massively parallelizing agents will improve fast.
• 循环和批处理——代理的大规模并行化将快速提升。
• Computer use — slow today, but Claude 4.7 made it substantially better. The catchall for software without an API.
• 计算机使用——现在还很慢，但 Claude 4.7 让它好了很多。没有 API 的软件的最后一道万能入口。

But his deepest answer was about philosophy. When asked about local vs. cloud models, Boris said: "It doesn't matter."

但他最深层的回答是关于哲学的。被问到本地模型 vs 云端模型时，Boris 说："这无关紧要。"

In a couple of years, the model will decide whether to use local compute or cloud, start the agents, build the environments. The engineering decisions dissolve. What's left: knowing what to ask for.

几年后，模型自己会决定用本地算力还是云端算力，自己启动代理，自己搭建环境。工程决策会自然消解。剩下的只有：知道该要什么。