在 AI 圈,几乎所有人都在讨论 AI 什么取代我,取代数学家、外观师、经济学家,但被「杀死」最多次的一定是程序员。

Anthropic 给出了一个谁都没料到的答案:在程序员被取代之前,先替换掉是程序员用了二十年的工具。
由于 Claude 写代码的速度实在太快,在 Anthropic 内部真正的瓶颈反而是人类过去几十年发明的软件工具,它们已经开始处理不了 AI 7×24 生产力的代码。

没办法,Anthropic 用一个 7 人团队,还是用 Claude,用 5 个月的时间,从零构建了一套全新的代码版本控制平台,把整个大仓库迁过去。
平均每 3 秒合入一次代码
所以,过去几个月里,Anthropic 内部究竟产生了多少代码。
今年 5 月,Anthropic 整个公司平均每天合并(Merge,即代码借助审核正式入库)的代码变更大约是 2500 次,这完全是一家大型科技公司相当活跃的水平。
但到了今年 9 月,短短四个月后,这个数字暴涨了 12 倍,达到了每天 30000 次。

如果换算成时间,意味着 24 小时不间断,平均每 2.8 秒,就有一段经过测试和审查的新代码被放进 Anthropic 的核心代码库里。
更可怕的是 「笔者预计这种增长速度短期内根本不会放缓。内部对合并代码的潜在需要几乎是无限的,笔者每扫除一个基础设施瓶颈,代码产量就会立刻迎来下一波暴涨。」

这套旧的代码版本控制平台到底怎么了。
统治编程界 20 年的 Git,接不住 AI
很多非工艺背景但关注 AI 的朋友,可能经常听说 GitHub 上多少多少个 Star 的项目,GitHub 就是全球最大的开源社区,也是无数 AI 项目的发源地,Meta 的 Llama、DeepSeek 的开源模型,代码都托管在上面。
GitHub 里的「Git」,就是全球几千万程序员每天都在用的版本控制平台。
或者如果你让 Cursor、Claude Code、Codex 改代码,最后「提交」(commit)走的也是这套平台。

如果把公司的代码库是一份所有人共同编辑的「超级在线文档」,那么 Git 就是底下的历史版本「时光机与管理员」。
它由 Linux 之父 Linus Torvalds 在 2005 年发明,负责记录谁在哪一秒改了哪一行字,确保几千人同时修改不会互相覆盖,随时还能一键回退。


在过去 21 年里,Git 是当之无愧的工业界绝对标准。它的底层基因里,写着一个属于 2005 年的假设:程序员可能在任何地方(主要是没有网络的地方)写代码。
为了让人在没网的飞机或咖啡馆里也能翻看历史记录、比较修改,Git 使用的是「去中心化」外观。每当笔者要干活,它默认把整个代码库的所有历史(在 Anthropic,这个包接近 100 GB),完完整整地复印一份下载到你的电脑上。
这个外观在「人写代码」的时代堪称完美,直到 AI 开始写代码。Git 引以为傲的「离线可用、人人持有完整副本」,在这个情形下纯属浪费。

每天 2500 万次的 Checkout,即在不同的分支之间切换、恢复文件、创建新分支等,每一次都要去拉取庞大的代码包,如果不加特殊优化,每天会产生超过 600 PB 的数据流量,这相当于每天在公司内网里来回搬运 1.2 亿部 4K 高清电影。
为了让 Git 服务继续,Anthropic 巅峰时期在内部维护了数百台 Git 镜像服务器做分流。但因为每秒钟都有新代码合入,镜像同步的速度根本赶不上 Claude 写代码的速度。
Anthropic 也发现笔者为什么要用一套为「断网人类」外观的平台,去服务一群「永远在线」的 AI 智能体?

为 AI 量身定制的 Scion
想通这一点后,Anthropic 决定不再给 Git 打补丁,直接重造轮子,推出了自研版本控制平台 Scion。

既然 AI 智能体永远在线,那就没必要在本地存 100 GB 的历史记录。谁最后改了这一行?两个版本差在哪?Agent 直接发个 API 请求问服务器就行,甚至连代码文件都不用下载到本地。
现在的 Scion 兼容「虚拟化 checkout」,AI 改哪个文件夹,就只拉取哪个文件夹;甚至只有当程序真正去读取某一行文件时,数据才会从内存缓存里流式传过去。
同时,Scion 砍掉了繁琐的本地打包和全局锁,AI 提交一次代码的耗时,只跟它改了多少行字有关,跟整个代码库有几百个 G 毫无关系。

相比之前那几百台支撑的 Git 镜像集群,Scion 消耗的 CPU 和内存规模降低 200 倍,各级缓存命中率超过 99%。
AI 和测试任务拉取代码要从几分钟缩短到了几秒钟。单单这一项提速,就帮整个公司的持续集成(CI)平台直接省下了 25% 的算力。

造这个平台的,也是 AI
如果放在五年前,一家科技巨头想要替换掉底层的代码控制平台,还要顺带手重写代码审查平台,Anthropic 同步自研了名为 Antique 的审查工具、代码搜索和权限平台,这意味着什么?
当年 Google 搞内部的 Piper 平台,Meta 搞 Sapling 平台,无一不是动用了数十甚至上百人的顶级基础架构团队,耗费数年时间才完成迁移。

但在 Anthropic,整个项目的团队规格是:7 个工程师,5 个月。
秘诀自然又是 Claude 深度参与了开发,他们使用一个内部早期模型把工程师手写的数据迁移管道重写了一遍,最终做到了内存占用降到原来的 1/20,速度快了 4 倍。

Anthropic 还在博客中说,这是一场奇妙的闭环:正是 Claude 强大的写代码能力把他们逼到了平台崩溃的绝境,也正是靠着 Claude 疯狂写代码,这 7 个人才能在 5 个月内造出这套高并发的分布式平台。
更有意思的是人的角色转变,将来人类顶尖工程师的真实工作状态会是怎么样呢。
这 7 位人类工程师既不用敲代码,也不是来审核 AI 写的代码,他们要搞清楚 AI 为什么要改动
他们发现,自己最常对 Claude 说的一个字,是「不」。
让 Claude 互相开展「对抗性代码审查」时,AI 往往会变得过于偏执,为了防范一些极小概率的假设性风险,写出过度复杂、过度工程化的代码。这时候,必须由人类架构师出手叫停,帮 AI 做减法。

Anthropic 表示,他们并不打算把 Scion 拿出来当商业软件卖,之所以把这套架构公开,是因为他们看到了不远处的将来。
当代码由成百上千个 Agent 在后台并行生成、测试、修复时,所有为「人类手速」外观的软件工程工具,从版本控制、代码审查,到测试流水线,全都会在瞬间成为创新的障碍。
Anthropic 最近已经发了太多「警告」,这次也不例外:随着 AI 编程工具在各行各业的渗透,其他公司也不需要等太久,大概在将来 6 到 12 个月内,就会撞上同一堵墙。
本文来源:https://www.ifanr.com/1684269
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
In AI circles, almost everyone is talking about what AI replaces me, replacing mathematicians, visualizers, economists, but the programmers must have been “killed” the most times.
Anthropic gave an answer that no one expected: it was the programmer who used the tool for twenty years before the programmer was replaced.
Because Claude writes code so quickly, the real bottleneck within Anthropic is the software tools invented by humans over the past few decades, which have begun to fail to handle AI 7 × 24 productivity code.

No way, Anthropic used a team of 7 people, or Claude, in 5 months, built a new code version control platform from scratch, and moved the entire large repository.
So how much code has been generated inside Anthropic over the past few months?
In May, Anthropic's average daily merger (Merge, that is, code formalized by audit) was about 2,500 code changes, which is a fairly active level for a large technology company.

But by September this year, just four months later, that number had skyrocketed 12-fold to 30,000 per day.
If converted to time, it means 24 hours of uninterrupted, on average every 2.8 seconds, a piece of new code that has been tested and reviewed is put into Anthropic's core codebase.

What is more frightening is that "I do not expect this growth rate to slow down at all in the short term. The potential need for merged code is almost unlimited, and every time I remove an infrastructure bottleneck, code production will immediately usher in the next wave of skyrocketing."
What happened to the old code version control platform?
Many friends with non-process backgrounds but concerned about AI may often hear about how many Star projects are on GitHub. GitHub is the world's largest open source community and the birthplace of countless AI projects. Meta's Llama and DeepSeek open source models are hosted on it.

Git on GitHub is the version control platform used by tens of millions of programmers around the world every day.
Or if you ask Cursor, Claude Code, Codex to change the code, the last "commit" goes to this platform.

If the company's codebase is a "super online document" co-edited by everyone, then Git is the historical version "time machine and administrator" underneath.
It was invented in 2005 by Linus Torvalds, the father of Linux, and is responsible for recording who changes which line in which second, ensuring that thousands of people do not overwrite each other at the same time, and can be reversed at any time with one click.
For the past 21 years, Git has been the industry's absolute standard. In its underlying genes, a hypothesis from 2005 is written: programmers can write code anywhere, mainly where there is no network.

Git uses a “decentralized” look to let people flip through history and compare changes in planes or cafes without a net. Whenever I have to work, it defaults to copying all the history of the entire codebase (in Anthropic, this package is close to 100 GB) and downloading a complete copy to your computer.
This appearance was perfect in the era of "people writing code" until AI began to write code. Git prides itself on being “available offline, with a full copy in everyone's possession,” which in this case is simply wasteful.

25 million checkouts per day, that is, switching between different branches, restoring files, creating new branches, etc., each time you have to pull a huge code package. Without special optimization, more than 600 petabytes of data traffic will be generated every day, which is equivalent to moving 120 million 4K HD movies back and forth in the company's intranet every day.
To keep the Git service going, Anthropic at its peak maintained hundreds of Git mirror servers internally for offloading. But because new code is being merged every second, the speed of image synchronization simply can't keep up with the speed at which Claude writes code.
Anthropic also found out why I use a set of platforms for the appearance of "disconnected humans" to serve a group of "always-on" AI agents?

After thinking about this, Anthropic decided not to patch Git anymore, rebuilt the wheels directly, and launched the self-developed version control platform Scion.
Since the AI agent is always online, there is no need to keep a 100 GB history locally. Who ended up changing the line? What's the difference between the two versions? The Agent can directly send an API request to the server, and even the code file does not need to be downloaded locally.

Now Scion is compatible with "virtualization checkout". If the AI changes the folder, it will only pull the folder. Even when the program actually reads a line of files, the data will be streamed from the memory cache.
At the same time, Scion cut out the tedious local packaging and global locks. The time it takes for the AI to submit the code once is only related to how many lines it has changed, and has nothing to do with the hundreds of Gs in the entire code base.
Compared with the previous hundreds of supported Git image clusters, Scion consumes 200 times less CPU and memory, and the cache hit rate at all levels is more than 99%.

AI and test tasks pull the code from a few minutes down to a few seconds. This single acceleration directly saves 25% of the computing power of the entire company's Continuous Integration (CI) platform.
If five years ago, a technology giant wanted to replace the underlying code console and rewrite the code review platform by hand, what does it mean that Anthropic developed its own review tool, code search and permission platform called Antique?

At that time, Google was engaged in the internal Piper platform, and Meta was engaged in the Sapling platform. It took dozens or even hundreds of top infrastructure teams years to complete the migration.
But at Anthropic, the team specifications for the entire project are: 7 engineers, 5 months.
The secret is naturally that Claude was deeply involved in the development, using an internal early model to rewrite the engineer's handwritten data migration pipeline, and finally reduced the memory consumption to 1/20 of the original, 4 times faster.

Anthropic also said in the blog that it was a wonderful closed loop: it was Claude's strong coding ability that drove them to the breaking point of the platform, and it was thanks to Claude's crazy coding that these seven people were able to create this highly concurrent distributed platform in 5 months.
What is more interesting is the transformation of people's roles. What will the real work status of top human engineers be in the future?

These 7 human engineers neither typed nor audited the code written by the AI to figure out why the AI was changing
They found that the one word they said most to Claude was “no.”
When Claude conducts "adversarial code reviews" with each other, AI often becomes too paranoid, writing overly complex and over-engineered code to guard against hypothetical risks with minimal probabilities. At this time, the human architect must stop and help the AI to subtract.

Anthropic said they didn't intend to sell Scion as commercial software, and they made the architecture public because they saw a not too distant future.
When code is generated, tested, and repaired in parallel by hundreds or thousands of agents in the background, all software engineering tools with a "human hand" look, from version control and code review to test pipelines, all become obstacles to innovation in an instant.

Anthropic has been issuing too many “warnings” lately, and this time around is no exception: as AI programming tools penetrate all walks of life, other companies won't have to wait long, and will hit the same wall in the next six to 12 months.
This article is from: https://www.ifanr.com/1684269 For copyright issues, please write to: 13951431913#139.com or submit a review on the right.