# The AI-pilled compounding startup
**作者**: Ann Miura-Ko
**日期**: 2026-04-14T04:25:48.000Z
**来源**: [https://x.com/annimaniac/status/2043908558115438687](https://x.com/annimaniac/status/2043908558115438687)
---
Over the last month I've been doing office visits with AI-native companies in San Francisco and we’ve had the chance to actually see how they work today and we saw a lot that surprised us. Here’s a short summary of what an AI-native company might mean for the future:
## The Product Manager is disappearing
On a single day where we visited five companies, we found only one full-time PM across all of them (even though one company had as many as 40 employees). Engineers talk to customers daily and own product decisions end-to-end.
The PM isn't being "augmented." With many startups we saw, the role is being absorbed into engineering and design.
## The most dangerous side effect: the feature factory
When you can build anything a customer asks for in a day, the temptation to build everything is overwhelming. Multiple companies told us this is their biggest strategic risk right now.
The ones winning this fight have hard constraints. One company's agents can only configure existing features through JSON — they literally cannot create new application code. Another uses squad-level North Star metrics to kill ideas before they ship. Several emphasized that the founder has to decide where the product has opinions and where it's flexible.
When execution is nearly free, taste becomes the moat but how a company organizes to make taste evident is still being decided.
## The stack is converging
Almost every company we visited runs the same core: Slack, Claude Code, GitHub. Codex for code review and Linear. Linear has not only survived the SaaSpocalypse, they are creating a roadmap for how to thrive.
Slack has also become a central orchestration layer for agents. Emoji reactions auto-create tickets. Bots report diagnostics and triage customer issues. Agents get tagged in threads and start working on a fix.
Six months ago, Cursor came up in every conversation. Today it gets mentioned sporadically. Today engineers are living in Claude Code. One researcher told us he'd run Cursor alongside Claude and kept asking himself why he even needed the second window. Troubling for all of these coding platforms: engineers don't seem terribly loyal or attached to any particular tool, which raises the question of how a coding platform maintains value over time unless they get the benefit of the data being generated by engineers and train the model - advantage Anthropic especially with news of Mythos.
## The people across the org are empowered to build real things
An enterprise account manager had been asking her product team for months to automate account uploads. Nobody prioritized it. Then she asked an AI agent in Slack. It was done in an hour.
An accounting team is writing database queries and using MCP to interrogate their own business data. A Chief of Staff is producing direct mail and marketing materials in under 30 minutes.
The most underestimated shift isn't what AI does for engineers. It's what it does for everyone else.
## The cost of experimentation has collapsed creating compounding impact
A researcher tests 10 interface designs, runs each for a day, and throws 9 away. A designer generates multiple competing iterations in separate tabs in under 6 minutes. A growth PM with zero coding experience built a full Meta Ads pipeline (strategy briefs, AI-generated video ads, automated posting to Meta) in two days.
Companies are using AI to simulate customers before real ones ever touch the product. One team built AI agents that play different user personas to stress-test their product without waiting for real feedback. Another runs hundreds of research interviews in a week instead of 50 in a quarter. One company built customer personas with full negotiation histories, communication preferences, and decision-making patterns and uses them to prepare for sales calls.
These companies are iterating 3-5x faster and that speed shows up in two ways. For some it means getting through a single experiment faster so they can run more experiments over the same period. For others it means running multiple experiments in parallel. In either case, the surface area of unknowns gets covered faster. Both the build and learn steps are compressing across the organization. Knowledge compounds.
But this is fundamentally a different way of operating. It reminded us of how warfare has shifted from fighter jets to swarms of drones and the impact on strategy in warfare. Something similar is happening in how companies operate.
## What comes next
We're continuing to visit companies and will publish deeper case studies with more specific examples in the future. But the pattern is already clear: the gap between companies that have internalized these practices and those still debating "AI strategy" is enormous — and it's widening every week.
If you're a founder building this way, we'd love to hear from you.
## 相关链接
- [Ann Miura-Ko](https://x.com/annimaniac)
- [@annimaniac](https://x.com/annimaniac)
- [61K](https://x.com/annimaniac/status/2043908558115438687/analytics)
- [Upgrade to Premium](https://x.com/i/premium_sign_up)
- [12:25 PM · Apr 14, 2026](https://x.com/annimaniac/status/2043908558115438687)
- [61.4K Views](https://x.com/annimaniac/status/2043908558115438687/analytics)
- [View quotes](https://x.com/annimaniac/status/2043908558115438687/quotes)
---
*导出时间: 2026/4/14 23:44:03*
---
## 中文翻译
# 深受 AI 影响的复合型初创公司
**作者**: Ann Miura-Ko
**日期**: 2026-04-14T04:25:48.000Z
**来源**: [https://x.com/annimaniac/status/2043908558115438687](https://x.com/annimaniac/status/2043908558115438687)
---
在过去一个月里，我走访了旧金山的多家 AI 原生公司，有机会亲眼目睹了他们如今的工作方式，许多发现让我们感到惊讶。以下是关于 AI 原生公司对未来可能意味着什么的简短总结：
## 产品经理（PM）正在消失
在我们一天走访的五家公司中，我们发现它们总共只有一名全职 PM（尽管其中一家公司拥有多达 40 名员工）。工程师每天与客户沟通，并全权负责产品端到端的决策。
PM 并没有被“增强”。在我们看到的许多初创公司中，这一角色正在被工程和设计职能所吸收。
## 最危险的副作用：功能工厂
当你可以在一天内构建出客户要求的任何功能时，构建所有功能的诱惑简直难以抗拒。多家公司告诉我们，这是它们目前面临的最大的战略风险。
那些在这场斗争中获胜的公司都有严格的约束条件。一家公司的智能体只能通过 JSON 配置现有功能——它们实际上无法创建新的应用程序代码。另一家公司使用小组级别的北极星指标（North Star metrics）来在想法落地前将其否决。有几家公司强调，创始人必须决定产品在哪些方面要有鲜明观点，在哪些方面要保持灵活。
当执行成本几乎为零时，品味就成了护城河，但公司如何组织架构才能让品味显现出来，这一点尚无定论。
## 技术栈正在趋同
我们走访的几乎每家公司都运行相同的核心组件：Slack、Claude Code、GitHub。Codex 用于代码审查，以及 Linear。Linear 不仅在 SaaS 末日中幸存下来，它们还在为如何蓬勃发展制定蓝图。
Slack 也已成为智能体的中央编排层。Emoji 表情符号会自动创建工单。机器人报告诊断信息并对客户问题进行分类。智能体在对话线程中被标记，并开始着手修复问题。
六个月前，Cursor 在每一次谈话中都会被提及。如今它只是偶尔被提及。现在工程师们工作在 Claude Code 中。一位研究员告诉我们，他曾将 Cursor 与 Claude 并行运行，并不断问自己为什么还需要第二个窗口。这对所有这些编码平台来说都是一个令人头疼的问题：工程师似乎对任何特定工具都不太忠诚或执着，这就引出了一个疑问：除非编码平台能从工程师产生的数据中获益并训练模型——尤其是 Anthropic，凭借 Mythos 的消息占据了优势——否则编码平台如何随时间维持其价值？
## 组织内的每个人都被赋予了构建真实事物的能力
一位企业客户经理几个月来一直要求她的产品团队自动化账户上传工作。没人重视这个需求。后来她在 Slack 里问了一个 AI 智能体。问题在一小时内就解决了。
一个会计团队正在编写数据库查询，并使用 MCP 来查询他们自己的业务数据。一位幕僚长（Chief of Staff）在不到 30 分钟内就制作出了直邮和营销材料。
最被低估的转变不是 AI 为工程师做了什么，而是它为其他所有人做了什么。
## 实验成本的崩溃创造了复合效应
一名研究员测试了 10 种界面设计，每种运行一天，然后丢弃了其中 9 种。一名设计师在 6 分钟内在不同的标签页中生成了多个相互竞争的迭代方案。一位毫无编码经验的发展（Growth）PM 在两天内构建了一个完整的 Meta Ads 流水线（策略简报、AI 生成的视频广告、自动发布到 Meta）。
公司正在利用 AI 在真实客户接触产品之前模拟客户。一个团队构建了扮演不同用户角色的 AI 智能体，以此来给产品施加压力测试，而无需等待真实的反馈。另一个公司在一周内进行了数百次研究访谈，而不是一个季度进行 50 次。一家公司构建了具有完整谈判历史、沟通偏好和决策模式的客户画像，并利用它们来准备销售电话。
这些公司的迭代速度提高了 3 到 5 倍，这种速度体现在两个方面。对于一些公司来说，这意味着更快地完成单个实验，以便在相同的时间内运行更多的实验。对于另一些公司来说，这意味着并行运行多个实验。无论哪种情况，未知的覆盖面都会更快地被探索。整个组织的构建和学习步骤都在压缩。知识在复利增长。
但这根本上是一种不同的运作方式。这让我们联想到战争是如何从战斗机转向成群的无人机，以及这对战争战略的影响。类似的情况也正在公司的运营方式中发生。
## 接下来会发生什么
我们将继续走访公司，并将在未来发布更具体的案例研究和更深入的例子。但模式已经很清晰了：那些内化了这些实践的公司与那些仍在辩论“AI 战略”的公司之间的差距是巨大的——而且这种差距每周都在扩大。
如果你是一位以这种方式构建产品的创始人，我们很乐意听到你的声音。
## 相关链接
- [Ann Miura-Ko](https://x.com/annimaniac)
- [@annimaniac](https://x.com/annimaniac)
- [61K](https://x.com/annimaniac/status/2043908558115438687/analytics)
- [升级至高级版](https://x.com/i/premium_sign_up)
- [12:25 PM · Apr 14, 2026](https://x.com/annimaniac/status/2043908558115438687)
- [61.4K 次观看](https://x.com/annimaniac/status/2043908558115438687/analytics)
- [查看引用](https://x.com/annimaniac/status/2043908558115438687/quotes)
---
*导出时间: 2026/4/14 23:44:03*