Alex Liu

Writing · 2026年8月22日 · 6 分钟阅读

两个飞轮：AI 创业公司的管理模型

从第一次实习时偶然换办公室开始，我看到了一种让当下建设与面向未来的探索同时运转的组织方式。

本文基于我此前发布于 AirJelly 的英文文章修改。 查看 AirJelly 原文 ↗

组织观察两个飞轮

当下

  1. 建设
  2. 发布
  3. 支持

未来

  1. 观察
  2. 整理
  3. 探索

同一家公司，两个不同的时间线

我第一次实习是在一家美国怀俄明州的 AI 创业公司。我的职位是 AI Research，但实际做的是 AI PM 的工作。

有一天，我申请从大办公室搬到小办公室。原因很简单：大办公室的人下午才到，而我每天早上已经开始工作；他们下班晚，我总觉得自己早走不太合适。小办公室的人来得更早，我想，也许搬过去能早点结束一天。

这个看起来无关紧要的决定，让我看见了一种之后一直在思考的组织结构。

两个办公室，两条时间线

搬过去之后我发现，两个办公室不只是物理上分开，它们像是活在公司的两个时间线上。

假设当前产品正在 v1.0。大办公室的人在运行 v1.0、开发 v2.0；小办公室的人已经在讨论 v4.0，甚至开始设想 v5.0。

大办公室由成熟团队构成，负责把已经成立的产品继续建设、发布和支持。小办公室则是几位 AI PM 实习生，我们的工作是探索。

它们不是重复的两套团队，而是同一家公司里的两个飞轮：

  • 当下的飞轮：建设、发布、支持，让已成立的产品稳定向前。
  • 未来的飞轮：观察、整理、探索，为还没有进入主路径的问题保留空间。

一种刻意保留的不对齐

这听起来有些反直觉。大多数团队都在追求对齐，但这个结构刻意保留了一点不对齐。

如果探索与建设被放进同一套日常协作里，探索很容易被交接、排期和线上问题吞掉。紧急的事几乎总会先被处理，而重要但不紧急的事，往往没有真正属于它的时间。

小办公室的价值不在于制造更多“未来概念”，而在于保护一种不同的工作节奏：不必立刻交付，也不必先证明自己会被采用，但必须持续接近新的技术与真实问题。

没有人分配过的三个角色

当时小办公室原本有两位实习生，我加入后变成三人。名义上，我们的任务是给创始人提供灵感；实际上，创始人是公司里最重要的 PM。

因为每个人的时间安排更灵活，关注点也不同，我们慢慢形成了一种没有预先设计过的分工：

  1. Scout：持续追踪前沿 AI 进展，把原始信息翻译成团队能理解的输入。
  2. Systematizer：把零散输入整理成可复用的 SOP、方法或工作流。
  3. Bridge：把前沿探索带回一线，也把真实用户和社区反馈带回探索中。

这三个角色不是职位名称，更像是一条自然形成的链路：

前沿 → 系统 → 一线

一个办公室守住当下，另一个跑向未来

我并不认为每家公司都需要真的开两个办公室。但在 AI 创业公司或正在经历 AI 转型的团队里，这个模型值得被认真对待。

它不一定需要额外的人头或复杂的组织设计。更重要的是，要给面向未来的探索一块不被日常工作完全覆盖的空间；同时，让探索者、系统化的人和贴近一线的人之间形成持续的往返。

一个飞轮守住当下，另一个飞轮跑向未来。两者都不能停。

My first internship was at a Wyoming AI startup. My title was AI Research, but in practice I was doing AI PM work.

One day, I asked to move from the big office to the small one. The reason was simple: the big-office crowd did not arrive until the afternoon, while I was already there every morning, and I felt awkward leaving early. The people in the small office came in earlier, so I thought moving there might let me finish my day earlier too.

That unremarkable decision led me to an organizational structure I have been thinking about ever since.

Two offices, two timelines

After I moved, I realized the two offices were not merely separated by space. They seemed to inhabit different versions of the company.

Imagine the current product at v1.0. The big office was operating v1.0 and building v2.0. The small office was already discussing v4.0 and imagining v5.0.

The big office consisted of experienced full-time teams responsible for building, shipping, and supporting an established product. The small office consisted of AI PM interns whose job was to explore.

They were not duplicate teams. They were two flywheels inside one company:

  • The present flywheel: build, ship, and support what has already earned its place.
  • The future flywheel: observe, systematize, and explore questions that have not yet entered the main product path.

A deliberate misalignment

This sounds counterintuitive. Most organizations optimize for alignment. This structure deliberately preserved a small amount of misalignment.

When exploration and delivery share the same daily operating system, exploration is easily consumed by handoffs, deadlines, and production issues. Urgent work nearly always gets handled before important work that has no immediate deadline.

The small office was valuable not because it generated more futuristic concepts, but because it protected a different working rhythm. Its work did not need to ship immediately or prove in advance that it would be adopted. It did, however, need to stay close to new technology and real problems.

The roles nobody assigned

There were two interns in the small office before I joined and three afterward. Officially, our job was to give the founder inspiration. In practice, the founder was the most important PM in the company.

Because our schedules were flexible and our interests differed, three roles emerged without being designed in advance:

  1. The Scout: follows frontier AI developments and translates raw information into inputs the team can understand.
  2. The Systematizer: turns scattered inputs into reusable SOPs, methods, and workflows.
  3. The Bridge: carries frontier exploration back to operating teams and brings real user and community feedback into exploration.

Together they formed a natural chain:

Frontier → systems → frontline

One office holds the present; the other runs toward the future

I do not think every company needs two literal offices. But AI startups, and organizations moving through an AI transformation, need some protected space for work that is important without yet being urgent.

This does not necessarily require extra headcount or elaborate organizational design. What matters is preserving a different time horizon and maintaining a continuous exchange between the people who scout, the people who systematize, and the people closest to operational reality.

One flywheel holds the present. The other runs toward the future. Neither can stop.

返回 Writing