Never 王昆｜社区 × AI 产品

x.com/NeverToyJoy/st… 汇总成了一篇文章便于阅读

@rubickguo1 OKK，下次发文章

@NeverToyJoy 昆哥可以搞个长文出来，更适合分发和传播

一个中层，如何在传统互联网公司推动 AI 转型：7 条实战经验（0/7） 过去一年，Claude、Codex 等产品让 Agent 不再只是 Vibe Coding 工具，而逐渐成为了真正“会使用电脑的工作者”。 AI 辅助（或者说替代？）白领工作的时机，已经到了。 一方面，我忧心忡忡未来自己、行业和人类群体整体受到的AI冲击。 另一方面，我也在快速学习、使用Agent，成为最先适应的人之一。 因为还在公司打工，摆在所有非AI native互联网公司面前的挑战都一样： 到底该如何进行AI转型？ 这里说的AI转型并非指的自身的产品如何引入AI能力。 很多公司里的很多人正在强行拙劣地把AI能力嵌入到自己的产品里，提供不匹配他们薪资的💩一般的体验。 传统互联网产品如何被AI重构，是另一个值得单独展开的话题。以后我会分享自己经历过的失败、成功，以及一些不太主流的判断。 这个系列想讨论的，是另一种转型： 如何把 AI 尽可能嵌入公司的日常工作流，替代更多工作，并帮助真正掌握方法的个人和组织，实现 10X 产出。 我所在的公司并不以效率著称，有些组织机制甚至天然反效率。 想在这样的环境里推动 AI Native 改革，阻力重重。 尽管如此，到今天为止，我们依然取得了一些结果：产研速度更快了，交付质量也更高了。 比如我从 0 到 1 推动的 Toy平台（bilibili.com/toy/intro）一个提供给Vibe Coding创意者，便捷部署和分发自己网页的平台：起初只有一名研发，没有设计，产品工作则由我这个产品总监兼职承担；截止目前，已经有数百万用户使用过Toy。 接下来，我会用 7 篇短帖（真的非常短），分享自己在真实组织里推动 AI 转型的具体经验。 不讲“拥抱 AI”之类的正确废话，只讲踩过的坑，以及真正有效的方法。

还有一类平台类，比如Loopit、马卡农、灵光，希望依靠AI能力带来全新供给，做出来新的平台。不过平台的机会更属于新产品，还是旧平台升级不好说。

（7/7）一个中层，如何在传统互联网公司推动 AI 转型：7 条实战经验——减少系统里的交接节点 How a Middle Manager Can Drive AI Transformation Inside a Non-AI-Native Internet Company: 7 Lessons from the Front Lines — Reduce Handoffs Across the System AI 真正改变组织，相比让原来的每个岗位各自快 20%，更有价值是让更小的团队能够端到端完成交付。 AI doesn’t truly transform an organization by making every existing role 20% faster. It does so by enabling a smaller team to deliver end to end. 交接越少，对齐成本越低，人的判断也越不容易在流水线上损耗。 The fewer the handoffs, the lower the cost of alignment—and the less human judgment gets lost along the assembly line. 所以，我现在理解的 AI Native，并非“所有员工都装上了 AI”。 So my current definition of AI-native isn’t “every employee has an AI tool installed.” AI Native = AI 开始改变公司的成本结构、协作接口和组织拓扑。 It’s when AI begins to reshape the company’s cost structure, collaboration interfaces, and organizational topology. 前者是工具推广，后者才是真正的转型。 The former is tool adoption. The latter is real transformation.

（6/7）一个中层，如何在传统互联网公司推动 AI 转型：7 条实战经验——给 Agent 一个可衡量的 Goal How a Middle Manager Can Drive AI Transformation Inside a Non-AI-Native Internet Company: 7 Lessons from the Front Lines — Give the Agent a Measurable Goal 结果定义得越清楚，中间过程就越可以成为黑盒。 The more clearly the desired outcome is defined, the more the process itself can remain a black box. 在明确的边界内，人和机器怎么完成任务都可以，最后用统一的标准验收。 Within clear boundaries, humans and machines can complete the task however they choose. The final result should be evaluated against a shared standard. 如果结果不好，就回到模型、上下文、流程和人的能力上寻找原因。 If the result falls short, look for the cause in the model, the context, the workflow, or the human capabilities involved.

（5/7）一个中层，如何在传统互联网公司推动 AI 转型：7 条实战经验——把个人经验沉淀成组织资产 How a Middle Manager Can Drive AI Transformation Inside a Non-AI-Native Internet Company: 7 Lessons from the Front Lines — Turn Individual Know-How into Organizational Assets 好用的 SOP、上下文、检查表、Skill 和 Harness，都应该沉淀并留在团队里。 Useful SOPs, context, checklists, Skills, and Harnesses should be captured and retained within the team. 否则，每个人都要从零学一遍。所谓的提效只发生在个体身上，并没有转化为组织能力。 Otherwise, everyone has to learn the same lessons from scratch. The productivity gains remain individual instead of becoming an organizational capability.

（4/7）一个中层，如何在传统互联网公司推动 AI 转型：7 条实战经验——从 Prompt 升级到 Loop How a Middle Manager Can Drive AI Transformation Inside a Non-AI-Native Internet Company: 7 Lessons from the Front Lines — Move Beyond Prompts and Build Loops 专业任务很少能靠一问一答完成。更稳定的结构是：人定义目标和标准 → Agent 先做 → 人反馈 → Agent 修正 → 用 Evals 验收。 Professional tasks are rarely completed with a single prompt and response. A more reliable structure is: humans define the goals and standards → the agent takes the first pass → humans provide feedback → the agent revises → evals verify the result. 关键不是找到一句神奇的 Prompt，而是这个循环能不能持续收敛。 The key isn’t finding a magic prompt. It’s whether the loop can keep converging toward a better result.

（3/7）一个中层，如何在传统互联网公司推动 AI 转型：7 条实战经验——既要给出可复制的示范，也要展示能力上限 How a Middle Manager Can Drive AI Transformation Inside a Non-AI-Native Internet Company: 7 Lessons from the Front Lines — Make It Easy to Copy, and Show What’s Possible 身边同事展示一个今天就能复制的工作流，会让人觉得：“他能做到，我也能。” When a colleague demonstrates a workflow that others can copy today, people think: “If they can do it, so can I.” 外部高手则负责展示能力上限，让大家意识到：“原来还可以这样。”两种分享缺一不可。 External experts show what’s possible and make people realize: “I didn’t know AI could do that.” You need both kinds of examples.

（2/7）一个中层，如何在传统互联网公司推动 AI 转型：7 条实战经验——对结果挑刺，不考察表演 How a Middle Manager Can Drive AI Transformation Inside a Non-AI-Native Internet Company: 7 Lessons from the Front Lines — Judge the Results, Not the AI Theater 我更关心交付有没有变好，而不是一个人每天问了多少次 AI。 I care more about whether the work has improved than how many times someone uses AI each day. 看到一份“AI 味”很重的材料，我会继续追问：你自己的判断是哪部分？哪些结论是你愿意负责的？ When a deliverable has that unmistakable AI-generated feel, I keep asking: Which parts reflect your own judgment? Which conclusions are you willing to stand behind?

一个中层，如何在传统互联网公司推动 AI 转型：7 条实战经验（1/7）——宁可少给人，也要给最好的模型 How a Middle Manager Can Drive AI Transformation Inside a Non-AI-Native Internet Company: 7 Lessons from the Front Lines (1/7) — Give Fewer People Access, but Give Them the Best Models 如果核心任务只能使用明显落后一档的模型，团队很容易用糟糕体验证明“AI 不行”。 If people are forced to use models that are clearly a tier behind for core tasks, they will quickly take the poor experience as proof that “AI doesn’t work.” 预算有限可以限制人群和任务，不要把差模型平均发给所有人。 If the budget is limited, limit access to fewer people or use cases. Don’t give everyone an equally mediocre model.

EN. verson： How a Middle Manager Can Drive AI Transformation Inside a Non-AI-Native Internet Company: 7 Lessons from the Front Lines (0/7) Over the past year, products like Claude and Codex have pushed agents beyond vibe coding tools. They are gradually becoming real “workers who know how to use computers.” The time for AI to augment—or replace?—white-collar work has arrived. On one hand, I’m deeply worried about the impact AI will have on my own future, my industry, and society as a whole. On the other, I’m learning and using agents as quickly as I can, trying to become one of the first people to adapt. For those of us still working inside companies, every non-AI-native internet company faces the same challenge: How do we actually make the transition to AI? By AI transformation, I don’t mean adding AI capabilities to existing products. Far too many people at far too many companies are awkwardly forcing AI into their products—and delivering 💩 experiences that hardly match what they are paid. How existing internet products should be rebuilt around AI is a separate topic worth exploring. I’ll eventually share my failures, successes, and a few unconventional opinions about it. This series is about a different kind of transformation: How do we embed AI as deeply as possible into everyday company workflows, replace more tasks, and help the individuals and organizations that truly understand it achieve 10x output? The company I work for isn’t exactly known for efficiency. Some of its organizational mechanisms are almost anti-efficiency by design. Trying to drive an AI-native transformation in an environment like this comes with enormous resistance. Even so, we’ve achieved tangible results. Product and engineering cycles are faster, and the quality of what we deliver has improved. One example is TOY（bilibili.com/toy/intro）, a platform that helps vibe coders easily deploy and distribute the web experiences they create. I drove TOY from zero to one. At the beginning, we had just one engineer, no designer, and I handled the product work part-time alongside my role as product director. To date, millions of users have used TOY. In the next seven posts—and I promise they’ll be very short—I’ll share practical lessons from trying to drive AI transformation inside a real organization. No generic advice about “embracing AI.” Just the mistakes we made—and the methods that actually worked.