From Sanity to Madness

从理智到疯狂

📝 Originally published (in Japanese) at forge.workstyle.tech. In a series of satirical videos, I had an AI production team (a planning meeting of 12 roles) write the scripts for the third episode. 📝 本文最初发布于 forge.workstyle.tech（日语）。在一系列讽刺视频中，我让一个 AI 制作团队（由 12 个角色组成的策划会议）编写了第三集的剧本。

After reviewing the finished video, the client sent back the following feedback: Mark Zuckerberg doesn’t just test things when he’s unsure; he tests everything. We should emphasize “trying everything” in the copy, right? Otherwise, the content feels like, “Wait, you’re testing that?” Where in the video is the part about Bill Gates’ “obsessive close reading”? The humor comes from the idea that he over-reads too much, so the content and copy should reflect that absurdity, like, “Wait, you read that too and are concerned about it?” Turn inability to make up your mind into madness. 在审阅完视频成品后，客户发回了以下反馈：马克·扎克伯格不仅在不确定时进行测试，他测试一切。我们应该在文案中强调“尝试一切”，对吧？否则内容会让人觉得：“等等，你连那个也要测试？”视频里关于比尔·盖茨“强迫症式精读”的部分在哪里？幽默感源于他读得太多这一设定，所以内容和文案应该反映出那种荒谬感，比如：“等等，连那个你也读了，而且还在意？”要把这种无法下决定的状态转化为疯狂。

Where Did the AI Script Actually Land?

AI 剧本最终落脚点在哪里？

The evolution of the Zuckerberg version went like this: 扎克伯格版本的演变过程如下：

Initial Draft (Me) 初稿（我） Sam: Which one do you prefer? Sam：你更喜欢哪一个？ Zuck: Both. Zuck：两个都要。 Sam: I need to make a decision. Sam：我需要做一个决定。 Zuck: Already shipped it. Zuck：已经发布了。 Sam: How about B? Sam：B 方案怎么样？ Zuck: Marginal difference. We already have a million users. Zuck：差别微乎其微。我们已经有一百万用户了。

Final Script by the Production Team 制作团队最终剧本 Sam: Which one do you prefer? Sam：你更喜欢哪一个？ Zuck: Let’s try both. Zuck：两个都试试。 Sam: We already tried. Sam：我们已经试过了。 Zuck: Next up: deciding. Zuck：接下来：做决定。 Sam: Deciding? Sam：做决定？ Zuck: Let’s try that too. Zuck：那个也试试。

The team cut the “1 million users” line because it implied unauthorized large-scale user experimentation. What remained was a coherent story contained entirely within the boardroom: “Let’s test the decision-making process too.” It was enough for a smart person to say something slightly different; there was no insanity involved. 团队删掉了“一百万用户”那句台词，因为它暗示了未经授权的大规模用户实验。剩下的就是一个完全局限在会议室里的连贯故事：“我们也测试一下决策过程吧。”对于一个聪明人来说，说出一些稍微不同的东西就足够了；这里面并没有疯狂的成分。

Skipping a Step

跳过逻辑步骤

Here is the idea the client pitched: Touching a boiling kettle while saying, “Huh? It’s hot? Let’s test it.” But because that’s completely normal for him, his expression doesn’t change. Then, they added this: What if at the end, he gets burned and his hand swells up as big as his face? 这是客户提出的想法：摸着烧开的水壶说：“嗯？烫吗？测试一下。”但因为这对他来说完全正常，所以他的表情没有任何变化。然后，他们补充道：如果最后他被烫伤，手肿得像脸一样大呢？

Here is how the dialogue turned out in the final draft: 最终定稿的对话如下： Sam: Which logo? Sam：选哪个 Logo？ Zack: Put out both. Zack：两个都发出去。 Sam: That kettle is hot, you know. Sam：你知道那水壶很烫吧。 Zack: Hot? Zack：烫？ Zack: Let’s test it. (Screen flashes to “Post-Test” for a split second) Zack：测试一下。（屏幕闪过“测试后”字样） Zack: It was hot. ← Visual of swollen hand appears Zack：确实烫。← 出现手部肿胀的画面 Sam: Told you so. Sam：我早跟你说了。 Zack: Hearing about it and testing it aren’t the same thing. Zack：听说和测试是两码事。

From comparing logos to a steaming kettle. The subject of verification leaps straight from work to his own body. He doesn’t bat an eye. And with that final line, you realize that in his own mind, it makes complete sense. 从比较 Logo 到烧开的水壶。验证的对象直接从工作跳跃到了他自己的身体。他连眼皮都不眨一下。而通过最后那句台词，你意识到在他自己的逻辑里，这一切完全说得通。

I pushed the other episodes further, too

我也进一步推动了其他剧集

The Gates episode. The first draft only had the document page numbers jumping from 128 → 3 → 1. The madness of the deep dive wasn’t showing through yet. The client’s instructions were this: Pack it in up to the fourth character on page 42. 盖茨篇。初稿中只有文档页码从 128 跳到 3 再到 1。深度钻研的疯狂感还没体现出来。客户的指示是：把内容压缩到第 42 页的第四个字符。

Sam: How’s the document looking? Sam：文档看起来怎么样？ Gates: The footnote on page 17. Gates：第 17 页的脚注。 Sam: What about the main text? Sam：正文呢？ Gates: The source is from 2019. Gates：来源是 2019 年的。 Sam: …What about the main text? Sam：……正文呢？ Gates: The fourth character on page 42 has a different font. Gates：第 42 页的第四个字符字体不一样。

I hadn’t even mentioned the main text once. I wasn’t even looking at pages; I was looking at individual characters. 我甚至一次都没提过正文。我看的根本不是页面；我是在看单个字符。

The Nadella episode. At first, it ended on a good note: “Let’s connect with our competitors.” Even with enemies, we want to collaborate. The punchline was: “Ah, a mosquito!” “A mosquito? Let’s connect.” 纳德拉篇。起初，结尾很正面：“让我们与竞争对手建立联系。”即使是敌人，我们也想合作。笑点是：“啊，一只蚊子！”“蚊子？让我们建立联系。”

Sam: Competitors are using the same features. Sam：竞争对手正在使用相同的功能。 Nadella: Let’s connect. Nadella：建立联系吧。 Sam: They’re coming to crush us. Sam：他们要来摧毁我们了。 Nadella: Even more so, let’s connect. Nadella：那更要建立联系。 Sam: …Ah, a mosquito. Sam：……啊，一只蚊子。 Nadella: A mosquito? Nadella：蚊子？ Nadella: Let’s connect. Nadella：建立联系吧。

The right side of the collaboration diagram changes from “Competitors” to “Mosquito.” 合作图表的右侧从“竞争对手”变成了“蚊子”。

Why the AI Didn’t Leap

为什么 AI 没有产生跳跃性思维

Reading the team’s meeting logs makes the reasons clear. Editorial prioritizes ensuring that actions are not mistaken for those of real people. Scriptwriting focuses on tightening the consistency of the dialogue. Direction adheres to the principles of “not adding explanations” and “avoiding clichéd one-liners.” All of these are valid points. However, when everyone insists on what is “correct,” the conversation converges on a safe middle ground that offends no one and contains no contradictions. Madness can only be born by shattering that consistency. 阅读团队的会议记录，原因就很清楚了。编辑工作优先确保行为不会被误认为是真实人物的行为。剧本写作专注于加强对话的一致性。导演遵循“不添加解释”和“避免陈词滥调”的原则。这些都是合理的观点。然而，当每个人都坚持什么是“正确”的时候，对话就会收敛到一个不得罪任何人、没有任何矛盾的安全中间地带。疯狂只能通过打破这种一致性来产生。

Division of Roles

角色分工

Here is the setup that worked best: 以下是效果最好的分工模式：

Humans: Make the intuitive leaps. A kettle, the fourth character, a mosquito. Leaps that logic alone could never produce. 人类： 进行直觉上的跳跃。水壶、第四个字符、蚊子。这些仅靠逻辑永远无法产生的跳跃。

The AI Team: Make those leaps work. Fit everything within the runtime, decide the on-screen presentation order, ensure factual consistency, and draw the line so viewers don’t mistake it for an action the person actually took. AI 团队： 让这些跳跃变得可行。将所有内容控制在时长内，决定屏幕上的呈现顺序，确保事实一致性，并划定界限，以免观众误以为这是当事人真实采取的行动。

For the kettle episode, the team proposed skipping ahead to the instant “after the test,” bypassing the touch scene entirely. When image generation models couldn’t properly draw a swollen hand, we had another AI generate it as an SVG. When it came to turning creative leaps into concrete realities, the AI did an exceptional job. 在水壶那一集中，团队建议直接跳到“测试后”的瞬间，完全绕过了触摸场景。当图像生成模型无法正确绘制肿胀的手时，我们让另一个 AI 将其生成为 SVG。在将创意跳跃转化为具体现实方面，AI 的表现非常出色。

Summary

总结

When everyone plays the voice of reason, the conversation simply converges toward consistency. Madness is born from skipping a logical step in focus (e.g., work → one’s own body, a page → individual letters, competitors → bugs). The human takes the leap, and the AI makes it work. It didn’t work the other way around. You can draw the safety line at: “Keep it confined to the individual’s own actions, and don’t drag others into it.” 当每个人都扮演理性的声音时，对话只会趋向于一致性。疯狂源于跳过逻辑焦点（例如：工作 → 自身身体，页面 → 单个字母，竞争对手 → 虫子）。人类负责跳跃，AI 负责让跳跃落地。反过来则行不通。你可以划定安全界限：“将其限制在个人行为范围内，不要牵扯他人。”