slop的定义是『把成本转嫁给接收方』：The AI Daily Brief 操作指南——如何去slop化你的AI

先说结论

这期 The AI Daily Brief（11 月 3 日，Operators Cut 实操系列）讨论的问题不是"AI 内容好不好"，而是给 slop（AI 口水内容）下了一个可操作的定义：嘉宾 Nufar（转写如此，本系列旧文亦作 Noufar/Nufar Gasbar）的定义是——本该由人完成的思考被部分或全部跳过，成本转移到了接收这份工作的人身上。这个定义直接给出标准：接收方付出最少、收获最大就是非 slop。围绕它，两人给出机器侧与人侧的成因分析、五问自检、五组工作习惯，以及主持人 Nathaniel（节目内自称 NLW）自己长内容流水线的完整人机分工实案。共同判断是：“会用 AI 量产"这个技能已经民主化，下一阶段的差异化是"用得好”——即有品味、有意图。以下均为节目内容转述，数字与判断未经独立核实。

slop 的规模：从年度词到"slop 工厂"

主持人开场把现象拉满：全员邮件"听起来不像任何人写的"、太长太顺、句式节奏可以预测（短句、三段列表、再一句"这其实关乎什么"）；金色光线过头的食品广告；一句就能听出没人写的电台广告；“now with AI"的营销活动；信息流里每条帖子用同一个钩子开头、同一个问题结尾。节目引用：Merriam-Webster 2025 年度词就是 slop（低质量、通常由 AI 批量生产的数字内容；更早的含义是喂猪的泔水）。而 2026 年变得更糟的原因很具体：公司把 agentic 工具铺给了所有人，过去一次一稿的工作，现在由 agent 批量产出文档、幻灯和代码。办公室外的尺度感来自节目引用的 Deezer 数据：每天上传的音乐过半是完全 AI 生成的，但 AI 生成内容只占下载收听量的约 1–3%——供给与需求完全倒挂。

让 Nufar 决定做这期的导火索是一场培训：他非常尊敬的一位产品负责人强推团队全员用 AI，采用率极高，结果团队变成了对方自己口中的 “slop 工厂”——没人读就发出去的长产出、每家公司听起来都一样的文案，最糟的是几乎没有原创思考。对方的原话是"这让所有人都变浅了，人和产物都浅”。数字侧的佐证：节目引用 Stanford 与 BetterUp 的研究称约 40% 的美国桌面工作者上个月收到过 slop，接收端为此耗费近 2 小时；另一项六千人的调查（转写中机构名存疑）称约 70% 的人发出工作成果前根本没有读过它。

机器为什么产 slop，人为什么容忍

机器侧的解释很干脆：工具在做它被造出来做的事。语言模型被训练去预测最可能的答案，通用问题得到的大致就是"互联网对这个话题的平均意见"；记忆和上下文功能帮不上忙——如果你今天没有把自己的思考喂进去，它无从知道你怎么想。后训练还在加剧：人类评分员奖励"熟悉感"，Stanford 研究者称之为 typicality bias（典型性偏好），被折叠进模型后输出范围收窄，默认答案又长又列表化——因为"长度看起来像努力"，而人恰恰这样奖励它。指望换模型解决？节目引用去年 NeurIPS 获奖论文的发现：不同公司的模型在收敛到相同的答案——换工具买不来原创性。图像模型同理，每家有自己的 house style（食物图的金光就是典型），不主动描述就不会离开默认风格。Nufar 的现场实验：把 ChatGPT（Sol）、Claude（Opus）、Gemini（转写作 Allegan，疑为 Argon）无痕模式同时打开，给同一道"写一条 AI 如何改变营销的 LinkedIn 帖子"——连他都惊讶于结构与内容的相似度；而把自己的观点、经验、喜欢的结构、永不使用的词重新喂回去后，稿子立刻上一个台阶（虽然仍不能直接发）。

人侧的容忍度来自三个机制：心理学家 80 年代起就把人脑称为认知吝啬鬼（cognitive miser）——能省思考就省；飞行员称之为自动化自满（automation complacency）——机器通常对，人就停止检查；研究还发现越信任 AI 的人，对输出应用的批判性思考越少。slop 的三种成因几乎都不是恶意：①省事，看起来还行，有压力就转发了；②不懂行——不懂装懂的 vibe coder 提出天真的请求、得到天真的结果，资深工程师提同样的需求却能拿到更好的活，因为知道要什么；③最麻烦的：外包思考——还没决定自己怎么想，就先问工具怎么想。Nufar 把阻止第三条称为他近几周的"运动"：找回思考这门失传的手艺。

五问自检与两类不靠谱的技术方案

判断一份工作（自己的或别人的）是不是 slop，五个问题：① 不重读的情况下，发出者能为每一句话辩护吗？② 发出前读完全文了吗——四十页报告读了两页觉得"还行"就把阅读外包给对方，是典型路径；③ 有没有只有这个人能贡献的东西——观点、决策或真实数据？谁都能署名的东西不会是高质量工作；④ 有没有更有效的传达方式？整体上帮读者省时间还是花时间？⑤ 最直觉的一问：愿不愿意把自己的名字全大写签在上面？“有时说不清为什么不舒服，但我们知道”——Nufar 说这是他最喜欢的一问。

技术方案他评价不高：AI 检测器在 LinkedIn 上"什么都能点亮"，量的是结构、冒号密度、列表密度、句长一致性，测不出质量与原创性——最多拿来对照自己的稿子是否落进平均数；**humanizer（拟人化工具）**被他总结为"slop 尽量不像 slop"——拟人化后的文本仍有约四分之三会被检测器标记，而读者真正的痛点从来是缺独特观点，不是具体文风。

Nufar 的五组习惯

①开工前（最重要的一步，也是人们跳过时下游代价最大的一步）：先有观点再开工具——散步、读材料，直到形成自己的看法，多乱都行；然后把它口述倒给 AI（brain dump，不管结构，告诉它你听到什么、担心什么、确定不要什么），本期节目本身就是这样起步的——他先对 AI 乱讲了十五分钟。同时声明置信度并要求顶撞：“我认为 60% 是 onboarding 的问题，同意我之前请使劲反驳”——模型天生顺从，带着强观点进来它会帮你论证每一条；即使很确定，也要反方观点来磨利自己。还要先想清楚受众、想给对方的体验、长度/格式/语气标准。②执行中：go wide——第一个答案就是最可能的答案、也就是最平均的答案；一次要六个不同方向（方法、角度、结构、视觉各异，可以跨域借力：游戏设计师会怎么做？新闻编辑部会怎么做？），让模型各自论证再挑；并且预设拦截点（大纲→第一节→初稿→成品），别让 agent"端到端跑完"——成品一摆出来沉没成本就登场，“就这样发吧"是最贵的决定。③评审：开一个 persona 董事会——让 AI 分别以怀疑的目标读者、领域专家、最忙的接收者、魔鬼代言人（论证"这根本不该发出去”）来审；尊重阅读预算——你自己都不会读的东西别发，否则对方会丢给他的 AI 去总结，“变成两个模型在对话、两个人假装在沟通”；把一页蒸馏成五句会逼你决定自己到底怎么想；按受众刻意选择格式（幻灯、播客、网页、图、视频皆可）。④沉淀：教 AI 你的品味——你不需要有语言学家或艺术总监的词汇，把最好的作品、欣赏的他人作品、讨厌的东西喂给强模型，让它替你说出偏好；视觉就做 mood board。把它沉淀成 skill，作为"要像我"的起点——代价是同质化（他自嘲受众已经听腻了他的口头禅，转写存疑），所以要定期主动跳出模具。⑤文化：收到 slop 就点名，尤其是管理者——“这不是对我时间的好用法，用五句话把你的观点发给我”；对客户向内容绝不放行；但不必藏 AI——Figma 已公开其 AI 写作规则：用 AI 写不丢人，但在顶部标注出来，让读者带着正确的预期进入（比如一份四十页报告的目的可能是"一起翻找有意思的部分"而非成品交付）。

（说明：本期中段有一整块赞助商口播——KPMG（广告引用其与 UT Austin 的"AI amplifiers"研究：500 多名早期职业者中表现最好者的共性是引导、评估、修正 AI 输出）、Robots and Pencils、CData、Blitzy。以上均按广告处理，不作独立事实背书。）

Nathaniel 的实案：How We Got Here 的人机分工

主持人用自己新开的深度史播客（每期 45–60 分钟；第二集刚上线，讲"东西变便宜、生活变贵"的"错误的丰裕"；首集回溯美国对科技幻灭的十五年并关联当下数据中心之争）展示了这套原则的满配版。他不可能是每个题目的专家（未来要做 GLP-1 药物史时更只是"翻译者"），所以分工底线画得极清：研究全部交给 AI——多个模型各出综述，他要的不是谁更好，而是各自强调的差异；但观点绝不外包。他常让 AI 同时 steelman 一个问题的两极立场并"以最深信念"激烈互辩——结论是"它终究不是一个有视角的人，只是它能检索到的所有论点的总和"，所以观点必须自己带着来。叙事架构也不外包：他几乎从不认同 AI 想要的论证结构，多年长内容经验养成的直觉不交给模型。操作上他不写稿、即兴口述，AI 把关键事实与数字做成简报（与 AI Daily Brief 的人机协作档案同款）。每期还有一篇 6000 字配套长文：AI 直写这个体量必是 slop，他的解法是两步——不喂大纲而是喂自己的口述转写稿（保留措辞与语气，允许 AI 扩展背景），然后逃不掉的五到十小时逐段重写（他甚至为这在网站里专门做了个内容管理系统）：“有些写法英语老师会判错，但听起来像我”。他给出的三个要点与 Nufar 完全咬合：对"哪部分是我的、哪部分是 AI 的"有明确观点；不外包观点；接受没有完美效率——把去 slop 的时间当作做生意的成本。Nufar 补充适用边界：按 stakes 定投入——内部一行邮件随便 AI（“简洁高效地把事办了就不算 slop”），代表品牌、面向客户、高影响的工作才需要全套仪式，最低限度是"想一下流程"。

往哪看：手工溢价与 top 1% 的重新定义

Nufar 的预测：模型越好，通用答案越诱人；agentic 热潮退去后，手工感（handmade）会成为溢价品，人们会为"人类指纹"付更多——哪怕乱、哪怕语法不完美，但像你、让人有直觉反应。纪律性的高质量工作就是所有人的 job security；把 slop 挡在工作流之外（尤其对外）的公司会以别人做不到的方式脱颖而出、吸引客户。主持人收尾的"自利视角"：听众都想当 AI 时代的 top 1%，但 top 1% 的定义已经换过一轮——早期是"知道用 Gamma 做更漂亮的幻灯"，如今这些技能已完全民主化；差异化从量产转向"用得好"——有品味、有意图。第一阶段"到场就有分"，第二阶段"产品本身要过硬"——难得多，但在无尽的内容海洋里，好东西反而更突出，人会不由自主地变擅长过滤噪音。他自己的注脚：这期本想用"load bearing"这个词，“但 AI 用得太多，我用不下去了”。

机制与启发（可迁移的链条）

• 评分偏好（typicality bias）+ 模型间收敛 → 通用提示词的输出同质 → “会用 AI"本身不再稀缺 → 差异化移向独有观点与品味。观察指标：同题多模型输出的相似度、AI 检测器误报率。启发：提示词里最值钱的部分是你自己的判断，不是咒语。
• slop=成本转嫁 → 接收方时间被持续燃烧 → 组织信任损耗，管理者开始"把门” → 制度化点名与 AI 标注（Figma 规则）。观察指标：所在公司是否出台 AI 写作/标注规范。启发：把读者的时间当自己的预算来花，这是反 slop 的第一性原则。
• 品味可编码（教 AI 你的偏好→skill）→ 个人产出提速 → 风险是 sameness 被受众记住 → 需主动定期换模具。观察指标：你的风格是否被一秒识别、且隔段时间有新意。启发：skill 是起跑器，不是风格的棺材。
• 手工溢价若成立 → 类似工业品与手工艺品的定价分叉 → “有人真的想过"成为稀缺品。观察指标：是否出现"人类创作"认证/标签市场。启发：AI 时代写作的护城河是署名权背后真的有思考。

ASR 校正表

转写原文（修正版）

附录为 ASR 转写全文（22 段，0s–2536s），仅应用校正表中已确认的替换，无删节；每段先英文原文、后中文翻译；广告口播按原文收录并标注。

[0s - 120s]

The AI slop is everywhere, man. Look, when the word slop started being used, it was perhaps reasonable to write it off as an angry response from folks who just didn’t like AI in general. However, there is no doubt now that as AI writing, and image generation, and video generation, and code generation have gotten ubiquitous, we are living in a mass of content. Inside work, outside of work, it is just content everywhere, all the time, and it actually comes with a cost. If a first phase of AI for work was learning how to use these tools to produce just volumes and volumes of new content, a next phase is how to distinguish yourself by using them well. And so, friends, let’s talk about deslopping your AI. The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI. All right, friends. Quick announcements before we dive in. First of all, thank you to today’s sponsors, KPMG, Robots and Pencils, CData, and Blitzy. To get an ad-free version of the show, go to patreon.com/AIDailyBrief, or you can subscribe on Apple Podcasts. And if you want to learn more about sponsoring the show, send us a note at sponsors@AIDailyBrief.ai. Now, two more quick things. First, it is Monday, which means the new episode, just the second so far of How We Got Here, is out. This one is called How We Made the Wrong Things Cheap or the Wrong Abundance, and it’s about how stuff got cheap while life got expensive, and all of the attendant consequences of that change. That is now available on all major podcast apps. Again, the show is How We Got Here. Second, I am on the road today, so we are doing an operators cut, also one that is very timely for just about everyone. But I will be back with a normal episode for tomorrow. All right, Nufar, welcome back with another operators cut. How are you? I’m good. I’m excited about this one. I’m very passionate about this topic. Yeah, this is even even within the framework of operators cuts, which are sort of the more practical oriented AI Daily Brief episodes. There’s still a band of how conceptual.

【译】AI 口水内容（slop）真是无处不在，朋友们。听着，当 slop 这个词刚被使用的时候，把它当成一群本来就不喜欢 AI 的人的愤怒反应、一笑了之，或许还说得过去。但现在已经毫无疑问：随着 AI 写作、图像生成、视频生成、代码生成变得无处不在，我们活在内容堆里。工作之内、工作之外，到处都是内容、时刻都是内容，而它是有代价的。如果说 AI 用于工作的第一阶段，是学会用这些工具量产一批又一批新内容，那么下一个阶段，就是靠「用得好」来让自己脱颖而出。所以朋友们，我们来聊聊「给你的 AI 去 slop」。The AI Daily Brief 是一档关于 AI 领域最重要的新闻与讨论的每日播客和视频。好，朋友们，正式开始前几个快速通告。首先感谢今天的赞助商：KPMG、Robots and Pencils、CData 和 Blitzy。想要无广告版，去 patreon.com/AIDailyBrief，或在 Apple Podcasts 上订阅；想了解赞助事宜，给我们发邮件 sponsors@AIDailyBrief.ai。再两件小事。第一，今天是周一，意味着新节目《How We Got Here》目前总共第二集上线了，这一集叫《我们如何把错误的东西变便宜了，或者说错误的丰裕》，讲东西怎么变便宜了、生活却变贵了，以及这一变化带来的一切后果，现已在各大播客平台上线。第二，我今天在路上，所以这期是 Operators Cut（实操版），而且对几乎所有人来说都非常及时。明天恢复正常节目。好，Nufar，欢迎回到又一期 Operators Cut。你怎么样？——我很好，这期让我很兴奋，我对这个话题非常有热情。——是的，即便在 Operators Cut 这种偏实操取向的 AI Daily Brief 节目里，也存在一个从概念到应用的频谱，而这一期大概应用到了极致：基本上就是一份带结构的事做/不做的清单。所以我很期待。希望大家能带走一些非常具体的技巧、想法和思路；其中很多你们可能凭直觉早就在做了——我猜这档播客的听众比大多数人想这些想得更多——但同时也希望它能为作为听众的你和你身边的人提供资源。总之，等不及要开始了。

[120s - 240s]

Versus applied, they are, and this one is about as applied as it gets. It’s pretty much a list of things to do or to not do that have some structure around them. So I’m excited to dig in. Hopefully, I think folks will come away with some very specific techniques, ideas, thoughts, etc. A lot of which they will probably have already been doing intuitively, because my guess is that this is an audience that thinks more about this stuff than many others, but can also hopefully provide resources for both you as the listener and other folks around you. So, anyways, I’m excited to dive in. Right. So, before we talk about specifics, let’s bring us all to the same page, and I’m sure that you’ve all seen it this week and maybe even today. It might have been the all hands email that sounded like nobody, or it was too long and too smooth, slightly sterile, with a rhythm that you can hear in your own head by now, like short sentence, then list of three, then a line telling you what something is really about, and you’ve seen that in your inbox in multiple places. Maybe it was the especially believable image, or maybe it was a radio ad that you knew after one sentence that nobody really wrote, or maybe it could have been like the food ad with the lighting that is a little too golden and 20 more just like it, or it was a campaign for literally any company whose whole idea was now with AI. And one of the worst places is the feed where every post opens with the same hook and ends with the same question, and it’s everywhere, and it looks the same, and after a while you stop reading any of it. And in all honesty, I think we all contribute, and we’re all part of the thing. So, Merriam-Webster word of the year for 2025 was “slop.” That was almost a year ago, and they defined it as digital content of low quality that is produced usually in quantity by means of artificial intelligence. And before that, it meant scraps you feed the pigs. So this year, I think it got lots worse because 2026 is.

【译】在讲具体做法之前，先把大家拉到同一页——你们这周、甚至今天肯定都见过它了。也许是那封「听起来不像任何人写的」全员邮件：太长、太顺滑、有一点无菌感，那个节奏你脑子里都能预演——短句，然后三项的列表，然后一句告诉你这件事「其实真正关乎什么」，这样的邮件你在收件箱多个地方见过。也许是那张「特别可信」的图片；也许是一句就能听出没人真写过的电台广告；也许是光线金得有点过头的食品广告、外加二十张一模一样的；也许是随便哪家公司「现在加上 AI 了」的营销战役。最严重的重灾区之一是信息流：每条帖子用同一个钩子开头、同一个问题结尾，到处都是，长得一模一样，看了一阵之后你谁的内容都不读了。老实说，我认为我们人人有份，我们都是这个问题的一部分。「Slop」是 Merriam-Webster 的 2025 年度词——那差不多是一年前的事了。他们的定义是：通常借助人工智能批量生产的低质量数字内容。在那之前，它的意思是喂猪的泔水。而今年我认为情况糟糕了很多，因为 2026 年是公司把 agentic 工具铺给所有人的一年：过去一次只来一稿的工作，现在是从 agent 那里成批涌来的文档、幻灯片和代码。要看办公室之外的规模：Deezer 最近的统计称，每天上传到该平台的音乐超过一半是完全 AI 生成的；但看下载，AI 生成内容只占人们收听量的约 1% 到 3%。这是背景和定义。但真正让我最近对这个话题燃起热情的，是一场与管理者们的培训，其中有一位我非常尊敬的产品负责人：他们非常懂 AI、思维非常锋利，也在非常用力地推动员工使用 AI。所有人都采用了，但团队变成了一个 slop 工厂——这是他们自己的定义：没人读就发出去的超长产出、听起来像同一个模型的文案、任何公司都能产出的同款结果，而最糟的是几乎没有原创思考。他们告诉我：这让所有人都变浅了——人和产物都浅。

[240s - 360s]

When companies rolled out agentic tools to everyone, and the work that used to arrive one draft at a time comes from agents producing documents, decks, and code in in volume. And if you want even a scale outside of the office, a recent stats from Deezer reported that more than half of the music uploaded to the platform every day is fully AI generated. But if you look at the downloads, the AI generated downloads make up about one to three percent of what people listen to. So that’s the background or the definition. But the thing that got me to be very passionate about this topic recently is that I had a training session with the managers and one product leader that I really respect a lot, and they’re very AI savvy and very sharp thinkers. And they were kind of pushing their employees very hard to use AI. Everybody adopted that, but the team turned into a slop factory. That’s their definition: long outputs that nobody read before sending the copy that sounds like the same model and the same result from any company. And worst of all, very little original thinking. What they told me is it made everybody shallow, the people and the artifact. The way I define that, given these conversations and many others that I’m having with leaders, is the work where the person who should have done the thinking basically skipped it, either partially or altogether, and the cost moved to whoever received the work. And of course, humans produce slop way before we had AI. AI just made it cheap and in volume. So this definition also gives you the standard because the most effective output for the least effort on the other side will be non-slop work. And compacting it to the format, the length, and the quality that your audience needs is on you, the responsible AI users. A few numbers that agree also with that: Stanford and BetterUp they did a work slop study, and they said that about forty percent of U.S. desk workers had received in the past month something to be considered slop.

【译】结合这些对话、以及我与其他领导者的很多对话，我给 slop 的定义是：本该由某个人完成的思考，被这个人部分或完全跳过了，而成本转移到了接收这份工作的人身上。当然，早在 AI 出现之前人类就在生产 slop，AI 只是让它变得便宜、且成批。这个定义同时也给了你标准：对接收方而言，用最少的付出得到最高的效用，就是非 slop 的工作。把内容压缩到你的受众需要的格式、长度和质量，是每一个负责任的 AI 使用者自己的责任。几个与之一致的数字：Stanford 和 BetterUp 做了一项职场 slop 研究，称约 40% 的美国桌面工作者在过去一个月里收到过可被视为 slop 的东西，接收端为此耗费接近两个小时。还有 Glencelvie（转写如此，出处待核）——我记得你在某期节目里引用过——今年对六千名工作者的调查称，几乎 70% 的人把工作成果发出去之前从头到尾读都没读过。我猜你们中有人现在在说：「不，这些数字还偏保守了，在我们公司严重得多。」那么除了抱怨之外，我想在这一期真正帮到你，分三部分：第一，解释我们为什么走到了这里——机器和人各自产生 slop 的原因；第二，怎么判断一份工作（你自己的或别人的）是不是 slop；第三也是最重要的部分：让工作质量对得起任务分量的习惯有哪些。我要把一件事说清楚：这不是一期反 AI 的节目。我用 AI 帮我做我做的每一件事，当然包括这一期。我只是想确保你以一种真正产出原创作品的方式去做。

[360s - 480s]

And it cost close to two hours on the receiving side. In the Glencelvie, I think you quoted it in one of the episodes. Six thousand workers this year said that almost seventy percent shipped work without ever reading it to someone else. And I suspect some of you are now saying something like, “No, these numbers are even on the lower side. In our company, it’s much more.” So beyond kind of complaining and so on, I want to help you. And the way I want to help this episode is in three parts. First, explain why we got here, why slop happens in the machine and in in us. Also, how do you know that something is slop, either in your own work or someone else’s work? And finally, and the most important part, what are the habits that keep the work you all scale to the stake of the task? And I want to make one thing very clear: this is not an anti-AI episode. I use AI to help me with everything that I do, including of course this session. I just want to make sure that you do it in a way that actually produces original work. So I want to start with why the machine produces slop, and the answer is that the tools are doing what they were built to do because a language model is trained to predict the most likely answer. You give it a generic question, and you get roughly the average of what the internet thinks about this topic. And if you add some memory and context to the tool, they help, but they still can’t tell the tool what you actually think about it if you haven’t shared your thinking today. And then there’s of course another thing that contributes to that, and that is the tuning after the training. The way it works is that people rate the answers, and raters reward what feels familiar. Stanford researchers named this the typicality bias, and that is being folded in many cases into the model, and the range narrows of the answers, and the default comes out long and very listy, like we all know that AI create because length looks like effort, and in many cases that’s how people reward it. And if you think that swapping models can solve it, then the best paper last year from NeurIPS they found that different.

【译】先说机器为什么产 slop：答案是工具正在做它被设计出来做的事——语言模型被训练去预测最可能的答案。你给它一个通用的问题，你得到的大致就是互联网对这个话题的平均意见。给工具加上记忆和上下文会有帮助，但它们仍然无法替你告诉工具「你」到底怎么想——如果你今天没有把你的思考分享出来的话。训练之后的调优也在推波助澜：流程是人类给答案打分，而打分者奖励「感觉熟悉」的东西。Stanford 研究者把它命名为「典型性偏好」（typicality bias），它在很多情况下被折叠进模型，答案的范围随之收窄，默认输出变得又长又满篇列表——我们都知道 AI 会这样，因为「长度看起来像努力」，而很多时候人就是这么奖励它的。如果你觉得换模型能解决——去年 NeurIPS 的最佳论文发现，不同公司的不同模型会收敛到相同的答案。所以换工具买不来原创性。顺便说，图像模型也不更好：每个模型都有自己的「家族风格」，你不描述想要别的，它就退回那个风格，食物图上的金光就是这种家族风格之一。

[480s - 600s]

Models from different companies converging the same answers. So switching tools won’t buy you originality. And by the way, with image models, it’s not better. It’s kind of the same story because every model has their own house style, and it falls back to that if you don’t describe something else that you want to do. So the golden light on the food is one of these things that is like a house style. And in order to prove it, and I strongly recommend to run similar exercise yourself, I did like an experiment. And what I did, I opened ChatGPT, Sol, Claude, Opus, and Gemini, Allegan, the latest versions in an incognito mode. So that means that memory and history were off, and I gave them all a very generic and simple prompt: write a LinkedIn post about how AI is changing marketing. And I have to say that even I was surprised by how similar the results. Not only did they have very similar structure, as you can see here, and I think the structure by now will be very identifiable by all of you, but even more so, the content was very similar. And then when I reran the same task now with my point of view, what I believe about AI in marketing, what my experience, and also what I like in terms of structure and words that I never use, all of a sudden I got a much better draft. Could I have posted that on LinkedIn as is? Not at all. But at least I had a much better starting point for editing. So that’s the machine side. The question is why we as humans tolerate the slop, and the answer is that psychologists have called our brain a cognitive miser since the I guess early 80s because we are prone to save thinking and energy wherever we can. Even if you talk to pilots, they call the downside automation complacency and something that when the machine is usually right, you just stop checking. And even if you go into knowledge workers, research found that the more people trusted the AI, the less critical thinking they applied to the output, and the models make it very easy to feel very confident.

【译】为了证明这一点——我也强烈建议你们自己跑一遍类似练习——我做了个小实验：把 ChatGPT（Sol）、Claude（Opus）和 Gemini（Allegan，转写如此）的最新版本全部用无痕模式打开，也就是说记忆和历史都关掉，然后给它们同一个非常通用、非常简单的提示词：写一条关于「AI 如何改变营销」的 LinkedIn 帖子。我得说连我都对结果的相似程度感到惊讶：不仅结构非常相似——我想这个结构你们现在都能一眼认出来——内容更是如此。然后我用「我的观点」重跑同一任务：我对 AI 在营销中的看法、我的经验，还有我喜欢的结构、以及我永远不用的词——突然之间我得到了一份好得多的草稿。能直接发 LinkedIn 吗？完全不能。但至少我有了一个好得多的编辑起点。这是机器那一侧。问题是：我们人类为什么容忍 slop？答案是心理学家从大概八十年代初就把我们的大脑称为「认知吝啬鬼」——因为我们在任何能省的地方都倾向于省下思考和能量。问问飞行员，他们把这种副作用叫「自动化自满」：当机器通常是对的，你就不再检查了。再看知识工作者：研究发现人们对 AI 越信任，对输出应用的批判性思考就越少——而模型让你非常容易感到非常自信。

[600s - 720s]

So if you think about how slop gets made, there are basically three ways, and almost none of them are bad intent. It’s because it’s easy, like it looks fine, and you’re under pressure, so you just forward it, and not knowing the craft. For example, if you compare someone who vibe code without knowing much what they’re doing, the request is going to be very naive, and they will get in many cases naive results. While a more senior engineer asking for the same thing, they get better work because they know what to ask. And lastly, and that’s the most troubling part, that is because we outsource the thinking. We’re asking the tool what it thinks before we decide what we think, just because it’s easy and the tools are so opinionated. So literally preventing this last one is my crusade over the last few weeks. I call it recovering the lost art of thinking. And before I want to give you some ideas on how to prevent slop, I want to also give you like a set of questions to detect slop in yours or in other people’ work. All right. So to tell slop from good AI-assisted work, these are some questions that you might ask yourself or when you review someone else’s work. So the first thing that I want to ask is, could the sender defend each and every sentence without rereading that? In many cases, in all honesty, we cannot. The second thing is, did the sender read all of it before it went out? The forty pages report seems okay after you read one or two pages, and then you just get tired, so you offload the reading to the other side. The other question that I want you to ask is, there anything only this person could have contributed, like a point of view, a decision, or real data? If anybody could put their name to this output, that means that it’s not going to be relevant or high quality work. Is there a more effective way to convey this information, or does this one still do the job the best it can? Namely, does it overall save the reader time or cost them time instead of helping them? And lastly, and that’s the most

【译】所以 slop 是怎么被制造出来的？基本三条路径，而且几乎都不是恶意。一是图省事：它看起来还行，你有压力，于是直接转发；二是不懂行：比如一个不太清楚自己在干什么的 vibe coder，请求会很天真，很多时候就得到天真的结果；而一个资深工程师提同样的需求会拿到更好的工作成果，因为他知道该要什么；第三条也是最让人担心的一条：我们外包了思考。我们在决定自己怎么想之前，先问工具怎么想——只因为它容易，而且这些工具个个「观点鲜明」。阻止最后这一条，就是我最近几周的「运动」：我称之为找回思考这门失传的手艺。在给出一些防止 slop 的想法之前，我想先给你们一组检测 slop 的问题——检测你自己的工作，或者别人的工作。要把 slop 和好的 AI 辅助工作区分开，这些是你可能问自己、或在审阅别人工作时问的问题。第一：不看稿重读，发出者能为每一句话辩护吗？老实说，很多时候我们自己就做不到。第二个问题：发出者把全文读完了吗？四十页的报告读了一两页觉得「看起来还行」，然后你就累了，于是把阅读外包给另一边。第三个问题：这里面有没有只有这个人才能贡献的东西——一个观点、一个决策、或真实的数据？如果任何人都能把自己的名字署在这份产出上，那它不会是相关的、高质量的工作。第四个问题：有没有更有效的传达信息的方式？这一份是不是仍然把这件事做到最好了？也就是说，它整体上是帮读者省时间，还是让他们花时间？最后一个也是最重要的、最直觉的一个：你愿不愿意把自己的名字用全大写签在这份产出上？如果你觉得不舒服，那就说明还有活要干。有时候我们说不清为什么不愿意署名，但我们就是知道——所以这是我最喜欢的一问。

[720s - 840s]

Important thing and the most intuitive one: Would you feel comfortable putting your name in all capital letters to this output? If you don’t feel comfortable putting your name, that means that there is more work to be done. And sometimes we cannot explain why we don’t feel comfortable putting our name, but we know. So that’s why I like this test the most. There are a couple of technological solutions that try to address some of these problems. In my opinion, they are not that great, but I have to mention them. So one is using AI writing detectors, and especially if you turn them on on LinkedIn, everything lights out as being AI generated, fully or almost fully. So that’s the funny part. But what they’re trying to do is measure structure, colon density, bullet density, same length sentences, a few more things. But you can use a detector on your own work as means to try and move out of the average. But I wouldn’t give them too much weight, as I don’t think that they detect quality or originality, which is what I think we should care more. And also humanizers, which are like AI skills or AI processes trying to make text sound more like human. Mostly, I think that they are slop trying not to look like slop, because humanized text is still flagged about three quarters of the time by those AI detectors. And the reader’s problem was always the lack of human unique point of view, not the specific style. So you can try them to improve things that are already interesting, but I wouldn’t use them as a means to skip the the thinking. And speaking of skipping the the thinking, I want to have a fine line between outsourcing everything to AI and still getting the efficiency boost where human’s judgment matters. So some places not only AI can help us, but it can also create many more opportunities that without AI we cannot do. While in other places, I am wholeheartedly under the belief that those should remain human. And you have a partial list here, but I am very curious because I.

【译】有几个试图解决这些问题的技术方案。我的看法是它们都不太行，但必须提到。一个是 AI 写作检测器：尤其是在 LinkedIn 上打开它们，什么都会被点亮为「AI 生成」或「几乎全 AI 生成」——这是它好笑的部分。它们尝试测量的是结构：冒号密度、列表密度、等长句子等等。你可以把检测器用在自己的工作上，作为把自己挪出「平均值」的手段，但别太当回事——我不认为它们能检测出质量或原创性，而那才是我们更应该关心的。然后是「拟人化工具」（humanizer），就是那些让文本听起来更像人写的 AI 技能或流程。我的看法，它们大多是「努力让自己看起来不像 slop 的 slop」——拟人化后的文本仍然有大约四分之三的概率被那些 AI 检测器标记出来。而读者的问题从来都是缺少人类独有的观点，不是某种具体文风。所以你可以用它们去改进本来就有趣的东西，但别把它们当成跳过思考的手段。说到跳过思考，我想划一条线：一边是把一切外包给 AI，一边是在人类判断重要的地方仍然拿到效率提升。有些地方 AI 不仅帮得上忙，还创造出没有 AI 就不存在的机会；另一些地方，我发自内心地相信应该留给人类。你这里有一份不完整的清单，我也很好奇——我知道你对这件事非常有主见，Nathaniel：你怎么决定哪些留在自己手里、哪些交给 AI？

[840s - 960s]

I know that you’re very opinionated about it, Nathaniel. So, how do you decide what to keep in your hands and what to delegate? Yeah, so I think first of all, what you said about understanding that you’re going to have different approaches to this for different types of things that you produce is really important, right? There is a wildly different burden on some original content that you’re trying to produce into the world versus an email, and in fact, I don’t think anyone has any problems anywhere with slop AI emails as long as they’re concise and efficient and just getting to the point. It’s almost like it’s not slop by virtue of the fact that it just does its damn job and gets out of the way. That it doesn’t matter if all the same words are used because that’s its purpose. Its purpose isn’t to make you think or provoke ideas or whatever. So, there’s a huge band of ways that I think about this, and I guess maybe let me use the example of this new podcast how we got here because I’ve sort of had to get intentional about all these processes that I’ve evolved naturally over time with the AI Daily Brief, and I had to put them into practice in a very specific way with this new show. So, this new show, by the way, speaking of things that are completely out of sync with people’s attention habits, for those who haven’t heard it yet, it’s a 45 minutes to an hour long deep dive on some topic about the entire cultural, technological, social, economic history that got it that way. First episode is about the 15 year history that led to America falling out of love with tech that shapes the data center debate, and so this is an extremely information dense, long form content piece that has all sorts of potential problems. There’s the problems of where the information comes from. There’s the problem of keeping people interested in it. There’s the problem of the point of view which drives it along. There’s the problem of the narrative architecture, and this is a show that would be impossible for me to do as I do it without AI, but also has all the potential to fall into every single trap of AI if not done well.

【译】是的，首先你说的这一点非常重要：对不同类型的产出，你要有不同的处理方式。你要向世界发布的原创内容，和一封邮件，承担的责任完全不同。事实上我认为没有人会介意 AI 写的内部邮件——只要它简洁、高效、直奔主题。它几乎「因为老老实实把事办完就退场」而不算 slop：用没用那些同样的词根本不重要，因为邮件的目的本来就不是让你思考或激发想法。所以我处理这件事有一个很宽的谱带，也许用新播客《How We Got Here》来举例最合适：我不得不把在 AI Daily Brief 上自然演化出来的那一整套流程，非常刻意地在一个新节目上落地。顺便介绍下这档新节目——说到和大众注意力习惯完全错位的东西——没听过的朋友：这是每期 45 分钟到一小时的深度节目，讲某个话题的文化、技术、社会、经济史，讲它是如何走到今天这一步的。第一期讲的是让美国「不再爱」科技、并塑造了当下数据中心之争的十五年历史。所以这是一个信息密度极高的长内容，会踩到所有可能的坑：信息从哪来的问题、让人保持兴趣的问题、驱动它的观点问题、叙事架构的问题。这是一个没有 AI 我不可能以现在的方式做出来的节目，但也是一个如果做得不好就会掉进每一个 AI 陷阱的节目。

[960s - 1080s]

And this is not a show where I can. I’m not an expert on each of these topics. The first ones that I’m doing, I’m much more, much closer to an expert in because that’s why I’m starting with them. But as I get farther out and I’m doing the history of GLP-1s or something like that, like it’s not. I’m going to have to sort of turn myself into the translator, even, even without being deep in that space. So the way that I think about deslopping that, or making sure that there isn’t slop that enters into that process, is one, having a point of view first on which parts are mine and which parts are AI’s and how those go together, right? Knowing that it is going to be an intentional plan. That’s the foundation upon which it stands. Two first things that I know when I’m thinking about that plan are that one, there is no way for me to do all of my own research, so that has to be AI. But two, I am unwilling to allow AI to form my point of view on a topic. So that’s those are my starting points. And so the first thing that I do with any episode is I am taking full advantage of AI’s ability to very quickly read everything that exists and has ever existed about a topic and turn that into any medium of getting myself up to speed that’s useful. I can have the five-minute overview report. I could have it be the 50-page research report, whatever it is. But I’m producing research reports of all different varieties using all sorts of different models because there are differences I find in what things they choose to emphasize. Not so much better or worse. I just want all of it. And I’m using that to make sure that I either have a point of view that that is my own or that I’m sort of challenging it and building it out from whatever point I’m starting with. I think a great example of why you literally cannot outsource your point of view in content to AI is one part of that process that I often do is I will have an AI model, Steelman.

【译】而且这不是一个我能装内行的节目：我对每个题目都不是专家。最先做的这几期我还算接近专家——这也是我先做它们的原因；但越往外走，比如做 GLP-1 药物史的时候，我就不得不把自己变成一个「翻译者」，哪怕我并不深入那个领域。所以我对这件事「去 slop」——确保 slop 不进入这个流程——的思考方式是：第一，先有一个观点：哪些部分是我的、哪些是 AI 的、它们怎么拼在一起。知道这将是一份有意的计划，这是整个事情站立的地基。第二，思考这个计划时我最确定的两件事是：一，我不可能自己做完全部研究，所以研究必须交给 AI；二，我不愿意让 AI 替我形成对一个话题的观点。这就是我的出发点。于是我做每一期的第一件事，就是充分利用 AI 极快地「读遍一个话题曾有过的全部内容」的能力，把它变成任何——

[1080s - 1200s]

Two wildly different perspectives, and have it argue vociferously for each side of a particular issue or phenomenon, and that means that it’s even if in one run it happens to decide one versus another, it’s completely able to argue as though it is deepest conviction, whichever thing you want it to. And so the fact that it ultimately it is not a a person that has a perspective; it just has the sum total of all the arguments that it can go out and source means that you really you have to have that that sort of that point of view. Everything builds from there. So I’ve got the research, which is largely in the hands of AI, although I do actually have a research team that that helps interact with it, and then a point of view. I also tend to be very strong for myself in architecting the narrative and the flow of the argument. That’s something that is extremely important to me, and that I find I almost never agree with how AI wants to construct an argument. I don’t necessarily think that’s something that I every person has to follow. You know, I’ve been doing this type of long form content for so long now that I have very specific instincts and feelings about how arguments should be built. But that’s one where again I’m not willing to outsource it. So the human side, I’ve got the point of view and the architecture of the argument. Then I’ve got the AI side. I’ve got the research base. From there, I don’t script anything that I do. I do it extemporaneously. But the way that I’m able to do that is I architect that narrative flow, and then I put all the sort of materials and talkover visuals that I’m going to use in the sequence that they’re going to come. And that once again, I’ve now brought AI back in. So I’ve got the narrative architecture in very sort of precise, specific outlines, and then I have AI do bullets basically that have either key facts or key numbers, like basically the things that I want to hit. Same type of research dossier that we actually produce with the human for the AI daily brief. That’s a combined human AI effort.

【译】——能让我快速进入状态的媒介：可以是五分钟概览报告，也可以是五十页研究报告，什么都行。我用各种不同模型生成各种规格的研究报告，因为我发现它们选择强调的东西确实有差异——不是谁好谁坏，我全都要。我用这一切来确保：我要么拥有一个属于自己的观点，要么正从我的起点出发去挑战它、构建它。一个「你为什么完全不能把内容里的观点外包给 AI」的好例子：我流程里常做的一步，是让一个 AI 模型对两个截然不同的视角做最强论证（steelman），让它就某个议题或现象的每一方都激烈辩护。这意味着：哪怕某一次运行里它恰好倾向了一边，它也完全能够以「最深的信念」为你指定的任何一边辩护。它终究不是一个有视角的人，它只是它能检索到的所有论点的总和——所以你真的必须自己带着那个观点来，一切都从那里生长。于是我有了研究——大部分在 AI 手里，虽然我确实也有一个研究团队帮忙与它互动——和一个观点。我在给叙事和论证流程做架构这件事上对自己也非常强势：这对我极其重要，而且我发现我几乎从不认同 AI 想怎么搭论证。我不觉得每个人都必须照做——我做这类长内容太久，对「论证应该怎么搭」有非常具体的直觉和感受——但这是另一个我不愿外包的东西。所以人类这边：观点加论证架构；AI 那边：研究底座。在此之上，我不写稿，全部即兴。我能这么做的方式是：架构好叙事流，把要用的材料和解说配图按出场顺序排好——AI 又在这里进场：我有精确具体的叙事架构大纲，再让 AI 做带关键事实和关键数字的要点清单，也就是我必须命中的那些点——和我们在 AI Daily Brief 用人工加 AI 一起做的研究档案是同一类东西，那是人机合作的产物。

[1200s - 1320s]

For the new show, how How We Got Here, and then I do my normal extemporaneous talk over that thing. So that’s sort of it’s basically this never ending back and forth where the most important parts of perspective and narrative architecture remain with mine, and I use AI to make that all viable to do it in the speed at which it needs to get produced. The one other piece that I thought was worth noting, though, because I’ve sort of given this basically, there’s no slop from the ground up. I don’t use AI to produce the script of the content, but for each of these shows, I’m producing a companion essay, and there is simply no way for me to write a six thousand word companion essay in the same week that I’m doing these shows on top of the AI daily brief. Maybe someone could, but I can’t, and so that inevitably, there is no way to have AI write something that’s six thousand words without it being slop. And so I basically had to go figure out how to handle this, and so the two pieces to deslopfy that particular piece of content because I did think it was incredibly important because a lot of people want to read something long form like this rather than just listen to it, or they want both. There’s a chance to have charts and visuals that are totally different with that type of written content, and so the two things that I did were one, instead of having it just write from the outline that we had concocted, I actually give it the raw transcript of my talk over text, so it has more than just the outline. It has the actual way that I phrase things and intone things, and it can hew fairly closely to that, even if I’m giving it permission to sort of expand and bring in more details, context, background, or whatever. That was part one, but even with that, part two was I found that there is just a can’t escape it five to ten hour rewrite process that I’m going to go through with those essays. I actually ended up building a content management system into the How We Got Here website so that once I’ve printed that highly visual essay, I can just go through paragraph by paragraph. I can reorder paragraphs.

【译】有了这些，基本上就是一场永无止境的来回：最重要的部分——视角和叙事架构——留在我手里，我用 AI 让这一切以必须的生产速度变得可行。还有一点值得一提，因为我刚才给的版本基本上是「从地基上就没有 slop」——我不用 AI 生成节目内容本身——但每期节目我要产一篇配套长文，而我根本没有办法在做节目、做 AI Daily Brief 的同一周里，亲手写出一篇六千字长文。也许有人可以，我不行。而不可避免的事实是：让 AI 写六千字的东西、还不成为 slop，是不可能的。所以我得自己去想清楚怎么处理这块内容——我确实认为它极其重要，因为很多人想读这种长文而不是只听，或者两者都要；而且书面内容还有机会配完全不同的图表和视觉元素。我做的两件事：第一，不是让 AI 从我们拟的大纲直接写，而是把我口述解说的原始转写稿喂给它，所以它拿到的不只是大纲，还有我实际的措辞和语气，它可以相当贴近地跟着走——即便我允许它扩展、补充更多细节、背景之类。这是第一步。但即便如此，第二步是：我发现有一段怎么都逃不掉的五到十个小时的重写过程在等我。我最后干脆在《How We Got Here》的网站里建了一个内容管理系统：把那篇高度视觉化的长文弄出来之后，我可以逐段过、重排段落、

[1320s - 1440s]

Graphs and I edited just a huge amount because my writing, my talking doesn’t sound like the average AI writing. There’s things that that are literally wrong, like an English teacher would say I do poorly, but they sound like me. And so this is a concession where I think accepting that there is a limit to the efficiency here, to basically embrace the fact that without AI, there’s no way that this 5,000 word essay even exists. But there is still a cost to it, and that cost is the five to ten hours that I’m going to take editing it to make it not this sort of unbearable AIisms, even if it follows my own narrative architecture. So those are the versions of this that I’ve been thinking about a lot recently. I think you know, in terms of big sort of takeaway points, the key elements are one, having a perspective and point of view on which parts are supposed to be yours versus AI. Two, I think not outsourcing one’s point of view ultimately. That the key thing that you are trying to achieve or come back to, and then three, accepting that there is no perfect efficiency, even if AI can do it all, and you have to figure out with whatever time budget you have left how much desloppifying you’re willing to do, and consider that part of the cost of doing business. I might think you might have scared away some people with your five to ten hours editing, but let’s try and give like a more general perspective that I think agrees with everything that you’re saying, but gives maybe a few other point of thinking for people with other positions and so on. So amazing to hear how you run high quality with like AI first and AI native work. And one thing that I want to say, especially to the people who feel that what you’re saying sounds intimidating, that it needs to be proportional to the stakes. So like you said, the internal email, the one line that nobody cares about, can definitely be fully AI made, and I think nobody is looking for like artisanal meeting transcript notes or summaries. But I do think that anything that represents your brand, that’s supposed to be your original.

【译】调整图表，并且做大量的编辑——因为我说话的方式、我写东西的方式听起来不像平均水平的 AI 写作。有些地方按英语老师的标准甚至是错的，但它们听起来像我。所以这是一种让步：接受这里的效率是有上限的——坦然接受「没有 AI，这篇五千字长文根本不存在」；但它仍然有成本，这个成本就是我那五到十个小时的编辑，把那些难以忍受的 AI 腔去掉——哪怕它遵循的已经是我自己的叙事架构。这些就是我最近一直在想的版本。总结成要点：一，对「哪些部分应该是你的、哪些是 AI 的、它们怎么拼」有一个视角；二，最终不外包自己的观点——那是你要不断回锚的东西；三，接受不存在完美的效率——就算 AI 能全干，你也要用手头剩下的时间预算决定愿意做多少「去 slop」，并把这部分当作做生意的成本。——我猜你那个「五到十小时编辑」可能吓跑一些人，那我们试着给一个更一般的视角：它与你说的全部一致，但也许能给处境不同的人几个别的思考点。听你讲怎么用 AI 优先、AI 原生的方式跑高质量，真的精彩。我特别想对那些觉得「你说的听着吓人」的人说一句：投入要与事情的分量成比例。就像你说的，内部邮件、那行没人在意的字，完全可以让 AI 全包；我也认为没人在等「手工艺级的会议转写笔记或摘要」。但任何代表你品牌的东西、本该属于你原创的——

[1440s - 1560s]

Original work that is customer-facing and has high impact to your company. This is the type of work that requires this intentionality, like and Nathaniel just pointed out, and that everybody should kind of carve their own belief processes and priorities. At the very minimum, just think about the process, and that’s my overall playbook on how I believe that everybody should deslop critical and important work. A new study from KPMG and the University of Texas at Austin found that when people work with AI, similar skills don’t guarantee similar outcomes. Researchers studied more than 500 early career professionals and found that the best performers consistently amplified the value of AI by guiding, evaluating, and refining its outputs. These top performers, called AI amplifiers, weren’t defined by what they knew alone, but by how they worked with AI. Learn more about what separates AI amplifiers from everyone else at KPMG.com/us/AI-amplifiers. The best teams don’t have a single star carrying everyone else. They know their own strengths and each other’s weaknesses and play to both. That’s the team Robots and Pencils has built on purpose. Nobody there is grinding through busy work to pad a headcount number. People come for the hard problems and they stay because everyone around them is leveling up at the same time. In a market full of companies that are just trying to hire fast, that’s worth a look. Check out RobotsandPencils.com/careers. This episode is brought to you by CData. Companies rolling out AI keep hitting the same wall. Token spend is climbing, security is holding agents back, and no one knows how AI is truly being used. The root cause: AI is cut off from the systems and context your business runs on. That’s why CData built Connect AI Gateway, one control point between your AI and your systems. Built on the connectivity layer powering Palantir, Google, Salesforce, and hundreds more, each agent gets exactly the data and tools it needs. It governs down to the record and lets you finally see your org’s complete AI usage in one.

【译】——面向客户、对公司有高影响的工作，就是需要这种刻意程度的类型——就像 Nathaniel 刚才指出的，每个人都应该刻出自己的流程与优先级；最低限度，至少想一想流程。这就是我关于「每个人都该如何给关键而重要的工作去 slop」的整体打法。（以下为本期集中插入的赞助商广告口播，按原文收录。）KPMG 与德克萨斯大学奥斯汀分校的新研究：与 AI 协作时，相近的技能并不保证相近的成果。研究者研究了五百多名早期职业专业人士，发现表现最好的人始终通过引导、评估和完善 AI 的输出来放大它的价值。这些被称为「AI amplifiers（AI 放大者）」的顶尖表现者，其定义不只在于他们知道什么，而在于他们如何与 AI 工作。详见 KPMG.com/us/AI-amplifiers。最好的团队不靠单个明星扛所有人——他们清楚自己的长处和彼此的短板，并且两头都发挥。这是 Robots and Pencils 有意打造的团队：没人在那里靠杂活凑人头；人们为难题而来，留下是因为身边每个人都在同步升级。在一个都在拼命快招人的市场里，这值得一看：RobotsandPencils.com/careers。本期节目由 CData 带给你。铺开 AI 的公司不断撞上同一堵墙：token 开销在涨，安全把 agent 拴住，没人知道 AI 到底被怎么用。根因是：AI 与你企业赖以运转的系统和上下文被切断了。所以 CData 做了 Connect AI Gateway——你的 AI 与你的系统之间的单一控制点。它构建在为 Palantir、Google、Salesforce 等数百个平台供能的连接层之上：每个 agent 精确拿到它需要的数据和工具，治理细到记录级，让你第一次能在一个地方看到组织完整的

[1560s - 1680s]

Place accurate answers, a fraction of the token cost, and a gateway that learns your business with every interaction. See how at cdata.ai. Every AI coding tool on the market does the same thing. First, it starts writing code. Blitzy does the opposite. Before writing a single line, Blitzy spends days reverse engineering your entire codebase. Thousands of agents ingest millions of lines, mapping every dependency, every undocumented constraint, every architectural decision made over the last decade. The result is a dynamic knowledge graph that understands your software the way a principal engineer would after thirty years in the building. Other tools guess at context with grep searches and markdown files. Blitzy never guesses. It builds true understanding first, then delivers over eighty percent of entire software epics autonomously, validated, end-to-end tested, production grade pull requests. That’s why Fortune five hundred engineering teams trust Blitzy with the codebases that matter most. See for yourself at blitzy.com. That’s b l i t z y dot com. I want to divide it into stages. And stage one happens before you open the tool. And the very first thing, like you just heard, that Nathaniel saying is that I think that you have to know what you’re thinking before you open the tool. Go take a walk, read the research until you you have a view of your own, no matter how messy it is. My favorite version is to brain dump everything that you’re thinking once you have some original thinking. Dictate ramble to the AI tool. Don’t care about the structure. Just tell it everything. What you heard. What worries you. What you’re thinking. What you’re sure you don’t want to do. And then let the model and the tool figure out the structure. This episode, for example, for me started with fifteen minutes of me talking to the AI tool. Very messy, but I just told it everything I’m thinking about to help to to help write this down. And then the AI tool will help you shape many of the thinking or speak back to you what it heard. And I think that in many cases, this is funnily enough the the one thing that people skipping that has the highest downstream impact. So that’s.

【译】AI 用量：准确的答案、零头的 token 成本、一个随每次交互学习你业务的网关。见 cdata.ai。市面上的所有 AI 编码工具都一个样：第一件事就是开始写代码。Blitzy 反着来：在写下一行代码之前，Blitzy 先花几天逆向工程你的整个代码库——数千个 agent 消化数百万行代码，映射每一处依赖、每条没写进文档的约束、过去十年做过的每个架构决策。产出是一张动态知识图谱，像在这栋楼里干了三十年的主任工程师那样理解你的软件。别的工具靠 grep 搜索和 markdown 文件猜上下文，Blitzy 从不猜。它先建立真正的理解，然后自主交付整个软件 epic 的百分之八十以上：经过验证、端到端测试、生产级的 pull request。这就是财富五百强工程团队把最重要的代码库托付给 Blitzy 的原因。自己去看看：blitzy.com，b-l-i-t-z-y 点 com。（广告口播结束。）现在我想把它分成几个阶段。第一阶段发生在你打开工具之前。而第一件事，就像你刚才听到 Nathaniel 说的：你必须先知道自己在想什么，再打开工具。去散步、去读研究，直到你有了一个自己的看法——不管多乱都行。我最喜欢的版本是：一旦有了一些原创思考，就把脑中一切倒出来——对着 AI 工具口述、随便说，不管结构，把一切告诉它：你听到了什么、担心什么、在想什么、确定不要什么，然后让模型和工具去理清结构。这一期节目就是这样起步的：我对着 AI 工具讲了十五分钟，非常乱，但就是把我想到的一切都告诉它，让它帮我写下来。然后 AI 工具会帮你把很多思考理出形状、或者把它听到的复述给你。有趣的是，这恰恰是人们跳过的一环里，下游影响最大的一环。这就是——

[1680s - 1800s]

Number one, you have to think before you open the tool. Habitually, I want you to also tell the tool how confident you are in whatever it is that you’re thinking, because the model is built to be agreeable. So if you’re alive with a very strong opinion, it will probably help you justify every one of your strong opinions. As you said, it will still menu. So say how sure you are and how hard you want it to challenge you. For example, I’m about 60% sure the problem is onboarding pushback. Hard before you agree with me, that starts a completely different conversation. Even when you are sure, ask for the counter view. Often it will help sharpen your own point of view. The other thing that I want you to do before you start with the tool is to think who the audience is and what you know about them and what experience that you want to for them to come out with whatever it is that you’re creating. It can be the feeling, the the action that they want to take, or anything else that matters, including the standards of length, of format, tone, anything that will make it yours and appropriate to the task at hand. It gives the model a target, and it also helps you validate against something once you do the work together. And of course, all of that is done prior to the work that you actually kick off with the tool. And then comes the the actual let’s call it shepherding of the execution with AI, and that’s gonna be the second phase. And here, one thing that I think that you can do is to go wide, meaning that the first answer is probably the most probable one, and that’s often the most average one. However, if you ask for six directions that differ in approach, angle, structure, visuals, whatever it is you’re producing, it push it further. You can ask for directions that borrow from other domains, for example, how a game designer or a newsroom would handle this, and then make the tool argue for each direction on its own. And then you squint at the different directions, look at what you like about each, what’s missing from all of them.

【译】——第一条：打开工具之前先思考。习惯上，我还希望你告诉工具你对自己想法的置信度，因为模型天生讨好。如果你带着非常强的观点进来，它多半会帮你把每一条强观点都论证到底。所以要说清楚你有多确定、要多用力地要它挑战你。比如：「我认为问题在 onboarding 的把握是六成。同意我之前，请狠狠反驳。」这就开启了一场完全不同的对话。哪怕你很确定，也要反方观点——它常常能磨利你自己的观点。开工前还有一件事：想清楚受众是谁、你对他们的了解、你想让他们带着什么离开——可以是情绪、可以是想引发的行动，任何重要的东西都算，还包括长度、格式、语气这些标准——任何能让这份东西「是你的」、且匹配当下任务的东西。这给模型一个靶子，也让你在真正开工之后有一个校验的基准。这些都应该在与工具真正启动之前做完。然后是第二阶段：对 AI 执行的「牧养」。这里你能做的一件事是「横向展开」：第一个答案大概是最可能的答案，而那往往就是最平均的答案。但如果你要六个在方法、角度、结构、视觉上彼此不同的方向，就把它推得更远。你可以要借鉴其他领域的方向——比如一个游戏设计师、或一间新闻编辑部会怎么处理这件事——然后让工具为每一个方向各自论证。之后你眯起眼审视这些方向：每个方向你喜欢什么、它们全都缺什么。

[1800s - 1920s]

And then follow the ones that you are most drawn to that represents something that you feel that will serve the purpose at hand. After you do that, I want you to be very intentional, like you heard Nathaniel doing about your interception point, the workflow that might look the most efficient. Here’s what I need go and do it end to end. It does the most damage mostly because when you get like a polished and finished piece sitting in front of you, the sunk cost gets in, and then you, in many cases, will say, okay, I don’t want to go back to the actual narrative. Let’s just ship it like that. Also, with the tools, it’s often harder to move it from this. So decide in advance what are the points where you want to step in, and I want you to do that in very strategic points and decide them in advance, like the outline, and then the first section, and then maybe the the first narrative, and then the draft, and then the finished good. So there are intentional strategic points where you’re reviewing that when the cost of fix or a complete pivot is as low as possible. And a few additional review habits that I think that will go a long way, and which I really really like, is to what I call to convene a board, meaning that before any important work goes out, you have AI review that from distinct personas. So it can be the reader seeing it called a scepting target audience, a domain expert, the busiest person who will receive it, and in many cases also include the devil’s advocate because that can also be the person who is arguing it shouldn’t go out at all. Fix based on the different perspectives that this board of reviewers can provide you, and of course it can also be different segments of your target customers. And of course only use the feedback that you find valuable. And I think that once you use that, you leverage AI to really give you more perspectives and improve the confidence before it ships. So of course this can be done both early on in critical junctions as well as on the finished good. And there are a few additional review habits.

【译】然后跟着那些你最被打动、觉得最能服务当下目的的方向走。做完这个之后，我想让你像刚才听 Nathaniel 讲的那样，对你的「拦截点」非常刻意。那个看起来最高效的工作流——「这是我要的，端到端去干完吧」——造成的伤害最大，多半是因为：当一份打磨完的成品摆在你面前，沉没成本进场了，很多时候你就会说「算了，不想回炉了，就这样发吧」；而且在工具里，从成品往回改往往更难。所以要提前决定你要在哪些点介入，而且要选得战略、提前定好：大纲、然后第一节、然后也许第一版叙事、然后草稿、然后成品——这些是你刻意安排的评审点，选在「修改成本或彻底掉头成本最低」的位置。另外几个评审习惯能让质量走很远的，我特别喜欢的一个是「召集一个董事会」：任何重要工作发出去之前，让 AI 用不同的 persona 来审——可以是把内容当作怀疑目标受众的读者（转写此处含混，按此理解）、领域专家、收到它的最忙的那个人，很多时候还要加上魔鬼代言人——因为那也可能是那个论证「这东西根本不该发出去」的人。基于这个评审团的不同视角去修改；也可以换成目标客户的不同细分，并且只用你觉得有价值的反馈。用过之后，你就是在借 AI 拿到更多视角、在出货前提高信心。这既可以做在关键节点，也可以做在成品上。还有几个评审习惯——

[1920s - 2040s]

That will take you even further with the quality. The first thing that I want you to do with regards to review is to respect the read budget, and I think the most important part of it that you never send something that you wouldn’t read yourself. So if it’s too long or not clear for you, then your audience for sure will not read it either. They’ll just hand it over to their own AI to summarize, and then now we have two models that are talking while two people pretend to communicate. So distilling a page into five sentences forces you to decide what you think, or generating that as a deck or as a podcast or as an HTML or as an image or a video. You have so many options with AI. So choose the format on purpose, and before you even start the production, ask yourself about the audience what they will consume in the most effective way, and then go wild with that. And the best thing that you can do is if you will enjoy reviewing that in that format, or that your target audience will also enjoy consuming that. So that’s how we all get a single output out of the door in a higher quality with like human taste and so on. But I want to also propose a few ideas on how to make it easier and more personalized over time. And one thing that will help a lot is to teach AI your own taste. And of course, you don’t need to be able to describe your own taste in words because most of us just don’t have the vocabulary to be like a linguist or an art director that explains what it is that we like. But if we show a very capable model either our best work, the work by others that we admire, and the work that we dislike, and the things that we dislike, and then ask what the good ones have in common, or how will it profile your own taste and voice and so on, it will put words to your preferences much more precisely than what you can. And for visuals, I really like to create like a mood board of a bunch of images and colors and the feelings that you want to get out of that, and that beats any adjective that you will be able to use. Once you have a good profile of your preferences with regards to visuals or narrative.

【译】——能把质量再往上带一层。第一件事是尊重「阅读预算」，其中最重要的原则：永远不要发出你自己都不会读的东西。如果对你来说太长、不清楚，你的受众肯定也不会读——他们会把它丢给自己的 AI 去总结，于是我们就有了「两个模型在对话、两个人假装在沟通」。把一页蒸馏成五句话，会逼你决定自己到底怎么想；而把它生成为一份幻灯、一期播客、一个网页、一张图、一支视频——AI 让你选择极多——所以要刻意选格式，甚至在开工生产之前就问：我的受众用什么方式消化最有效？然后放开手去做。最好的信号是：你自己会享受用那个格式评审它、你的目标受众也会享受用它消费。这样我们才能让每一份产出带着人的品味、以更高质量出门。但我也想提几个让它随时间更省力、更个人化的办法。有一个会帮大忙：教 AI 你自己的品味。你当然不需要能用语言描述自己的品味——我们大多数人都没有语言学家或艺术导演那样的词汇。但如果我们把自己最好的作品、我们欣赏的他人的作品、我们不喜欢的东西和我们不喜欢的地方，拿给一个非常强的模型看，然后问它：好的那些有什么共同点？它会怎么给你的品味和声音画像？——它能把你的偏好表达得比你自己精确得多。视觉上我特别喜欢做 mood board：一堆图片、颜色、你想要唤起的感觉——这胜过你能用上的任何形容词。

[2040s - 2160s]

Or tone of voice and so on, you can turn it into a skill or something usable with the AI tool, and then that can be the beginning of any work that is intended to sound like you. Still, you need to do the thinking, but at least it will help you sound much more like you. And of course, the downside of having such a skill is that potentially you will start to produce just the one thing because the skill will help you produce it again and again. And audience does remember that you have sameness. So, for example, maybe in my case, you’re very much fed with my falcons. That might be a downside of me codifying my preferences. But what you can do is to break it down on purpose. So, whenever the audience, the scope, or the goal changes, or just enough time has gone by, bring fresh idea from new inspirations and try to kind of go outside the mold by still teach it what you prefer and so on. This is where we want to create a cultural different. The first thing that I want you to do is every time you get slop from someone else, name it, and especially by the way, if you’re in a leadership position, because when something long, generic, and clearly unread arrives, just say so. You can do it kindly or not so kindly. This isn’t a good use of my time. I want you to send me your take in five sentences, or I want you to send me work that you think about more carefully. Just hold the gate and don’t allow for slop to come your way. And of course, when it’s customer facing work, this is where you never ever allow to do because we don’t ever put our team name when we were producing an external work. And lastly, I don’t want you to necessarily hide the AI. If you remember a while back, and Nathaniel, you did an episode on writing, and you were debating on the Stanley Druckenmiller AI written op-ed in The Wall Street, and the question was whether or not we should remove all the em dashes and remove all the AI telltale. And your point of view there was that primarily that was a proxy to let the audience feel as if you didn’t invest the time. So, in my opinion, you shouldn’t completely try to.

【译】一旦你有了关于视觉、叙事或语气的偏好画像，就可以把它变成一个 skill 或 AI 工具里可用的东西，让任何「要听起来像我」的工作都从它开始。思考仍然得你自己做，但它至少能让你听起来更像你。当然，这种 skill 的代价是：你可能会开始只产那一种东西，因为它帮你一遍又一遍地产出同一个——而受众是记得住你的「千篇一律」的。比如拿我来说，你们可能早就被我的那些口头禅喂饱了（转写存疑）——这可能就是我把我偏好固化成 skill 的代价。但你可以刻意打破它：每当受众、范围或目标变了，或者仅仅过了足够久，就从新的灵感里带进新鲜想法，刻意走出模具——同时仍然教它你的偏好。接下来是文化层面。第一件事：每次从别人那里收到 slop，就点名它——尤其是你在领导位置上的时候。当一份又长又平庸、明显没被读过的东西到达时，直说就行，可以温和也可以不那么温和：「这不是对我时间的好用法。把你的观点用五句话发给我」，或者「把你更认真想过的工作发给我」。守住这道门，别让 slop 找上你。而客户向的工作，是永远、永远不许放行的地方。最后，我不想让你反过来刻意藏起 AI。如果你记得前阵子——Nathaniel，你做过一期关于写作的节目，当时在讨论《华尔街日报》那篇 Stanley Druckenmiller 的 AI 执笔评论文章，问题是该不该把所有破折号、所有 AI 痕迹都去掉。你当时的观点是：那本质上是一个「让读者觉得你没投入时间」的代理信号。

[2160s - 2280s]

Hide the AI. You just need to make it feel like something that you originally produced. Any anything to add on that? Because you were very opinionated back then. I was going to use the phrase “load bearing” in today’s episode, and I literally just couldn’t anymore because of how how much it’s been used by AI. So maybe another example of this that’s practical from a company perspective: you’re starting to see companies that are on the the cutting edge formalize and codify their approaches to these issues. So Figma has shared their rules for AI writing, and one of them is if it’s written with AI, that’s not an issue, but label it as such on top so that people have that context to go into it. So, for example, there’s going to be lots of times where what you’re trying to do with that 40-page report is not say here’s the finished work, but to say let’s rip through this, find the parts that are interesting, you know, and so you just identify it as such. I think the point that you made at the beginning of one of the definitions of slop being transferring the cost of work from the person who produces it to the person who consumes it, I think, is acutely accurate. I think that basically almost wholly explains why people have such a visceral reaction to it. Is that you are now imposing a cost upon them? So it’s not even just a dereliction of your responsibility. You’re leaving them holding the work bag in some ways, and so when people are smart enough to know if you are prioritizing yourself over them in a context where you’re supposed to be presenting something to them, and I think that’s what folks have an issue with. So I think transparency, disclosure, understanding that there are different contexts and different needs for different types of work, all have a common through line of respecting other people’s time as much as you respect your own. Yeah, and I think with that we can kind of summarize. And I think looking forward.

【译】所以我的看法是：你不该试图把 AI 完全藏起来，你只需要让它感觉像一件你原创产出的东西。这边有什么要补充的吗？你当时可是非常有主见。——我本来今天想在节目里用「load bearing（承重）」这个词，但实在用不下去了——AI 把它用得太多。也许再补一个公司角度的实用例子：你开始看到处在最前沿的公司，把对这些问题的处理方式正式化、条文化。比如 Figma 公布了他们的 AI 写作规则，其中一条是：用 AI 写不成问题，但要把它标注在最上面，让人们带着这个背景进入。比如很多时候你拿那份四十页报告想说的不是「这是成品」，而是「我们把它过一遍、找出有意思的部分」——那你就把它标注成那样。还有，你开头给的那个定义——slop 是把工作的成本从产出者转嫁给消费者——我认为精准得刺眼。我认为它几乎完全解释了人们对 slop 的那种生理性反应：你现在是在把成本强加到别人身上。这甚至不只是失职，你某种意义上是把工作的包袱甩给了对方。人们足够聪明，能看出你在一个本该向他们呈现东西的场景里，把自己的优先级置于他们之上——我认为这才是大家真正不满的地方。所以透明、披露、理解不同工作有不同的语境和需求——它们有一条共同的线：像尊重自己的时间一样尊重别人的时间。

[2280s - 2400s]

For sure, the models will keep getting better, and the better they get, the more tempting the generic answer will become. So, I want to urge you to look at the five stages that we went through and acknowledge that they hold whatever ships next. Because my hunch is that once a lot of the current hype of the agentic tools and how easy it is to produce with AI, the handmade work will become the premium thing, and people will pay more for work with a human fingerprint. Like you said, even if it’s messy, if it’s not necessarily grammatically correct, but it sounds like you and it’s intriguing and it creates like a human gut feeling, and I do believe that disciplined, high quality work is the job security for everybody who is afraid about their job, regardless of what it is. Because original thinkers, I’m sure, will be rewarded more and more, and companies that keep the slop out of how work travels, especially outside of them, they will stand out and they will attract customers in a way that almost nobody else will be able to manage. Should they leave the generic output and the sameness out there? So, definitely use AI to produce more quality and more work. Push the envelope of the things that you can do, but I want you to also use it as a tool to think more sharply, to expand the horizon and the knowledge, and to see more directions, and then survive the thing that perhaps a tougher reviewer might do for you. We have everything on the companion page that you will probably put in the show notes in the AIDB website. And as always, I have to say in one sentence that part of what we teach in the various courses that we do with the Superintelligent training is exactly to teach you how to be an intentional AI user, like from a best effort to best in class, that does not outsource the thinking and is able to produce original work while still getting the immense value that the technology can give us, but doing so in a more strategic way. I guess maybe maybe I’ll add my my cynical take on this just to close it out. Or cynical is overstating it. The self interested take.

【译】好的，那我们可以总结一下，也往前看。可以肯定的是模型会继续变好，而它们越好，那个通用答案就会越诱人。所以我想请你回头看我们走过的这五个阶段，并且承认：无论接下来发布什么，它们都成立。因为我的预感是：一旦当前这波 agentic 工具和「用 AI 生产如此容易」的热潮退去，手工活（handmade）会成为溢价品，人们会为带人类指纹的作品付更多钱。就像你说的：哪怕它有点乱、未必语法正确，但听起来像你、引人入胜、能唤起人类直觉的东西。而且我真心相信：有纪律的高质量工作，就是每一个担心自己饭碗的人的 job security，无论那是什么工作。因为原创思考者一定会被越来越多地奖励；而那些把 slop 挡在工作流动路径之外、尤其挡在公司大门之外的公司，会以几乎别人都做不到的方式脱颖而出、赢得客户——至于把平庸与千篇一律放任在外面的公司，就留在那儿吧。所以，尽管用 AI 产出更多、更高质量的工作，把你能力的边界往外推；但也请把它当成一件让你思考得更锋利的工具——扩大视野和知识、看见更多方向，然后扛过一个更苛刻的评审者可能对你的拷问。所有内容都在配套页面上，应该会放在节目说明里、AI Daily Brief 的网站上。照例用一句话收尾：我们在 Superintelligent 各类培训里教的，恰恰就是做一个有意图的 AI 使用者——从「尽了力」到「最好的一档」——不外包思考、能产出原创作品，同时仍然拿到这项技术的巨大价值，只是以一种更战略的方式。最后我补一个……「愤世嫉俗」说得太重了，就说是「自利视角」的收尾吧。

[2400s - 2520s]

I believe that a lot of the folks who listen to this show are determined and hell bent on being of the group that uses this technology to get ahead, to not allow this technology happen to them, but to use it for whatever goals they’re trying to achieve. I think that’s mostly why people are here, even if they’re here for the news versus the practical stuff. Overall, they want to be in the top one percent of this new era that where AI is a part. And simply put, what that means has undergone a shift, and I think we’ve all recognized it. But at the beginning, using it well, being in that top one percent or even the top ten percent, meant realizing you could use Gamma to produce better presentations that were more visually consistent, more quickly, and having more comprehensive answers to things. But all of that has been completely democratized as a skill now, right? Everyone has access to those things. It’s no longer just the one percent; it’s everyone who’s touching these tools, unless they’re philosophically against them at this point. And so, the value, the differentiated value, if you’re trying to be in that top percent of users, shifts from just producing a bunch of stuff to using them well. And well, it turns out means with taste and with intention. And unfortunately, it’s it’s a much harder phase in a lot of ways than the first one was. The first one, we all got points for just showing up to the party early. Now, the the differentiation in the product is actually going to matter. But there are huge opportunities in a world of endless seas of content. The stuff that’s good will stand out even more, and people will get enormously good, even without trying to, at filtering out all of this noise. So it’s just the next opportunity, and it involves a different set of skills. But it’s the thing that’s next, so I think it’s what we’re all going to do. Awesome.

【译】我相信这档节目的大量听众，都是铁了心要做「用这项技术跑到前面去」的那群人——不让这项技术「发生在自己身上」，而是用它去达成自己想达成的任何目标。我想这基本就是大家在这里的原因，哪怕你是冲着新闻而不是实操来的。总体上，你们想做这个 AI 时代的 top 1%。而坦白说，「top 1%」的含义已经变了一轮，我想我们都承认了。一开始，「用得好」、进 top 1% 乃至 top 10%，指的是你意识到可以用 Gamma 更快地产出视觉上更一致的演示文稿、拿到更全面的答案。但这一切现在已经彻底民主化了：人人都能摸到这些——除非你在观念上抵触它们。所以，如果你想留在使用者的顶层，差异化的价值就从「产出一堆东西」转移到「用得好」。而「用得好」最终意味着：有品味、有意图。不幸的是，这个阶段在很多方面比第一阶段难得多：第一阶段，我们所有人光凭早到派对就得分；现在，产品本身的差异化才真正要紧。但在无尽的内容之海里也有巨大的机会：好东西会更加突出，而人们会不由自主地——哪怕不刻意——变得极其擅长过滤噪音。所以这只是下一个机会，它需要另一套技能，但它是接下来的事，我想也是我们所有人都要去做的事。太好了。

[2520s - 2536s]

All right, Nufar. Great to have you on the show again. Like Nufar said, we have a companion experience up on the website that has this, but even way, way more sort of practical and tangible, even than the the background slides that Nufar was talking over. So go check it out at AIDailyBrief AI. And thanks again for being here.

【译】好的，Nufar，很高兴你又来。就像 Nufar 说的，我们在网站上放了配套体验，比 Nufar 讲的时候用的那些背景幻灯更实操、更具体，大家去 AIDailyBrief.ai 看看吧。再次感谢你来。