来源:HN 热门榜(1038 分,449 条评论),9 月 21 日 Colin Breck 博客帖;同周另一帖 I don’t want the details(261 分,160 条评论),9 月 23 日 Michael Heap 博客。两个帖子互相不引用,但因为都在反对「请把这件事展开成一份读者可以读懂的长文」的姿态,被本周 HN 的同一条神经拉到一起。

原文概要

Colin Breck 的博客《I Don’t Want to Read What You Didn’t Write》是一篇长文,开头一句话立下基调:

“People who rarely produced original writing are suddenly producing extensive design proposals, business plans, documentation, presentations, tickets, pull requests, blog articles, and meeting summaries, all generated by AI. It is all unreadable.”

他接着点名两类折磨:把已经做好的事用 LLM 展开成 design document,把 PR diff 写成 LLM-to-LLM 摘要——「why are we doing this? What is the value? How risky or urgent is this work?」这些问题反而没人提。讽刺点是有人用 LLM 写的「私人消息」来回应他的长文——他看到时已经没有任何想读、想回的冲动。

Breck 的支撑数据来自 Cynthia Dunlop 的调研《Report: How developers react to AI-scented blog posts》:能看出 AI 味的读者中,78% 直接停止阅读,71% 之后避开作者;98% 的受访者更偏好作者亲笔写的原版。Bryan Cantrill 把这条反应总结成「revolt of the reader」——读者的脑子在「句子中间」就会拉下 eject 把手。Breck 同时引用了一段 Bjarne Stroustrup 在 panel 上的发言:

“When I read a book, I really feel that if I can hear the author’s voice ringing in my head—complete with weird accents and peculiarities of their speaking patterns—then I feel like something has succeeded.”

文章后半段 Breck 解释自己「写得更快但不爱读别人写的 AI 长文」的原因:他「用 AI 校对」、「让 AI 帮我检查某段 LaTeX 是否跟源码一致」、「让 AI 帮我补 BibTeX」——但他自己写 LaTeX 源、给 AI 提供 journal style guide、历史 paper、源码、生产配置、log、metric 这些上下文。结论是:AI 让写论文更快,但不能让 AI 替你写作——把 AI 当 collaborator、不当 ghostwriter。

Michael Heap 的博客《I don’t want the details》是一段故事:他在一次 post-incident call 上被叫到 SVP 那一级,准备复盘事故原因,SVP 在他开口前打断了他:

“I know that if we get into the details, the reasons will be perfectly reasonable. You’ll explain what happened, I’ll understand why everyone made the decisions they made, and I’ll empathise with you. Then it’ll happen again. So I don’t want the details. I want to know what we’re changing.”

Heap 起初以为这是傲慢,几秒钟后才意识到这是信任——「I already believe you. Now let’s talk about what happens next.」后文三条要义:第一,发生故障后默认会问「why did this happen?」,但「understanding an issue is not the same as fixing it」——一份好解释不一定是份好修复;第二,「如果所有人 involved 都「reasonable」,什么都不会改」;第三,组织需要的是「system prevents this class of mistake」而不是「we’ll be more careful next time」。最后一条是节制:「Not every failure deserves a new process.」有的时候接受「consciously accept this risk」比硬塞一个新流程更值。

两篇帖子共享一种姿态:默认对方有判断,扔掉无效的解释层,请求「能复用的那个动作」——前者让你少读 LLM 改写的长文,后者让你少讲已经不能再改变过去的复盘叙事。

讨论焦点

信息论倒转:发给对方的应是「prompt」,不是 LLM 输出

主帖里重展开最久的子线是 hatthew 提出的信息论 framing——把 AI 当语言而不是渠道:

“Writing is fundamentally the transfer of information from your brain to my brain. If you have 1000 bits of semantic information you want to transfer, you can’t give 300 bits of semantic information to an LLM and have it fill in the remaining 700, because it doesn’t know what those 700 bits are. If it’s able to guess those 700 bits correctly, then they aren’t true semantic information, and you really only have 300 bits you want to transfer. You might as well transfer those bits to me directly, rather than having the LLM add on an extra superfluous 700 bits that I then have to filter out.” — hatthew [c:49794483]

(译文:写作本质上是你脑子到别人脑子的信息搬运。你有 1000 比特想搬运,你不能自己留 300 比特把剩下 700 比特让 LLM 补——它不知道那 700 比特是什么。它猜对了,那 700 比特就不是新信息;那其实你只有 300 比特真要传。你不如把那 300 比特直接传给对方,不要让 LLM 加 700 比特的冗余让对方过滤。)

这条 framing 在讨论区被反复改造,selcuka 把信息论翻成工作流:

“Please don’t send me AI-generated text. Send me your prompt instead. It is highly likely that I will understand it without needing an LLM, and if not, I can do it myself.” — selcuka [c:49795523]

(译文:请别给我发 AI 写的成品。把你的 prompt 发过来。我大概率不需要再开 LLM 就能读懂;如果读不懂,我自己开 LLM。)

为什么这条路是省 token?japanuspus 把它说成「LLM-to-LLM roundtrip」:「To communicate with me, have your LLM expand your message to include relevant parts of your context, then I will have my LLM summarize it as briefly as possible with respect to my context.」feoren 一句话打断他:

“Just ask them for their prompt, and then put that in your own LLM with your own context. Why do you need their LLM’s fluff?” — feoren [c:49806604]

(译文:直接要 prompt,把 prompt 扔进你带自己上下文的 LLM 里跑一遍就行了。要对方的 LLM 输出干嘛?)

steveBK123 把这条压缩得更狠:

“We’ve invented the opposite of lossless compression” — steveBK123 [c:49795522]

(译文:我们发明了无损压缩的反义词。)

“A broken telephone.” — copperx [c:49795674]

(译文:一台坏掉的电话机。)

「好解释不是好修复」与好 AI 写作其实是同一件事

Heap 那条帖里有句几乎原话点名的座右铭:「A good explanation can make things worse. Once everyone agrees that the behaviour was reasonable, the urgency to change anything disappears.」主线帖里 hatthew 的信息论逻辑刚好和这句话是对仗的——「好解释」把同理心用完,「好 AI 写作」把同样的关切 article-grade 完成;两者都不解决问题、只完成共识。

讨论区里把两件事拉到同一根线上的,是 TheOtherHobbes 那一长串对「writing 不是「信息传输」而是触发读者 latent space 一段体验」的分析——把”形式”理解为不同 register 的堆叠,而 LLM 默认 register 是「营销 + ad speak」,「偶尔又有自己一套技术 register」,结果换到 design doc 时像「LinkedIn speak on a wedding anniversary」。

“The real problem with LLM writing is that it’s good at the technical layer — the grammar and spelling — and has some insights into the rest. But the default content style is marketing and ad speak. And recently it’s developed a weird and unique hybrid style which applies marketing fluff and pretension to technical content like code comments. So you get one register instead of all of them.” — TheOtherHobbes [c:49795680]

(译文:LLM 写作的真正问题是它在技术层(语法、拼写)已经及格,对剩下的层次也有些触觉。但默认内容风格是营销 + 广告话术。最近又叠出一套独特的混合风格——把营销的小话和装腔调弄到代码注释这种技术内容上。所以你看到的只是一套 register,而本该多套。)

“I don’t care how good the LLM gets. If I know some text was written by LLM I’d much rather know the prompt — the seed of intent.” — carbyau [c:49797232]

(译文:不管 LLM 写得有多好,只要我知道这段是它写的,我更想要 prompt——意图的种子。)

SVP 真的「不要细节」吗?

Heap 那帖里最尖锐的反方意见来自 FartyMcFarter,他把这套说辞一层层拆开:

“If you trust them completely, you don’t need to know what happens next either. Just trust them to do the right things, your leadership isn’t required. If you don’t trust them completely, how can you know if ‘what happens next’ is appropriate without knowing the details?” — FartyMcFarter [c:49815911]

(译文:你完全信他们,你也不用知道「接下来做什么」——信任就够了,不劳你领导。如果你没有完全信,你不知道细节怎么判断「接下来要做的」合不合适?)

dpark 顺着这条给出 bounded framing,把这条讨论从「全信不信」挪到「分层信任」:

“‘Trust’ is usually bounded. I trust my elementary aged son to get dressed for school without supervision. That doesn’t mean I trust him to start a campfire without supervision… Trusting that people are capable and well intentioned does not mean trusting that they will accomplish the same thing without you.” — dpark [c:49816464]

(译文:信任是有边界的。我信我小学的儿子可以没人看着自己穿校服——但我没信他可以没人看着点篝火……信他有能力且善意,不代表信他离开你也能搞定这件事。)

malvim 把伤疤说穿了:

“We’ll just fire joe — Ok cool. I agree with the parent comment. You need SOME level of detail, otherwise your decision might as well be a guess.” — malvim [c:49816201]

(译文:「我们准备把 joe 开掉。」「好的。」——我跟楼主观点一致,你总得知道点细节吧,不然你的决定跟瞎猜没差。)

drunken_thor 从工程政治视角把这条怀疑推到极致:

“It isn’t about trust, it is about performing accountability so the SVP can say they did their job. I think the author of the article is going out of their way to be charitable to the SVP for no reason. That meeting was a reassertion of power, not a wise move made by a sage.” — drunken_thor [c:49818269]

(译文:这跟信任无关,这是 SVP 在表演「履职」——回头他就能说这一轮事故他处理过了。我看文章作者对 SVP 的善意完全没来由;那场会议是权力的再确认,不是高人的远见。)

linster 在第一层就给文章定性:

“The quote box at the start of the article is the essence of it, the rest is AI-speak dressing it up.” — linster [c:49815896] [thread 2]

(译文:开头引言那段才是文章的全部,剩下的都是拿 AI 腔在包装它。)

bluGill 在元层面把这套工作流叫成「打分回路」:SVP 把每一次「what happens next」的回答当一次「得分样本」,对照自己心里的合格线判断:合理就过、不合理下次再追答,不指望单次答完美——把「完美信任」换成了「次数收敛」。

如果不能解释,就丢一个 prompt;如果不能改,就丢一个 change

主帖里 monkeydust 把「发 prompt + 摘要」做了一个公开承认版的工作流:

“When sharing summaries from calls that I attended but my team has not but I think they would find interesting and I simply don’t have time to handcraft it myself I use LLM and attach my prompt and the transcript. Attaching the prompt initially threw some people… ‘wait, your admitting to using AI…’ ‘err yea, unlike you with that PowerPoint you sent me last week’.” — monkeydust [c:49797950]

(译文:我开了一个我自己出席但团队没出席他们可能感兴趣的会,我没空自己写,我就把 LLM 摘要 + 我自己的 prompt + 会议 transcript 全发过去。开始有人吐槽:「你承认用 AI 了……」「是啊,至少我承认,不像上周你自己发给我那个 PowerPoint。」)

这条工作流把「我不读」变成「我发不了」。在 Heap 那帖里 bluGill 做了对偶的事——把 SVP 的「我不能每个事故都重读」换成「那我换个话术——『I’ll trust the next answer; next time, I’ll dig into the details』」:

“Used and phrased better: trust and only sometimes ask for the details, like cutting the just-shuffled deck in a card game, sometimes the player to the dealer’s right will just tap the deck instead of cutting it, in effect saying ‘I trust that shuffle’. Sometimes, certainly enough to both keep the SVP well-grounded in the details, and to keep the team knowing the SVP is both watching and cares about the details, the SVP MUST ask for the details. Trust and verify. In terms of phrasing, ‘I don’t want the details’ is all about the SVP sending the message s/he doesn’t want to be bothered. The phrasing should be: ‘I trust you on the details on this one; let’s go straight to what do we do to manage the next one of these?’.” — toss1 [c:49816683] [thread 2]

(译文:用得「更」对:时不时才问一次细节,像刚洗完一副牌让庄家右手的人切,他有时候只轻轻拍一下牌——意思是「我信这把洗牌」。但要经常够多地切牌,让 SVP 接地气、让团队知道 SVP 既在看也在乎细节。SVP 必须在某些时候亲自问细节。Trust and verify。话术上,「我不要细节」传达的是 SVP 不愿被打扰。应该这么说:「这次我信细节;下一步我们要怎么管同类的下一次?」)

LLM 被反向用:你阅读的麻烦正好是它的强项

这场讨论里最具体的反讽来自 ageitgey——他自己不用 LLM 写文章,他用 LLM 反查文章:

“Ironically, one of my favorite uses for Claude is to ask, ‘What study is this news article talking about?’ It’s pretty good at digging up the source and related sources. And most of the time, if you read the source, the article is nonsense and gets everything wrong.” — ageitgey [c:49799105]

(译文:讽刺吧,我用 Claude 最顺手的场景之一是问:「这篇新闻到底在讲哪篇论文?」它很会挖原文和相关来源。而且大多数时候,你点进去看原文,就会发现文章是胡扯、所有地方都写错了。)

典型观点一览

立场 用户 一句话
LLM 加噪声不加信号 hatthew [c:49794483] 1000 比特如果 LLM 能猜出剩下 700 比特,那 700 比特就不是新信息
不要给我 AI 输出,给我 prompt selcuka [c:49795523] 你的 prompt 我大概率不需要 LLM 也能读懂
发明了无损压缩的反义词 steveBK123 [c:49795522] 我们发明了 lossy expansion
写作不是「传输」是触发体验 TheOtherHobbes [c:49795680] LLM 默认 register 是营销 + ad speak,把它叠到代码注释是另一种灾难
不管 LLM 多好我都想看 prompt carbyau [c:49797232] 给我意图的种子,不要给我 grep 出来的 fluff
信任是有边界的 dpark [c:49816464] [thread 2] 信人员善意 ≠ 信他们离开你也能搞定;信是分层的
治理需配绩效 FartyMcFarter [c:49815911] [thread 2] 完全信就别问下一步,不信就别跳过细节
LLM 恰好干「反向挖出处」 ageitgey [c:49799105] 80% 时候点进去看原文,文章把研究讲错了
重新措辞比拒绝细节更管用 toss1 [c:49816683] [thread 2] 应该说的是「这次我信细节」,不要让团队感觉 SVP「不愿被打扰」
一切都是表演 drunken_thor [c:49818269] [thread 2] SVP 只是在重新确认权力,他没想真读细节

总体情绪

整体氛围是「先承认现状、再升级工作流」。Colin Breck 用 78%/71%/98% 的调研数据钉死了 AI 写作的「看见就逃」反应;Michael Heap 用一个 SVP 的一句话做出一份「不要这样复盘」的复盘。两帖的反对者都承认问题真实存在,但都指出了反对者在论证时的一个共同弱点:把「坏解释」当成「坏修复」的解药、把「发 prompt」当成「读 AI 写作的解药」,本身也可能变成立场表演。

讨论尾段的收敛点意外地相同——toss1 在线程 2 给出的话术修正、monkeydust 的「prompt + transcript」附件,把两件事并到了一条线上:默认对方有判断、把不能复用的解释层省略、把能复用的动作给齐。monkeydust 公开承认用 LLM 并公开 prompt,这一招的对手「上周你发的 PowerPoint 也是这么搞的」把 SaaS 时代默认偷懒的反讽一次性清掉。

“We’ve invented the opposite of lossless compression.” — steveBK123 [c:49795522]

引用帖子

# 标题 URL
1 I Don’t Want to Read What You Didn’t Write(主线, Colin Breck 博客) https://news.ycombinator.com/item?id=49794330
2 I don’t want the details(相关线, Michael Heap 博客) https://news.ycombinator.com/item?id=49815466
**免责声明**:本文是对 Hacker News 用户讨论的编译与提炼,不代表本网站立场。讨论内容版权归原作者所有。摘要内容仅反映 HN 社区及原文作者观点,不构成对任何博客、调研或方法论的推荐或反对。

本摘要由 AI 模型辅助生成:minimax-cn-coding-plan/MiniMax-M3