It’s interesting that while we are subjectively aware that agents can generate incredibly good and working code, seeing organizations remain very conscientious about details and use peer-to-peer pressure to verify and avoid slop is incredibly interesting. This peer-to-peer check also creates an internal profile of things that cannot be faked, and one must show the level of understanding needed to increase other peers’ trust so that they can push things forward; otherwise, a bad profile develops in people’s minds, which damages reputations. The ability to work on multiple things at the same time and still go into detail while mentally working at a higher level of abstraction seems to be one of the keys to still navigating the world, despite AI may be largely capable of bypassing human capabilities.
The following content is generated by LLMs and may contain inaccuracies.
Context
Your note is about a tension in software and knowledge work right now. Everyone can see that agents generate code that is “incredibly good and working,” yet well-run organizations stay very careful about details. They use peer-to-peer pressure to verify output and keep slop out. Your point is that this verification does two jobs. It filters quality, and it also builds an internal profile of each person that they can’t fake. To earn more trust from peers, and so the right to push things forward, a person has to show real understanding. If they don’t, a bad profile forms in people’s minds and damages their reputation. Your proposed survival skill is the ability to work on several things at once and still go into detail, while thinking at a higher level of abstraction. That’s how a person stays useful as AI becomes able to exceed human abilities in many areas.
This touches software engineering (code review), organizational behavior (trust and reputation), signaling theory and human–AI teaming. It matters now because generating output has become nearly free, so the costly step has moved to verification and attribution of understanding.
Key Insights
-
Agents can write impressive code, but that isn’t the same as code that ships in a real codebase. METR ran a randomized trial with 16 experienced open-source developers on 246 real issues. With early-2025 AI tools they took 19% longer, yet they believed AI had sped them up by 20%. METR suggests AI may do relatively worse “in settings with very high quality standards, or with many implicit requirements,” which are exactly the details that careful organizations enforce (METR, 2025). METR marks these results as outdated and has published a February 2026 update, which I couldn’t fetch to check. The more lasting lesson is the gap between how fast people feel and how fast they are. That gap is why outside peer checks matter: people judge their own AI-assisted work badly.
-
Slop is now an organizational cost. Niederhoffer, Hancock et al. call it “workslop” in HBR. They link it to a paradox: AI use at work has doubled since 2023, while an MIT Media Lab report found 95% of organizations see no measurable return (HBR, 2025). As I recall, the article also reports that people who receive workslop rate the sender as less capable and reliable. That would be direct evidence for your “bad profile” mechanism, but the fetched text was cut off before that section, so treat it as unconfirmed.
-
Peer verification only works when identity persists. Curl’s maintainer Daniel Stenberg reports that about 20% of 2025 security submissions were AI slop and only about 5% were genuine. Each report ties up 3–4 of the seven team members for 30 minutes to 3 hours. He also says plainly that “reputation doesn’t help” on HackerOne, because new users “can just create a new account next week” (Stenberg, “Death by a thousand slops”). This sharpens your claim. A profile “cannot be faked” only inside a group where people stick around and know each other, like a team or a long-running set of reviewers. Open, anonymous systems lack this, which is why commenters there suggested web-of-trust vouching, refundable fees, or screen recordings of a working reproduction. All of these add cost back in to stand in for a reputation the system can’t hold.
-
Code review was always partly about reputation and shared knowledge. Sadowski, Bacchelli et al. studied 9 million reviewed changes at Google. They found review motivations go beyond catching defects to include knowledge transfer and accountability (Modern Code Review: A Case Study at Google, ICSE SEIP 2018). Bacchelli & Bird’s earlier Microsoft study (ICSE 2013) found the same: the main outcome is understanding, not just bug-finding. So AI doesn’t create the need for peer checks. It removes the old proxy, “they wrote it, so they understand it,” and leaves the check for understanding as the main thing review does.
-
Why the profile is hard to fake: costly signals. Spence’s job-market signaling (1973) and Zahavi’s handicap principle apply here. A signal is believable when it’s expensive to produce without the real quality behind it. Polished code or prose used to be such a signal, and AI has made it cheap. Live behavior is still expensive: answering follow-up questions, defending trade-offs, spotting subtle edge cases. In Mayer, Davis & Schoorman’s trust model (1995), trust rests on ability, benevolence and integrity. Peer review mostly tests ability, but passing slop off as your own work also costs you on integrity.
-
Your multi-level abstraction skill matches observed changes in how people use AI. Lee, Sarkar et al. surveyed 319 knowledge workers (CHI 2025). They found GenAI shifts critical thinking toward “information verification, response integration, and task stewardship.” Higher confidence in GenAI predicted less critical thinking, while higher self-confidence predicted more (Microsoft Research, DOI). Being able to “go into detail while mentally working on a higher level abstraction” is stewardship. It also requires keeping enough self-confidence, grounded in real expertise, to keep checking.
-
Counterpoint: the multitasking claim needs care. Research on task switching (Monsell, 2003) and on “attention residue” (Leroy, 2009) shows real costs when people split attention. The skill you describe probably isn’t true parallel attention. It’s more likely fast, well-timed zooming between levels: knowing when a detail deserves a deep look. That makes it a judgment skill, not a bandwidth skill, and judgment is exactly what peer review tests.
-
A possible failure mode: reputation systems can be gamed or become cliquey. Trust profiles can harden into in-groups that shut out newcomers, a worry raised in the curl comments. They also reward looking like you understand. As AI gets better at producing convincing explanations, even live answers in review could be scripted. That would push verification toward higher-friction formats: in-person discussion, reproduction, and owning outcomes over time.
Open Questions
- If understanding becomes the scarce thing being verified, will organizations formalize the “internal profile,” for example as tracked records of who reliably catches what in review? And would making it explicit destroy the property that makes it hard to fake (Goodhart’s law)?
- When agents become reliable reviewers themselves, does human peer pressure keep its function, or does trust move from “this person understands” to “this person knows which checks to run and when to escalate”? In other words, does reputation shift from knowing the content to judging how verification should be done?
有趣的是,虽然我们主观上知道智能体能够生成非常优质且可运行的代码,但看到一些组织仍然极其注重细节,并利用同伴之间的压力来验证代码、避免低质量产出(slop),这一点非常有意思。这种同伴之间的检查还会形成一种内部档案,记录那些无法伪造的东西,并且必须展现出足够的理解水平,才能提升其他同伴的信任,从而推动事情向前发展;否则,人们心中就会形成一份糟糕的档案,进而损害声誉。在同时处理多件事情的同时,仍能深入细节,并在更高的抽象层面上进行思考,这似乎是在世界中继续航行的关键之一,尽管我们已经心知肚明人工智能在很大程度上已经超越人类的能力。
以下内容由 LLM 生成,可能包含不准确之处。
背景
你的笔记讨论的是当下软件与知识工作中的一种张力。大家都看到,智能体生成的代码"好得令人难以置信,而且能跑",但管理良好的组织仍然对细节非常谨慎。它们依靠同伴之间的压力来验证产出,把粗制滥造的东西挡在门外。你的观点是,这种验证承担两项职能:它既筛选质量,也在构建每个人的内部档案,而这份档案是无法伪造的。一个人要想赢得同伴更多的信任,从而获得推动事情前进的权利,就必须展现真正的理解。如果做不到,人们脑中就会形成一份糟糕的档案,损害其声誉。你提出的生存技能是:同时处理多件事情,既能深入细节,又能在更高的抽象层次上思考。随着AI在许多领域超越人类能力,这正是一个人保持价值的方式。
这一话题涉及软件工程(代码审查)、组织行为学(信任与声誉)、信号理论以及人机协作。它之所以在当下尤为重要,是因为生成产出几乎已经免费,因此高成本的环节转移到了验证和理解力的归因上。
关键洞见
-
智能体能写出令人印象深刻的代码,但这并不等同于能够在真实代码库中交付的代码。 METR 对 16 位经验丰富的开源开发者进行了随机对照试验,涉及 246 个真实问题。使用 2025 年初的 AI 工具时,他们的耗时多了 19%,但他们却相信 AI 让自己的速度提升了 20%。METR 认为,AI 可能在"质量标准极高,或存在大量隐性需求"的场景中表现相对较差,而这恰恰是谨慎的组织所要严格把关的细节(METR, 2025)。METR 已将这些结果标注为过时,并发布了 2026 年 2 月的更新,但我无法获取该更新来核实。更持久的教训在于人们"感觉"的速度与实际速度之间的差距。正是这一差距使外部同伴检查变得重要:人们对自己借助 AI 完成的工作往往判断失准。
-
粗制滥造的产出如今已成为组织成本。 Niederhoffer、Hancock 等人在《哈佛商业评论》(HBR)中将其称为"workslop"(工作垃圾)。他们把它与一个悖论联系起来:自 2023 年以来,职场中的 AI 使用量翻了一番,而麻省理工学院媒体实验室的一份报告发现,95% 的组织没有看到可衡量的回报(HBR, 2025)。据我所记,该文还报告称,收到工作垃圾的人会认为发送者能力更差、可靠性更低。这将为你所说的"糟糕档案"机制提供直接证据,但我获取的文本在相关部分之前就被截断了,因此请将其视为未经证实。
-
同伴验证只有在身份可持续时才有效。 curl 的维护者 Daniel Stenberg 报告称,2025 年约 20% 的安全报告提交是 AI 生成的垃圾,真正有价值的仅约 5%。每份报告都会占用团队七名成员中的 3–4 人,耗时 30 分钟至 3 小时。他还直言,在 HackerOne 上**“声誉没有任何帮助”**,因为新用户"下周就能再注册一个新账号"(Stenberg, “Death by a thousand slops”)。这使你的论断更加清晰。只有在人们长期留存、彼此相识的群体中(比如一个团队或长期运作的审查者群体),档案才"无法伪造"。开放、匿名的系统缺乏这一条件,因此那里的评论者提出了网状信任担保(web-of-trust vouching)、可退还的押金,或是可工作的复现过程的录屏等方案。这些方案都在增加成本,以替代系统本身无法维持的声誉。
-
代码审查从来都在一定程度上关乎声誉与共享知识。 Sadowski、Bacchelli 等人研究了 Google 的 900 万次审查变更,发现审查的动机不仅在于发现缺陷,还包括知识传递与问责(Modern Code Review: A Case Study at Google, ICSE SEIP 2018)。Bacchelli 与 Bird 早前在微软的研究(ICSE 2013)也得出了相同结论:审查的主要产出是理解,而不仅是发现缺陷。因此,AI 并没有创造同伴检查的需求,而是消除了旧有的替代指标,即"他们写的,所以他们理解",于是对理解力的检验便成为审查的核心。
-
为什么档案难以伪造:高成本信号。 Spence 1973 年提出的劳动力市场信号理论与 Zahavi 的让度原理(handicap principle)在此适用。只有当一个信号的生成成本高昂,且无法在缺乏真实质量支撑的情况下产出时,它才可信。精致的代码或文字过去就是这样的信号,而 AI 让它们变得廉价。而实时的行为仍然成本高昂:回答追问、为权衡辩护、发现微妙的边界情况。在 Mayer、Davis 与 Schoorman 的信任模型(1995)中,信任建立在能力、善意与正直之上。同伴审查主要检验能力,但将粗制滥造的产出冒充为自己的成果,也会损害你的正直形象。
-
你提出的多层抽象能力与人们使用 AI 的方式变化相吻合。 Lee、Sarkar 等人对 319 名知识工作者进行了调查(CHI 2025)。他们发现,生成式 AI 将批判性思维转向了**“信息验证、回应整合与任务统筹”**。对生成式 AI 的信心越高,批判性思维越少;而自信心越高,批判性思维则越多(Microsoft Research,DOI)。能够"深入细节,同时在脑中以更高的抽象层次工作",就是一种统筹能力。它同时也要求保持足够的自信心,而这种自信应建立在真实专长之上,才能支撑持续的检查。
-
反方观点:多任务的论断需要谨慎对待。 关于任务切换的研究(Monsell, 2003)以及关于"注意力残留"的研究(Leroy, 2009)表明,人们在分散注意力时会付出真实的代价。你所描述的技能很可能并非真正的并行注意力,而更像是在不同层次之间快速、适时地切换焦点:知道何时某个细节值得深入查看。这使它成为一种判断力,而非带宽能力;而判断力恰恰是同伴审查所检验的对象。
-
一种可能的失败模式:声誉系统可能被操纵,或变得排他。 信任档案可能固化为排斥新人的小圈子,这一担忧在 curl 的评论区中已被提出。它们还会奖励看起来像是理解了的人。随着 AI 越来越擅长生成令人信服的解释,即便是审查中的实时回答也可能被提前编排。这将促使验证转向更高摩擦的形式:面对面讨论、复现,以及长期对结果负责。
开放性问题
- 如果理解力成为被验证的稀缺对象,组织是否会将"内部档案"正式化,例如将谁在审查中可靠地发现了什么问题记录在案?而使其显性化是否会破坏使其难以伪造的特性(古德哈特定律,Goodhart’s law)?
- 当智能体自身成为可靠的审查者时,人类的同伴压力是否还能保持其功能?还是信任会从"这个人理解"转向"这个人知道该运行哪些检查、何时应当上报"?换言之,声誉是否会从"知道内容"转向"判断应当如何进行验证"?