AI as Cognitive Infrastructure
AI as Cognitive Infrastructure
AI 作为认知基础设施
Most teams are chasing a better model. I went the other way and built a better system around the model: a 21-role AI setup that persists context, plays defined roles, and pushes back when I’m avoiding the work. This is how it works, and why the scaffolding ended up mattering more than the model itself. 大多数团队都在追求更好的模型。我则反其道而行之，围绕模型构建了一套更好的系统：一个包含 21 个角色的 AI 设置，它能够保持上下文、扮演特定角色，并在我逃避工作时给予反馈。这就是它的运作方式，也是为什么最终脚手架（系统架构）比模型本身更重要的原因。
I’d planned to ship a homelab piece. Hit some blockers (separate post coming), and the energy went where it could actually land: meta. So I pivoted to refining how I work with Claude on these writings. The framework that came out of that pivot is what this article is about. It’s going to drive my homelab work next, then expand into the agent-network build I want to deploy on local infrastructure. The same rigor will apply to finances and investments, health and fitness, home improvement, operations and maintenance, and pretty much everything else I touch as a long-arc project. 我原计划发布一篇关于家庭实验室（homelab）的文章。但遇到了一些阻碍（后续会另发文章说明），于是我将精力转向了更实际的地方：元认知。因此，我转而优化与 Claude 在写作上的协作方式。这次转型所产生的框架正是本文的主题。它将推动我接下来的家庭实验室工作，随后扩展到我希望部署在本地基础设施上的智能体网络构建中。同样的严谨性也将应用于财务与投资、健康与健身、家居改造、运营与维护，以及我作为长期项目所涉及的几乎所有其他领域。
I. The frame
一、框架
The problem with how most people use AI 大多数人使用 AI 的问题所在
Most knowledge workers treat AI like an expensive search engine with better grammar. Ask a question, get an answer, move on. Each conversation starts cold. Each output is judged on its own. The system never gets smarter about the person using it, the work they’re doing, or the patterns in how they get stuck. That’s not a use of AI. That’s a use of autocomplete. 大多数知识工作者把 AI 当作一个语法更好的昂贵搜索引擎。问一个问题，得到一个答案，然后继续。每次对话都是从零开始。每个输出都被单独评估。系统永远不会变得更了解使用者、他们正在做的工作，或者他们陷入困境的模式。那不是在使用 AI，那是在使用自动补全。
What “cognitive infrastructure” means here 此处“认知基础设施”的含义
The shift: stop treating AI as a tool you call when you need something. Start treating it as a substrate the work runs on top of. The substrate persists, accumulates context, plays defined roles, gates its own actions appropriately, and pushes back on you when you’re avoiding the thing you should be doing. This requires building the substrate deliberately. It does not happen by accident, no matter how good the underlying model is. 转变在于：停止将 AI 视为需要时才调用的工具，开始将其视为工作运行的基底。这个基底能够持久存在、积累上下文、扮演明确的角色、适当地控制自身行为，并在你逃避本该做的事情时给予反馈。这需要刻意地构建基底。无论底层模型有多好，它都不会偶然发生。
The neurodiversity (ND) lens (why this matters more for some people) 神经多样性（ND）视角（为什么这对某些人更重要）
I have severe ADHD and I’m on the autism spectrum. “Remember to do X on a schedule” is a guaranteed failure mode for me. So is “stay in the right context window of attention across a multi-week writing project.” So is “consistently apply the same set of principles to every draft when energy is low.” The same things that make this hard for ND folks make AI most powerful for us, IF the system is built right. Offload everything that’s repeatable, time-critical, or tedious. Keep the writing, the voice, and the key decisions for the human. The whole point of the system is to make the parts that don’t need a human, not require one. This isn’t ND-only. Anyone over-committed enough to need this is going to benefit. ND professionals just feel the cost of NOT having this most acutely. 我患有严重的 ADHD（注意力缺陷多动障碍）并且处于自闭症谱系中。“按计划记住做某事”对我来说是注定失败的模式。在为期数周的写作项目中“保持正确的注意力上下文窗口”也是如此。在精力不足时“始终如一地将同一套原则应用于每一份草稿”同样困难。正是这些让神经多样性人群感到困难的事情，使得 AI 对我们来说最为强大——前提是系统构建得当。将所有可重复、有时效性或繁琐的任务卸载出去。将写作、语调和关键决策留给人类。系统的全部意义在于让那些不需要人类参与的部分，真正不再需要人类。这不仅适用于神经多样性人群。任何因过度投入而需要此系统的人都会受益。只是神经多样性专业人士对缺乏这种系统所带来的代价感受最为深刻。
Who this is for (and the honest pitch) 受众群体（以及坦诚的推介）
I’ve been doing AI-assisted writing for a while. Long enough to know what I’m good at and what I’m not. I am good at the writing itself: the voice, the lived-experience hooks, the technical depth, the judgment calls on what matters. I am not good at being my own developmental editor at 11pm on a Tuesday after a full day of senior architect work. I am not good at consistently applying an audience-proxy read to my own drafts. I am not good at remembering to run a quarterly review of whether my publication strategy is still coherent. The professional version of solving this is hiring people. A developmental editor, a brand strategist, an analytics consultant, a community manager, all on retainer. That’s how the professionals do it. It works. It also costs more than I make from this work, by a margin that does not get closer over time. I cannot afford that. Not as someone publishing on a free Substack with no revenue, and not as someone with a full-time job and finite evenings. That is what AI is for. AI democratizes access to expertise and process improvement. Not perfectly. Not effortlessly. But meaningfully, once you build the substrate underneath it. 我进行 AI 辅助写作已经有一段时间了。足够长的时间让我知道自己擅长什么，不擅长什么。我擅长写作本身：语调、生活经验的切入点、技术深度以及对重要事项的判断。我不擅长在周二晚上结束了一整天的高级架构师工作后，还要担任自己的发展编辑。我不擅长始终如一地以读者视角审视自己的草稿。我不擅长记住进行季度回顾，以检查我的发布策略是否仍然连贯。解决这个问题的专业做法是雇人。聘请发展编辑、品牌策略师、分析顾问、社区经理，全部按月付费。专业人士就是这样做的。这很有效。但它的成本超过了我从这项工作中获得的收入，而且差距并没有随着时间推移而缩小。我负担不起。作为一个在免费 Substack 上发布内容且没有收入的人，以及一个有全职工作、晚上时间有限的人，我负担不起。这就是 AI 的用途。AI 让专业知识和流程改进变得平民化。虽然不完美，也不轻松，但一旦你构建了底层的基底，它就具有了深远的意义。
Let me be real about who I’m displacing here: nobody. I was not going to hire a developmental editor for my Substack. I was not going to retain a brand consultant. I couldn’t afford them, and I wouldn’t have. AI isn’t taking business from those professionals in my case… it’s letting someone who never could have been their customer get a useful shadow of what they offer. I don’t want to minimize what AI is doing to people who do this work for a living. AI will negatively impact all of us in some form. But if you play your cards right, you can ride the AI wave to make yourself a better person, a more capable operator, a more thoughtful writer. That’s all any of us have control over. The framework is for writing here, but it generalizes. If you’re a knowledge worker, a creative, a small-business operator, anyone with a long-arc project on top of finite time and energy, this gives you a starting point. The specific roles are mine; the structure is general. 让我坦诚地说，我在这里取代了谁：没有人。我本来就不会为我的 Substack 雇佣发展编辑。我本来就不会聘请品牌顾问。我负担不起，也不会那样做。在我的案例中，AI 并没有抢走这些专业人士的生意……它只是让一个原本永远不可能成为他们客户的人，获得了一种有用的、类似于他们所提供的服务。我不想淡化 AI 对那些以此为生的人的影响。AI 会以某种形式对我们所有人产生负面影响。但如果你策略得当，你可以驾驭 AI 的浪潮，让自己成为一个更好的人、一个更有能力的运营者、一个更深思熟虑的写作者。这就是我们每个人所能控制的一切。这个框架虽然是针对写作的，但它可以推广。如果你是一名知识工作者、创意人员、小企业经营者，或者任何在有限的时间和精力下进行长期项目的人，这为你提供了一个起点。具体的角色是我的，但结构是通用的。
Why I think this matters now 为什么我认为现在这很重要
LLMs are starting to plateau. The capability curve from GPT-3 to GPT-4 was vertical. The curve since has not been. And let’s be real about what these things actually are: prediction engines trained on a large corpus of human knowledge. As George Carlin put it: “Think of how stupid the average person is, and realize half of them are stupider than that.” That’s, broadly, the corpus an LLM is the statistical aggregate of. That’s not nothing. It’s also not the brilliant collaborator the marketing suggests it is. Which means the trick is not waiting for the model to get smart enough. The model is what the model is. The trick is to build the system around it that provides guardrails and curates the expertise you actually want to tap into. Twenty-one roles, each grounded in a real discipline with real peer experts, with explicit gating bars and anti-patterns… that is the curation. The model is the powerful core capability. The system is what turns that capability into something that compounds. LLMs are not going to become AGI on their own. They will keep getting incrementally better, incrementally cheaper, and incrementally better integrated. 大语言模型（LLM）开始进入平台期。从 GPT-3 到 GPT-4 的能力曲线是垂直上升的。但此后的曲线并非如此。让我们认清这些东西到底是什么：在大量人类知识语料库上训练的预测引擎。正如乔治·卡林所言：“想想普通人有多蠢，然后意识到有一半人比那更蠢。”这大致就是 LLM 进行统计汇总的语料库。这并非一无是处，但它也不是营销所暗示的那种天才合作者。这意味着诀窍不在于等待模型变得足够聪明。模型就是模型。诀窍在于围绕它构建系统，提供护栏并筛选你真正想要利用的专业知识。二十一个角色，每一个都扎根于真实的学科，有真实的同行专家，有明确的准入门槛和反模式……这就是筛选。模型是强大的核心能力。系统则是将这种能力转化为复利的东西。LLM 不会自行演变成 AGI（通用人工智能）。它们将继续逐步改进、逐步降低成本，并逐步实现更好的集成。