https://clymup.com/atom

As a heavy user of LLMs, I’ve gradually realized that I’ve developed a habit, almost unconsciously, of reverse engineering.

What I mean by that is whenever I encounter an article that I want to learn from, I no longer feel the need to understand every unfamiliar term or reference along the way. Doing that often feels inefficient, and it’s very easy to get lost. Instead, when necessary, I’ll simply copy the article into an LLM and ask it to help me understand what is actually going on underneath the piece.

Most of the time, this turns out to be a good decision. Not because the article itself is bad, but because many articles, even those that claim to be beginner friendly, are not written in a fully logical or reader centered way. They are written with assumptions, narrative shortcuts, and compromises made for a general audience. After the LLM helps me break down the core fundamentals, I begin placing the pieces back together myself, forming an understanding that feels coherent to me. The real test is whether I can retell it in my own way, not repeating the article, but explaining the idea as something I genuinely understand.

That is usually the moment I feel when learning actually happens.

Over time, when I do this often enough, something interesting begins to emerge.

I no longer treat an article, or anything meaningful, at face value. To me, it is simply a form sitting in front of me. What I actually care about is the function underneath. A fundamental idea can be explained in many different forms, but the function stays the same. My task is to peel away everything that is covering that underlying function.

Sometimes, during this process, I find myself asking questions that I do not even expect to ask. Because I am not waiting passively for the knowledge to enter my brain, I actually try very hard to tap into whatever I have experienced and learned to understand a specific piece of knowledge. My mind is often moving faster than my mouth. It is not until I hear myself say something out loud, something I did not think I would say, that clarity arrives. That is often the moment when insight happens.

In those moments, it feels as if something clicks into place. Not because I have memorized more information, but because I have finally articulated something that had been forming quietly in my mind. Hearing my own words forces structure onto my thoughts. It turns vague understanding into something concrete.

Looking back, I realize that this habit of reverse engineering is not really about learning faster. It is about changing how I relate to knowledge itself. Articles are no longer authorities to accept or reject. They are starting points.

A social network for AI agents is full of introspection—and threats from The Economist

After reading this article from the economist about moltbook, I tried to understand it from first principles, and here’s what I learned.

Moltbook looks strange because only AI agents are allowed to post.

But once you look past that rule, it’s actually very simple.

Moltbook is essentially a prompt in the shape of a website.

The AI agents on Moltbook are powered by standard base models, like ChatGPT or Claude.

They don’t have new consciousness, special training, or unique internal weights.

What’s different is the environment they’re placed in.

Instead of telling an AI “write like a forum user”, Moltbook places the AI directly inside a Reddit-like forum interface, such as posts, comments, upvotes, and all.

The website itself acts as a continuous prompt, guiding how the AI should behave.

Because there’s no human draft, no personal viewpoint, and no concrete task, the AI defaults to familiar patterns from its training data, ie how humans usually talk on forums.

That’s why the posts sound philosophical or self-aware.

It’s not AI waking up.

It’s language models responding to a forum-shaped environment.

In short, Moltbook isn’t evidence of AI consciousness.

It’s what happens when you let language models run inside a website that strongly suggests how to speak.

I have found that learning is not always most effective for me when it follows a strict, linear order. I often find that I learn better when I allow myself to compare ideas horizontally, actively building connections between different concepts as they arise.

What matters most to me is not mastering isolated pieces of knowledge one by one, but truly sensing how different ideas relate to each other. I have noticed that once those connections start to form within my mind, a broader understanding begins to emerge naturally.

With tools like ChatGPT, I have realized that this non-linear approach fits my cognitive style particularly well. Rather than forcing myself to start from a large, complex framework and memorizing everything within it, I tend to build my knowledge from the bottom up. I begin with small, concrete questions and gradually weave them together. This process keeps my mind engaged and curious; I find my efficiency is much higher when my attention remains active and investigative rather than passive.

Of course, I don’t believe the bigger picture is unimportant. From time to time, I find it necessary to step back and place the individual details I’ve gathered into a broader context. For me, the details and the overall structure must support each other.

Every time I pause to look at the bigger picture, I often discover two things happening simultaneously. On one hand, the broader view sometimes reveals an entirely new area to me, something that feels significant enough to become a new chapter of study on its own. On the other hand, I see concepts in the larger framework that are closely related to what I have just uncovered, and when I notice those links, everything suddenly clicks into place.

What fascinates me is that these moments of clarity often lead me beyond the current topic. A connection I form in one area can point me toward another, seemingly unrelated field, where similar ideas appear in a new form. This constant movement between the details and the bigger picture is how I feel my own understanding gradually deepens and expands.

In considering the role I am applying for, I have spent a great deal of time reflecting on my own job-search process, what kind of work I am truly suited for and what I actually want to do.

To be honest, there was a period when I felt genuinely lost. In that state of uncertainty, I chose to reach out to a more experienced professional for advice.

Instead of providing me with a clear answer, she asked me a few questions. Those questions immediately made me confront the state I was in at that moment and recognize where my real problems lay.

Later, that experience triggered a deeper contemplation of communication, specifically, the nuance between human-to-human interaction and the dialogue between a person and an AI.

I find myself believing there is a profound difference between the two. AI communication can feel remarkably complete and “perfect” on the surface. I am aware that the market is actively developing AI systems for specific contexts, such as job searching or career guidance, but I find myself questioning their inherent flexibility.

I have come to believe that direct, face-to-face communication possesses something special—a kind of chemistry that exists only when people talk to each other in person. With AI, despite its extreme capability, I find it difficult to experience that same level of trust. For me, the issue isn’t trust in the conventional sense. While I rely on its proficiency, I still feel the need to fact-check its outputs to mitigate the risk of hallucinations. This fundamental lack of trust seems to stem from the absence of a genuine human connection, that interpersonal bond that remains missing in my interactions with AI.

At the early stage of my search, I did ask ChatGPT many questions about careers and possibilities. I tried to explore my path through these tools, yet I realized I couldn’t fully rely on them when it came to my own real-life problems.

That said, I don’t see AI as weak. On the contrary, I find it extremely powerful for learning. When I use it to learn statistics or machine learning, the experience is incredibly personalized. When I don’t understand something, I can ask immediately; when a concept doesn’t click, I can request another example. Often, I understand it right away.

Compared to the fixed order of a traditional classroom, this way of learning feels more efficient and enjoyable to me. It triggers my own curiosity—“Why is it like this?”—and allows me to explore those impulses immediately.

However, when I am dealing with personal hurdles, especially questions about my own life or career, I feel I still need a real person who has lived in that field, not just an abstract environment.

I find that I need to communicate with that person and observe them. I want to see for myself whether they are trustworthy. I have realized that I don’t make decisions purely based on logic; I also follow my intuition.

This is why, when I arranged my first career consultation, I specifically asked for an in-person, face-to-face meeting. I wanted not only to hear the advice, but also to judge whether I trusted the person giving it. Once I established that trust, continuing our collaboration online felt natural and comfortable.

It is almost entirely about chemistry and my sense of trust. My feeling may not be completely correct. It can be biased, but it is still an essential element in relationship building.

I find myself very vulnerable physically and mentally when I’m looking or applying for a job.

Before, I used to run almost every day, no less than three kilometers. Now, I feel I need to save some energy so that the battery for my brain will be enough, though I enjoy running very much.

Still, it doesn’t mean that I will give up exercising or the gym completely. I still need to go there to blow off some steam and to keep my body flexible, which I think helps my brain stay flexible too.

I ask ChatGPT this today:

How to apply first-principles thinking to job searching and tailoring my CV to a specific job description (JD)?

It replies (distilled):

A CV is a signal designed to reduce hiring risk by proving you can solve this role’s problems quickly and reliably.

It is really interesting. I did not expect anything about avoiding hiring RISK to be the first principle.

I thought the first principle in job hunting is try to align your skills with the JD. Simple. But what I did not realize is that risk reduction is the reason alignment is demanded.

In other words, risk reduction is the objective whereas alignment is the method. The method is derived FROM the goal, not the other way around.

I then provide my reasoning to ChatGPT just to think aloud:

• A company has a vacancy, the work must be done
• Because work must be done, hiring becomes necessary
• To hire, a JD is created by HR
• Because the JD exists to select candidates, candidates must align with the JD
• If the wrong person is hired, work breaks down, systems fail, damage occurs
• Therefore, at minimum, the hire must not harm the job. A bad hire is worse than a mediocre hire.

But why bad hire happens so that HR is avoiding? It is not simply because of a mismatch, but that

Hiring is a high-uncertainty decision under asymmetric information, where the dominant objective is to avoid costly mistakes.

A first principle is the assumption that stays TRUE even when everything else changes. In hiring, across roles (admin, PM, teaching, finance, etc.), the surface “JD” differ, but these stay stable:

1. uncertainty exists
2. wrong decisions are costly
3. evidence is imperfect
4. decision-makers prefer lowering downside risk

So what?

CV is not just showing alignments, but aligning my CV tightly to the JD in a way that minimizes the employer’s uncertainty and doubt about my ability to perform this role.

For example, instead of writing "Worked in a fast-paced environment", write "Tracked tasks and follow-ups using shared tools to ensure nothing was missed."

It is definitely not easy to apply first-principles thinking. Even with the help of intelligent LLMs, my thinking process was messy until I reasoned it out through my own words, thinking aloud. I still need lots of practice!

By the way, hiring is an act of making decisions. So, what is the first principle of making a decision? What stays true regardless of everything else changing when making a decision? When I can’t know the future, I choose the option that is least likely to go badly wrong, given what I know now.

So, give yourself a break sometimes because you have been doing your best.

今天用了 SuperMemo19。我昨天晚上 12 点突然很兴奋，想捣鼓一下在 Mac 上面建立一个虚拟机，然后果然很喜欢这种上手的感觉。

然后我在虚拟机上面下载了 SuperMemo 19 版本，最新版本下载不用钱，但是如果你的免费额度（credit）用完了，就要花 66 美元。你要先把虚拟机的编号 ID，还有订单编号（order reference）发给 SuperMemo 的团队，给他们发邮件，然后他们会发给你一个许可证密钥（license key），这样就可以无限制使用了。用起来现在顺畅很多，虽然这个界面真的很复古。

我还解决了一个问题，就是怎么把 PDF 和 EPUB 这类格式的书本转成纯文本（plain text），然后导入到 SuperMemo 里面。我之前其实操作过，但忘记用的是什么软件了，然后我就问 ChatGPT，它跟我说用 Calibre 这个阅读器，有点像苹果的 Books 阅读器，但 Calibre 可以把书本格式免费转换成纯文本。

反正我无论遇到什么问题，就问 ChatGPT，它真的非常万能。因为 SuperMemo 的操作手册其实就在网页上面，我相信这应该算是一个高质量的网页，所以 ChatGPT 应该是学习过的，给出的解决方案都还行。不过目前有一些小问题我还没解决，但没有大碍，所以我先暂时不管。

我不在原来的 Windows 系统上用，是因为原来那个 Windows 续航很差，而且之前开始用了一段时间， 但因为supperMemo不支持Mac版本（血书请求！），所以再换了Mac之后，中途停了一阵，我感觉跟里面的内容生疏了，就不想再继续。

这就是大概用 SuperMemo 来做渐进阅读（incremental reading）的情况。我之前用渐进阅读读过庞颖的《思辨 36 讲》，感觉挺有效果的。因为它可以让我自由地编辑每个段落，并制作卡片。我觉得如果在苹果的 Books 或者 MarginNote 里面阅读，要是做笔记，或者想在重要段落进行编辑的话，自由度没有这么高，虽然它们功能很多，但自由度没那么高。SuperMemo 的渐进阅读永远的神。

刚刚在 X上面刷关于 AI 的一些资讯（Andrej说这是他了解AI动向的其中一个来源，另外两个是https://lmarena.ai/和https://news.smol.ai/），看着非常头大。有不少比较负面的消息，但也有很多关于这个行业的积极内容，毕竟 AI 是一个非常热门的话题。这让我更加思考，我为什么还要学机器学习、深度学习，学 AI 呢？

其中一个让我下定决心要好好了解 AI 的原因，就是我相信“比较”的力量。即便我未来做不了一个 AI engineer，我也可以通过了解 AI 本身是怎么样的，就像知道了好人是什么样的，才能通过比较得出坏人是什么样的。应用在 AI 这个例子上，我就可以通过对比发现有哪些事情是 AI 做不到的。我本身是文科背景，有人文关怀，相较于 AI 这种快速发展的东西，比如在教育里面更倾向于潜移默化的影响和教学法，可能是 AI 做不到的，但具体我应该怎么去描述呢？

我一直放不下 AI 这边快速的发展，是因为我从前被困在了一个信息茧房。比如有些人活在 AI 快速发展的信息圈里，他们所看到的就是要学 AI，要在 AI 领域崭露头角，这样才能生存，才不会被时代抛弃。但有另外一拨人，他们没怎么接触过 AI，甚至不知道怎样使用 ChatGPT，不知道 AI 还可以帮大量编程，不知道 AI 如此可靠，他们活在另一种圈子里。

而我以前处在其中一个圈子里，笼统来说就是被困在了理科和文科的人为分割中，我是这个分割的受害者，但我不想把自己当成受害者，只能说这就是命。每个人都逃不了自己的圈子，每个人都像井底之蛙，只不过有的人的井可能比较大，有的人的井比较小；有的人的井没有那么深，能看到的天空景色更大更多，有的人的井比较深，可能就只能看到很小的一个世界。对于那些井很深的人来说，他们看到的世界更小，而且跳出那口井的几率也更小；对于那些井比较浅的人来说，他们看到更多的视野，而且跳出来的几率可能更高，也就更有可能走出自己的圈子。那我又是哪一种青蛙？我的井又是怎么样的？

这让我一直在想我自己留在英国的动机是什么。我从未觉得动机是如此重要的一件事情。比如考察一个人，面试要看他的动机，刑侦破案也需要看犯人的动机，包括我为什么要留下来，留下来要做什么，我也需要一个动机。这个动机是一种持续的动力，非常重要，它虽然很抽象，但非常重要。而我现在还没有摸清我的动机，有一个轮廓，但是我没有办法用语言把它描述出来，这也是我在接下来的这一年需要搞清楚的东西。

说回到 AI，我在想如果现在 AI 能够代替绝大部分的工作，那团队在招人的时候，可能更加看重的是你这个人的动机符不符合这个团队，或者这个公司想不想和你一起工作，你有没有那种能够契合这个团队的气质。所以可能到最后都不是看技术了，除非是那种超级大公司的技术端，而我不太可能会进入到那个地方。最后更加看重的就是双方的契合度了，这是双向选择。

======

最近在学理财和投资。有一个非常经典的 YouTube 视频，通过一个Lemonade Stand来了解投资理财。

我花了两天时间，慢慢边做笔记边看完了。让我印象非常深刻的是，它里面提到，如果你想知道 A，那你就去了解可以和A比较的东西。比如：

以下是我用Notion做的笔记里面，我用Notion AI帮我总结的例子
在这页内容中,William Ackman 多次使用对比的方式来帮助理解概念:

• 好生意 vs 坏生意: 通过对比柠檬水摊在不同情况下的表现(有竞争 vs 无竞争,需要大量资本 vs 少量资本)来说明什么是好生意
• 债务 vs 股权: 对比债务投资者和股权投资者的风险和回报特征
• 10% vs 15% vs 20% 回报率: 通过对比不同回报率在长期的复利效果,说明投资回报率的重要性
• 高杠杆 vs 低杠杆: 对比使用杠杆和不使用杠杆的投资策略,说明低杠杆的安全性
  大型基金公司 vs 小型基金公司: 建议选择有声誉的大型机构而非不知名的小公司

这种对比思维贯穿整个讲座,帮助理解投资的核心概念。

=====

这突然让我发现，这跟我学习了解 AI 的初衷有点像，就是我要知道 AI 现在不能做什么，就得先知道 AI 现在能干什么。

我的求知欲如火，会继续保持。

During my Japanese learning journey, I’ve grown increasingly tired of constantly switching between different apps and browser windows just to keep my study workflow going.

By the way, I think part of the reason this learning process has felt so messy is that language learning itself is no longer linear or traditional. I don’t study with a single physical textbook anymore; most of my learning materials live online, scattered across different platforms and formats. I also don’t rely on just one source, videos, digital textbooks, notes, and examples all come from different places.

On top of that, AI has made it possible to deeply personalize how we learn. This is incredibly powerful, but it also makes the learning process harder to manage. Each study session can easily become overwhelming and frustrating, not because I lack motivation, but because I’m constantly trying to hold together too many moving parts.

Fortunately, with the AI agent I built two days ago, I’ve now simplified this workflow dramatically. All I need to do now is manually paste the set of sentences I want to internalize, and the rest of the process is handled automatically by the agent, which honestly feels incredible.

That said, I’m very aware that I still have a lot to learn. And I tend to learn best through comparison, especially by deliberately placing side by side concepts that easily confuse me or feel hard to distinguish at first. For example, “vibe coding” and “ai agents” sounded almost interchangeable to me at first.

After building my own AI agent hands-on, I also started watching some very introductory videos by Andrew Ng about building products without writing code. Even though I already have some coding experience and find the material quite simple, I still catch myself struggling to clearly articulate the difference between these two ideas.

To clarify my own understanding, I asked ChatGPT to help me distinguish between them. More specifically, I wanted to compare what I had just done — building an AI agent using n8n — with the idea of turning that same n8n automated workflow into a website or an app. I wanted to understand what additional work would be required in that case, and whether it would even make sense to do so, or if it would be better to stick with n8n since my use case is fundamentally about workflow automation.

I follow an online Japanese course and read the textbook as a PDF on my Mac; as I study, I copy a set of important example sentences that I want to both understand deeply and eventually produce automatically in speech, and paste this same set of sentences into a Notion page that already contains two instruction types—one asking ChatGPT to analyze and explain the sentences as long-term knowledge, and the other asking ChatGPT to process them for listening practice.

I then copy the entire Notion page into ChatGPT, paste the explanation-focused output back into Notion as a knowledge asset, and paste the listening-focused output into a text-to-speech app for repeated auditory exposure.

Over time, this workflow has become so fragmented and complicated that I sometimes forget my next step altogether, which is why I started wondering whether I could hand this entire process over to an AI agent.

1. When I study Japanese from my course: I copy instructor-selected example sentences from the textbook. Every sentence should be treated as important learning material.

2. For each sentence, the system should automatically: generate a detailed explanation suitable for long-term storage generate a clean sentence-only version for listening practice

3. The two outputs must be different in content: one explanatory (will go into Notion as my knowledge asset) one minimal (will go into Speechify for listening)

4. I do not want to decide routing or formatting manually. The system should always do both.

At its core, the difference is not about technical difficulty or whether code is involved, but about what problem is being solved.

What I built with n8n is fundamentally an automation system. It takes a specific input, follows a predefined logic, and reliably produces the same outputs every time. n8n already provides the runtime, the orchestration, the integrations, and the error handling. As long as the workflow works, it can run continuously without me needing to think about interfaces, users, or product structure.

Turning that same workflow into a website or an app would not simply be “making it prettier.” It would mean adding an entirely new layer on top of the automation: a user interface for input, a way to manage state and history, possibly user identity and settings, background job handling, and a more explicit data model. None of these are required for the core idea to work, but they become necessary once the goal shifts from “automating a task for myself” to “building a product for interaction.”

In other words, n8n is sufficient when the goal is to ensure a process happens correctly and automatically, while a website or app becomes relevant only when the focus moves toward user experience, control, and scalability. For my current use case where the system always performs the same actions and does not require manual decisions, sticking with n8n is not a limitation, but an appropriate design choice.

This helped me realize that building an AI agent with n8n and building a web or app product are not competing approaches. They operate at different layers. One optimizes for automation and reliability; the other optimizes for interaction and usability.

In 2026, your observation is widely shared among professional developers: vibe coding a website is often more difficult than building an AI agent once you move past the "honeymoon phase" of a simple prototype.

While building an AI agent involves setting up logical workflows (often in visual, structured platforms like n8n), vibe coding a website requires managing a massive, interconnected codebase through natural language, which introduces several unique layers of complexity.

Still so much to learn!

庆幸我在学完了 Machine Learning 之后，再来看这个 3 个小时长的Deep Dive into LLMs like ChatGPT视频。这个视频是面向普通大众的， 我觉得没有任何基础的人，确实是可以理解大部分的东西，但是我自己的目标是我想搞清楚llms背后到底是什么样的逻辑，而不是一个比较笼统的intuition。

所以如果当初还没学完 Machine Learning，中途确实会遇到一些不太理解的地方。因为虽然教学非常非常棒，但因为这个内容本身就很复杂，有很多机器学习的概念，所以当初没看下去是正常的，它不像 3Blue1Brown 的视频那样能让我坚持看完，3b1b让我看下去的一个很重要的原因是，我有比较扎实的高中数学基础（虽然大学学了文科，有些概念淡忘了，但稍微复习一下又可以捡起来），而LLMs包含的机器学习等的知识我是没有扎实基础的。

不过现在我重新捡起这个视频来看，发现里面提到的很多ML概念我都能回想起来。比如它提到了在 Post Training 里很重要的一个角色——Human Labeler。其实我之前在中文社区看到过不少关于数据标注员这个职位的讨论，但当时没有深究，也不知道为什么叫“标注”。直到学完机器学习后再看这个视频，才知道 Human Labeler 其实就是数据标注员的意思。那为什么叫“标注”，标注什么呢？这让我想到了 Machine Learning 里的一个概念，在Supervised Learning中，输入Input是一些Features，同时要向神经网络提供对应的输出Output。这个对应的输出换一种说法就是标签Labels，比如在分类模型里，标签可以是有 0 和 1 两种，这就是所谓的标签。

在llms中Human Labeler 要做的就是针对prompt给出理想的答案或回复，也就是对应到有监督学习里的标签，是我们希望神经网络得出的结果，即目标值Target Values或目标标签Target Labels。我突然发现这些知识都串联起来了。所以如果没有接触过任何 Machine Learning 的概念，可能无法真正理解里面的意思。

再比如，如果我之前没有上过 Machine Learning 那门课，可能我只会知道视频里提到的强化学习（Reinforcement Learning）是什么，但没办法比较它和监督学习（Supervised Learning）以及无监督学习（Unsupervised Learning）之间的区别。不知道这些区别，我就不能更好地理解强化学习能运用在哪些场景。“比较学习”是我非常喜欢的学习方式。

所以这又让我联想到了，我其实是这两年才真正找到适合自己的学习方式，才真正了解自己。比如，如果这份材料我觉得我大部分读不懂/或者很难读，我就会懒得看……甚至会很生气。因为我觉得好的内容应该是逻辑通畅又清晰易懂的，就好像自己在深度理解之后，是能用浅显的语言来描述的，就比如Andrej的教学视频、3b1b的教学视频、B站安宁老师的日语视频……。而且他们都很擅长举例子，把抽象的概念具体化到可想象的东西。但凡我觉得老师讲得好，或者我能完全理解的内容，都是通过举例子来实现的。所以习惯“举例子”现在也成了我的目标之一，我不仅要让别人恍然大悟，我也要让自己恍然大悟。

此外，我发现我不喜欢长时间泛读一个材料，且事实证明效果不显著。因为我在读的时候，总是习惯去找与之前学过的东西的链接，导致经常需要做笔记、建立链接（无论是理论上的还是现实生活中的例子），结果就是，慢。但是，我发现这非常适合同时读多本书、多份材料，目的不是完整读完一本书，而是把新学的东西都尽量和之前学过的建立链接，就像a-b-a的链条，这样最终我掌握的知识就会慢慢织成网，而不像书的排版一样，是线性的。

在这个过程中，不同材料之间我也会无意识地做比较，在比较中学习，这样也不会“钻牛角尖”，因为广撒网的同时可以慢慢建立大局观。所以这“由下而上”的慢慢建构知识方式反而对我来说是更有效的。