<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Hong Chu on Medium]]></title>
        <description><![CDATA[Stories by Hong Chu on Medium]]></description>
        <link>https://medium.com/@yeesun.chu?source=rss-8151f2116118------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*dmbNkD5D-u45r44go_cf0g.png</url>
            <title>Stories by Hong Chu on Medium</title>
            <link>https://medium.com/@yeesun.chu?source=rss-8151f2116118------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 06:26:05 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@yeesun.chu/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Loop Engineering Is 70 Years Old]]></title>
            <link>https://medium.com/@yeesun.chu/loop-engineering-is-70-years-old-dc297438923b?source=rss-8151f2116118------2</link>
            <guid isPermaLink="false">https://medium.com/p/dc297438923b</guid>
            <dc:creator><![CDATA[Hong Chu]]></dc:creator>
            <pubDate>Wed, 22 Jul 2026 02:00:40 GMT</pubDate>
            <atom:updated>2026-07-22T02:00:40.002Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*92s2HCDWvWb_zdCGk2Fg5Q.png" /></figure><p><em>The hottest idea in AI right now was written down in 1954 by a rocket scientist.</em></p><p>There’s a new buzzword going around: <strong>loop engineering</strong>. The pitch is simple. Stop prompting your coding agent, and start building the system that prompts it for you. Give it a goal, a way to check its own work, a retry policy, and a stop condition. Then walk away.</p><p>I like the idea. I just can’t call it new.</p><h3>The evolution of AI engineering</h3><p>From 2022 to 2026, the whole story of working with LLMs goes from Prompt Engineering → Context Engineering → Harness Engineering → Loop Engineering.</p><p>The driver of this evolution is improving model capability: each jump removes the old bottleneck and exposes a new one further out.</p><ul><li><strong>Prompt → Context:</strong> Early models were one-shot and forgetful, so wording was the bottleneck. Once context windows grew, what mattered was what the model sees, not what you say.</li><li><strong>Context → Harness:</strong> Models could now act, not just answer. But they had no environment to act in. Tools, execution, feedback, and memory became the bottleneck.</li><li><strong>Harness → Loop:</strong> Agents could work for hours, but a human still reviewed every step and re-prompted. Human attention became the bottleneck. So we design the loop (verify, retry, stop) and step out of it.</li></ul><p>AI is new, and these buzzwords sound even newer. But nothing is new under the sun. One thing that fascinates me is finding the deep correlations between AI and ideas that existed long before it. And loop engineering hides a beautiful one: a whole field, seventy years old, that studied exactly this problem.</p><h3>What AI engineering is really about</h3><p>Strip away the buzzwords, and the essence of AI engineering is this: <strong>getting reliable behavior out of unreliable parts.</strong> The model is probabilistic. It will be wrong, sometimes confidently. The world it acts in is messy and keeps changing. Yet we want the system as a whole to be dependable. Every stage of the evolution above is a different answer to the same underlying question: how do you build certainty out of uncertainty?</p><p>Here’s the fascinating part: there is already a science of exactly this question, and it’s seventy years old. It’s called <strong>cybernetics</strong>, from the Greek <em>kybernetes</em>, the helmsman who steers a ship. In 1954, Qian Xuesen (钱学森), a rocket scientist at Caltech, distilled it into a book called <em>Engineering Cybernetics</em>. Its subject, in one line: how to design machines that steer themselves.</p><p>The core intuition is the helmsman himself. He doesn’t calculate the perfect rudder angle that will hold the course against every current and gust. Nobody can. He looks at where the ship is drifting, compares it with where it should be going, and nudges the rudder against the difference. Then again. Then again. <strong>You can’t compute the perfect action in advance, and you don’t need to. Correction beats perfection.</strong></p><p>That is the relationship between the two fields: AI engineering is cybernetics rediscovered. A thermostat holds your room within a degree while knowing nothing about weather, because the feedback loop, not the parts, carries the reliability. Loop engineering makes the same bet with a fallible model: accept that the agent will be wrong, and build the machinery that notices and corrects. Prompt engineering was the opposite bet, the “compute the perfect action in advance” approach. It lost, exactly as cybernetics would have predicted.</p><h3>Same loop, new names</h3><p>Put the 1954 loop and the 2026 loop side by side and it’s a little eerie. The setpoint became the loop contract. The controller became the agent. The plant became your codebase and tools. The sensor became the verifier, your tests and linters. The error signal became the failing output you feed back into context.</p><p>Even the slogans match. Loop engineers say “the verifier is the bottleneck, not the model.” Control engineers have been saying the same thing for decades: you can only control what you can measure. A bad sensor ruins the best controller, and a weak test suite ruins the best agent.</p><h3>Yes, gradient descent too</h3><p>If “correction beats perfection” reminds you of gradient descent, good instinct. Gradient descent <em>is</em> this loop, written as math. Measure the error, find the direction that shrinks it, take a small step, repeat. Nobody solves for the perfect weights in advance; a million small corrections find them.</p><p>And the mapping pays off in both directions. Learning rate? That’s how hard you turn the rudder. Too high, and you overshoot and oscillate, like a hotel shower swinging scalding-freezing-scalding while you frantically overcorrect. Momentum? That’s smoothing out noisy readings before reacting to them. Old ideas, wearing new clothes.</p><h3>Why bother with a 70-year-old book</h3><p>Because the old field already hit the problems the new one is just discovering.</p><p>Cybernetics didn’t stop at “feedback is good.” Its hardest-won lesson is that feedback can also <em>destroy</em>. Correct too aggressively, or react to stale information, and every fix overshoots. The system swings harder instead of settling. The field spent decades working out how much to correct, when to wait, and how to tell whether a loop will calm down or tear itself apart.</p><p>Sound familiar? Anyone who’s watched an agent “fix” a bug, break something else, fix that, and re-break the first thing has seen a loop oscillate. A control engineer would diagnose it on sight (too much gain, too much delay, no damping) and would have knobs to turn. Loop engineering doesn’t have that vocabulary yet. It’s all vibes and retry counts. In a real sense, it’s cybernetics waiting for its math.</p><p>We even know where this road goes, because aviation already walked it: stick-and-rudder, then instruments, then autopilot, then flight management. Each step moved the pilot one loop outward, from operator to supervisor. That’s exactly what’s happening to us.</p><h3>Old ideas, new bottleneck</h3><p>So here’s my actual recommendation. If you’re building agent systems, spend a weekend with cybernetics. Not for the trivia, but because it spent seventy years on the questions we’re about to spend the next five on: how much to correct, how to handle delay, how to work with noisy measurements, how to know when a loop will settle.</p><p>And if the pattern holds and the bottleneck keeps moving outward, ask yourself: what gets scarce once the loops run reliably? My bet is that the next discipline won’t be about running one loop. It’ll be about governing a fleet of them.</p><p>The helmsman never went away. We just keep rebuilding his job, one level higher.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=dc297438923b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[How I Passed the Claude Certified Architect Foundations (CCA-F) Exam in 2 Weeks]]></title>
            <link>https://medium.com/@yeesun.chu/how-i-passed-the-claude-certified-architect-foundations-cca-f-exam-in-2-weeks-6b967e6effb4?source=rss-8151f2116118------2</link>
            <guid isPermaLink="false">https://medium.com/p/6b967e6effb4</guid>
            <category><![CDATA[claude-architect]]></category>
            <category><![CDATA[llm-wiki]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[claude]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Hong Chu]]></dc:creator>
            <pubDate>Fri, 12 Jun 2026 07:45:17 GMT</pubDate>
            <atom:updated>2026-06-12T07:45:17.431Z</atom:updated>
            <content:encoded><![CDATA[<p>I just passed the <strong>Claude Certified Architect Foundations (CCA-F)</strong> exam with a score of <strong>843/1000</strong>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*L1wikAxC_f3grddlUlOpYw.jpeg" /></figure><p>I’m honestly very happy about it. Not only because I passed, but because the preparation process helped me build a much clearer mental framework for using Claude and thinking about AI-native workflows.</p><p>Before preparing for the exam, I was already using Claude almost every day, but had never learned it systematically. During the preparation process, I started to understand more deeply what the best practices are, why they matter, and how they can be applied to real product and business problems.</p><p>For me, this exam was not just about certification. It was part of my process of becoming more AI-native.</p><h3>Why I Wanted to Take This Exam</h3><p>My background is in Quant Trading, Data Science, and Product Analytics. I do have a master’s degree in Computer Science, but I’m not a traditional software engineer who writes production code every day.</p><p>My main interest has always been how AI can be applied to real business and product problems. And over the past year, I’ve been thinking a lot about what the next generation of <a href="https://medium.com/@yeesun.chu/if-ai-can-analyze-data-what-are-data-scientists-for-50671a6c0b0c">Data Scientists in the AI era</a>.</p><p>As AI becomes more powerful, I believe AI-native data scientists will need to go beyond coding, dashboards, experiments, and models. We will also need to understand how to design workflows, use agents, connect tools, manage context, and build systems that can solve problems reliably.</p><p>That was one of the main reasons I wanted to take the exam. I wanted a structured way to test and improve my understanding.</p><p>In total, I spent around <strong>60 hours over two weeks</strong> preparing for it.</p><h3>The Beginning Was Messy</h3><p>At the beginning, I felt a little lost.</p><p>Some concepts were familiar, but many details were not obvious to me. For example, I understood prompting, but I did not always know when a prompt was enough and when it was not. I understood tools, but I did not always think clearly about where safety logic should live. I understood agents, but I had to learn how to think about coordination, delegation, and context passing more systematically.</p><p>A lot of practice questions looked similar, but the answers depended on small wording differences.</p><p>I asked myself a couple of times:</p><blockquote>Why are these two questions almost the same but the answers are different?</blockquote><p>Then I realized I had not fully understood the underlying principles yet. That was when I changed my approach.</p><p>Instead of only asking, “What is the correct answer?” I started asking:</p><blockquote>Why is this the correct layer to solve the problem?</blockquote><p>That changed everything.</p><h3>The Biggest Lesson: Choose the Right Layer</h3><p>Before preparing for this exam, I often thought about AI workflow problems from the model’s perspective:</p><blockquote>How can I make the model behave better?</blockquote><p>After preparing, I started thinking more architecturally:</p><blockquote>What kind of failure is this, and where should it be fixed?</blockquote><p>That was probably the biggest mindset shift for me.</p><p>A few principles became especially important:</p><ol><li><strong>Prompts guide. Systems enforce.</strong></li></ol><p>Prompting is powerful when the problem is about interpretation, reasoning style, tone, or task framing. But prompting is not enough when the problem requires strict structure, exact labels, access control, or safety enforcement.</p><p>If something must be reliable, safe, or deterministic, it usually needs to be enforced by the system, not just described in the prompt.</p><p><strong>2. Criteria explain the rule. Examples teach the boundary.</strong></p><p>I liked this idea because it feels very practical.</p><p>In real work, humans also learn this way. We do not only learn from definitions. We learn from examples, edge cases, and feedback.</p><p>The same applies to AI workflows. If the model is inconsistent in judgment, examples often help more than another abstract rule.</p><p><strong>3. Don’t just compress context. Preserve decision-critical context.</strong></p><p>A good summary is not just shorter text.</p><p>It preserves the information that matters for future decisions, including goals, constraints, facts, assumptions, dates, sources, open questions, and decisions already made.</p><p>Otherwise, AI may sound coherent but make worse decisions in long context.</p><p><strong>4. Put the instruction where it naturally belongs.</strong></p><p>Claude Code has many configuration details: project files, rules, skills, commands, settings, tools, hooks, and more.</p><p>At first, these felt like a lot of separate concepts. Later, I realized the core idea is simple: different instructions belong in different places.</p><p>The more upstream you can prevent the problem, the better.</p><h3>How I prepared</h3><p>When I learn something new, I usually like to spend a meaningful amount of time researching before jumping directly into practice (typically 1/3, and 2/3 for complicated problems).</p><p><strong>So my first step was to ask my AI agent to search for other people’s experience and feedback about the exam.</strong> Reddit was surprisingly helpful for this. It gave me a better sense of what topics were important, where people got confused, what to be careful about, and where to find more practice questions.</p><p>I’ll list some of the most useful links at the end of this section.</p><p><strong>The second step was to start from the official materials.</strong></p><p>I think this is important. Community posts and practice questions are helpful, but the official exam guide should still be the anchor.</p><p>When I first read the official exam guide, it felt a bit abstract. So I put it into AI and asked it to explain each task more intuitively, with examples. That helped me turn abstract exam objectives into something I could actually understand.</p><blockquote>For this exam, understanding matters much more than memorization.</blockquote><p><strong>The third step was iteration.</strong></p><p>It took me about a week to study the official exam guide. Then I started doing practice exams.</p><p>The first practice exam did not go well. I used one of the easier practice sets and only got <strong>76.7%</strong>. That was a useful wake-up call.</p><p>After that, I screenshotted my wrong answers and asked AI to help me analyze what I had missed. For each practice exam, I tried to understand not only why the correct answer was correct, but also why my chosen answer was tempting but wrong.</p><p>After every one or two practice exams, I created a thinner version of my notes (8 practice exams and 4 versions in total). I removed things I already understood and kept only the important comparisons, confusing concepts, and recurring mistakes.</p><p>Another fun tool I used was my own implementation of <strong>Karpathy’s LLM Wiki</strong> (<a href="https://hong-chu.github.io/llm-wiki-starter/">what is LLM-Wiki</a>) for studying. I created an <a href="https://github.com/hong-chu/claude-certified-architect-foundations">LLM Wiki for this exam</a>, and every time I asked questions, the Wiki became a little smarter. It helped me build a personalized knowledge base instead of relying only on static notes.</p><h4>Resources I found helpful</h4><p><strong>Reddit discussions:</strong></p><ul><li><a href="https://www.reddit.com/r/ClaudeAI/comments/1to0xfc/just_passed_the_new_claude_certified_architect/">A very helpful discussion of someone who passed at 985/1000!</a></li><li><a href="https://www.reddit.com/r/claudeskills/comments/1t4ko4y/passed_the_claude_certified_architect_foundation/">Another discussion about which practice exams are the most helpful</a></li></ul><p><strong>Materials:</strong></p><ul><li>Please always read the <a href="https://everpath-course-content.s3-accelerate.amazonaws.com/instructor%2F8lsy243ftffjjy1cx9lm3o2bw%2Fpublic%2F1773274827%2FClaude+Certified+Architect+%E2%80%93+Foundations+Certification+Exam+Guide.pdf">official guide</a> first!</li><li><a href="https://drive.google.com/file/d/1luC0rnrET4tDYtS7xe5jUxMDZA-4qNf-/view">A very good architect playbook</a> (very well structured with most key points)</li><li><a href="https://github.com/hong-chu/claude-certified-architect-foundations">My LLM Wiki for CCA-F</a></li></ul><p><strong>Practice Exams:</strong></p><ul><li><a href="https://github.com/paullarionov/claude-certified-architect/blob/main/practical_test_en.html">https://github.com/paullarionov/claude-certified-architect/blob/main/practical_test_en.html</a> (free, medium)</li><li><a href="https://claudecertificationguide.com/mock-exam">https://claudecertificationguide.com/mock-exam</a> (free, medium)</li><li><a href="https://claude-certified-architect-mock-exam-cyberskill.vercel.app/">https://claude-certified-architect-mock-exam-cyberskill.vercel.app/</a> (free, hard, very helpful, please buy him a coffee!)</li><li><a href="https://www.certsafari.com/anthropic/claude-certified-architect">https://www.certsafari.com/anthropic/claude-certified-architect</a> (free, medium, with hundreds of questions)</li><li><a href="https://skillcertpro.com/product/claude-certified-architect-foundations-cca-f-exam-questions/">https://skillcertpro.com/product/claude-certified-architect-foundations-cca-f-exam-questions/</a> ($19.99, easy, honestly you may not need it if you already nailed the others)</li><li>I did the official practice exam last to check my readiness (free, easy)</li></ul><p>I saw a clear pattern in my scores. I was around <strong>720–750</strong> for the first two practice exams, then moved up to around <strong>780</strong>. After I adjusted my learning methodology, I quickly got over <strong>820</strong> consistently.</p><p>Two days before the real exam, I tested myself with the official practice exam and got around <strong>950/1000</strong>. That was when I felt ready.</p><h3>What I Took Away</h3><p>If I had to summarize what I learned from the exam preparation, I would say:</p><blockquote>Don’t trust the model alone. Design the system around the model.</blockquote><p>The model is important, but the workflow around the model matters just as much.</p><p>That includes:</p><ul><li>how instructions are written</li><li>how tools are described</li><li>how outputs are validated</li><li>how context is preserved</li><li>how agents coordinate</li><li>how humans are involved</li><li>and how safety is enforced</li></ul><p>This is the part I found most valuable.</p><p>The exam helped me move from “how do I use Claude better?” to “how do I design a more reliable AI workflow?”</p><p>That is a very different question.</p><h3>Final Thoughts</h3><p>Passing the CCA-F exam gave me more confidence, but it also made me more aware of how much there is still to learn.</p><p>I see it as one step in my journey toward becoming more AI-native in my work.</p><p>For me, the most meaningful lesson was this:</p><blockquote>Don’t only ask how to make the AI answer better. Ask how to design the system so the right answer is more likely, safer, and more reliable.</blockquote><p>That is the mindset I want to keep developing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6b967e6effb4" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[If AI Can Analyze Data, What Are Data Scientists For?]]></title>
            <link>https://medium.com/@yeesun.chu/if-ai-can-analyze-data-what-are-data-scientists-for-50671a6c0b0c?source=rss-8151f2116118------2</link>
            <guid isPermaLink="false">https://medium.com/p/50671a6c0b0c</guid>
            <category><![CDATA[product-analytics]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[data-scientist]]></category>
            <dc:creator><![CDATA[Hong Chu]]></dc:creator>
            <pubDate>Mon, 25 May 2026 02:16:28 GMT</pubDate>
            <atom:updated>2026-05-25T02:16:28.735Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o5SP3kc6dKCVagCEwepwqQ.jpeg" /></figure><p>When Meta coined the title “Data Scientist” in 2008, it kicked off a decade of arguments. What does a data scientist actually do?</p><p>Depends on who you ask. Some companies mean BI with a fancier name. Some mean product analytics. Some mean A/B experimentation and causal inference. Some mean ML engineering without the infra expectations.</p><p>The title never had a consistent definition, and honestly, that ambiguity was part of the appeal. You could be a data scientist doing almost anything with data.</p><p>None of that matters anymore.</p><p>Because here’s the uncomfortable question: if AI can write your SQL, build your dashboards, run your segmentation, fine-tune your models, and summarize your findings, what exactly are you for?</p><h4>What the Old Role Actually Was</h4><p>Let’s be honest about what the job actually looked like for most of us on the analytics track. Pull data. Build dashboards. Answer “why did retention drop.” Design an A/B test. Present findings in a deck that half the room skims on their phone.</p><p>Useful work. Hard skills. But vulnerable for two reasons.</p><p><strong>First, most deliverables are one-off</strong>. One query, one insight, one slide, forgotten by next quarter. One-off work is exactly what LLMs are best at killing.</p><p><strong>Second, Data Scientists are writers</strong>. We write SQL, Python, analyses, narratives. Writing is precisely what LLMs do. The junior craft that used to take years of training to master is now cheap.</p><h4>What Makes a Data Scientist Irreplaceable</h4><p>So what’s left? Not the technical skills. We have to admit AI does these better than us now. The hard skills we thought needed years of training to master, coding, modeling, querying, they melted.</p><p>But here’s the irony: <strong>what we used to call soft skills turned out to be the actual hard skills</strong>. And they fall into two categories.</p><p><strong>The human to AI interface.</strong> You translate business reality into analytical direction, and you judge whether the output is worth acting on. You know the pipeline broke last Tuesday. You know the PM’s “DAU drop” is actually a definition change. You know the 3% lift came from a segment you don’t want to keep. No model has access to any of this. You are the context layer, and you are the quality gate.</p><p><strong>The human to human interface.</strong> You reframe questions before anyone starts answering them. You push back when the org is optimizing the wrong metric. You sit in a room and tell a VP that the real problem isn’t what they think it is. This is not communication skills on a resume. This is the ability to change the direction of a decision before it gets made.</p><p>AI can produce the analysis. It cannot supply the context, and it cannot redirect the conversation.</p><h4>The New Job</h4><p>So the role survives. But it doesn’t stay the same.</p><p>If the irreplaceable value is context and framing, then the job is no longer about producing analyses. <strong>It’s about designing systems where your context and judgment are embedded into repeatable workflows</strong>.</p><p>The old job: someone asks a question, you answer it, you move on.</p><p>The new job: you build the system that answers that question every time it comes up, correctly, with your judgment baked in. You define what “active user” means so the agent doesn’t hallucinate a definition. You set the guardrails so the automated anomaly detection doesn’t page the team over noise. You design the evaluation so the org knows when to trust the AI output and when to escalate to a human.</p><p>You stop being the one who answers. You become the one who teaches the system how to answer, and when not to.</p><h4>The Paradox</h4><p>So the job isn’t dead. It just got a promotion nobody asked for.</p><p>And here’s the counterintuitive part: <strong>Data Scientists might be more important now than before AI</strong>.</p><p>When a junior analyst made a mistake, the blast radius was one bad slide deck. When an unguarded AI system makes a mistake, it makes it at scale, automatically, and confidently. The faster and more productive AI gets, the more expensive it is to let it run without judgment.</p><p><strong>Someone has to be the harness</strong>.</p><p>What does that look like in practice? We’ll get into that in the next article.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=50671a6c0b0c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Ontology × VSM — How Good Is Karpathy’s LLM Wiki?]]></title>
            <link>https://medium.com/@yeesun.chu/ontology-vsm-how-good-is-karpathys-llm-wiki-a8f98e1ce9d8?source=rss-8151f2116118------2</link>
            <guid isPermaLink="false">https://medium.com/p/a8f98e1ce9d8</guid>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[viable-systems-model]]></category>
            <category><![CDATA[harness-engineering]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ontology]]></category>
            <dc:creator><![CDATA[Hong Chu]]></dc:creator>
            <pubDate>Thu, 21 May 2026 06:21:47 GMT</pubDate>
            <atom:updated>2026-05-21T06:21:47.653Z</atom:updated>
            <content:encoded><![CDATA[<h3>Part 1 — Ontology: What Is There?</h3><p>In the AI era, building things is easy. Absurdly easy.</p><p>Everyone is a creator now, spinning up agents, workflows, knowledge systems in hours. The hard part isn’t building anymore. The hard part is: what did you just build? What is the relationship between the components? What is the relationship between the system and the human using it? These are ontology questions.</p><p>Ontology has three layers.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*l3FXqsW8LqXKLsm4QUERNw.jpeg" /></figure><p>Layer 1 asks: what is there? Layer 2 asks: how does it persist? Layer 3 asks: who counts as a participant? Layers 2 and 3 are intellectually interesting but increasingly academic.</p><p>Layer 1 is where the diagnostic power lives. Five rival lenses. Five different answers to “what is real.” Each one sees something in the system that the others miss.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*n8bOVpnvqdel2ltPSmbpPQ.jpeg" /></figure><p>Aristotle thought reality is things with properties <strong>[Lens 1]</strong>. A chair is a chair because it has legs, a seat, a back. In an AI system, this is the file, the row, the document sitting in storage. Clean and simple.</p><p>Whitehead disagreed. He thought reality is events, not things <strong>[Lens 2]</strong>. What looks stable is just a pattern that hasn’t broken yet. Your understanding of a topic last month is not your understanding today. A conversation is not a transcript. Process sees what substance can’t: things change.</p><p>Leibniz took it further. A thing means nothing in isolation. It’s defined entirely by its relations to other things <strong>[Lens 3]</strong>. A note that links to nothing is a dead note. A concept with no neighbors has no meaning. Relational ontology sees the graph, not the nodes.</p><p>Wheeler went more radical. Reality isn’t things or events or relations. Reality is information itself <strong>[Lens 4]</strong>. Not data. Distinctions. A schema that says “this field is a date” is making an ontological claim. Every API contract is an ontological contract. Most engineers live here without knowing it.</p><p>And Heidegger asked the question none of the others bothered with: real for whom? <strong>[Lens 5]</strong> Reality is what shows up for someone, in a context, with intent. Taste, judgment, refusal. The phenomenological lens sees the human. Not as an input source, but as the reason the system exists at all.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*R7RZMaroFBRiV5yVzxPRdA.jpeg" /></figure><h3>Part 2 — VSM: How Does It Stay Viable?</h3><p>Ontology answers three questions about reality. But there’s a fourth question it doesn’t touch: how does a system stay viable?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hl4sYgA4HjflpAo8KW7cqg.jpeg" /></figure><p>We covered Layer 1 — the five lenses. Layers 2 and 3 go deeper into persistence and participation. Intellectually rich, but increasingly academic. I’ll save those for a future piece.</p><p>The fourth question is different in kind. It’s not philosophy. It’s architecture. The Viable System Model comes from Stafford Beer’s cybernetics. It wasn’t designed for AI. It was designed for any system that needs to survive in a changing environment.</p><p>Think of a restaurant kitchen:</p><ul><li><strong>S1 Operations</strong> — line cooks actually cooking. The work gets done here.</li><li><strong>S2 Coordination</strong> — the expediter preventing collisions between stations.</li><li><strong>S3 Regulation</strong> — the head chef running tonight’s service. Optimizing what’s happening now.</li><li><strong>S3* Audit</strong> — the food-safety inspector. Independent check. Did we actually do what we said?</li><li><strong>S4 Scanning</strong> — the GM watching trends, weather, reservations. What’s coming from outside?</li><li><strong>S5 Identity</strong> — the owner’s vision. What kind of restaurant is this? What do we refuse to serve?</li></ul><p>Every viable system has all six. Remove any one and the system degrades. No S2, the kitchen collides. No S4, it gets blindsided. No S5, it drifts until it’s nothing in particular.</p><h3>Part 3 — Where Ontology Meets VSM</h3><p>Ontology is the diagnosis. VSM is the prescription. One tells you what kinds of reality your system is blind to. The other tells you how to organize so nothing essential is missing. They meet in the middle:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uOoiq8MpAM6zgUW550nrLw.jpeg" /></figure><p>S1 through S3 — operations, coordination, regulation — run on <strong>substance</strong> and <strong>informational</strong> ontology. Concrete things, structured rules, the system doing work. Most AI systems handle this well.</p><p>S4 — scanning — requires <strong>process</strong> and <strong>relational</strong> ontology. Noticing change, reading the environment, tracking how connections shift. This is where systems start to go blind.</p><p>S5 — identity — is <strong>phenomenological</strong>. What is this system for? What does it refuse? Without this, the system is capable but directionless.</p><p>A system built only on substance ontology can fill S1 through S3 but will always have empty S4 and S5. The ontology predicts the architectural gap. Now let’s test this on a real system.</p><h3>Part 4 — The Diagnostic: Karpathy’s LLM Wiki</h3><p><a href="https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f">Karpathy’s LLM Wiki</a> is arguably the most influential blueprint for personal AI knowledge systems right now. The insight is elegant: maintenance was always the bottleneck for personal wikis. LLMs solved maintenance. Suddenly a wiki that compounds knowledge over time becomes viable because upkeep cost drops to near zero.</p><p>The pattern: three layers (raw sources → wiki → schema), three operations (ingest, query, lint), two special files (index.md as catalog, log.md as audit trail). It deserves serious analysis, not just praise. So let&#39;s run both diagnostics.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QiSAbT5oqJqUJQmXrsd32Q.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VyM3B-kGfYl070sHXte9vQ.jpeg" /></figure><p>The two diagnostics converge on the same gap. The wiki is strong on substance, relational, and informational — which maps to strong S1 through S3. It’s partial on process and phenomenological — which maps to partial S4 and S5. The ontology predicted the architecture. The architecture confirmed the ontology.</p><p>This is not a criticism. Karpathy’s pattern is the best available blueprint for the knowledge-compounding core of a personal AI system. What the diagnostic reveals is not that it’s wrong, but that it’s deliberately scoped. Knowing where the scope ends is what lets you extend it intelligently rather than bolting on features that don’t address the real gaps.</p><h3>Part 5 — Closing the Gaps</h3><p>The diagnostic points to three extensions.</p><p><strong>Temporal awareness [Process, Lens 2].</strong> The wiki needs to know that knowledge evolves. Not just what you think now, but that you used to think differently. Version history in git isn’t enough. The system itself should distinguish current understanding from historical understanding. When you ask a question, it should know which version of your thinking to retrieve.</p><p><strong>External scanning [S4].</strong> Lint scans inward. Nothing scans outward. A viable system notices things without being asked. Heartbeats, watchers, scheduled checks. A new paper drops in your field. A dependency you rely on gets deprecated. A concept you wrote about six months ago has new developments. The system should surface these, not wait for you to manually drop in a source.</p><p><strong>Explicit identity [Phenomenological, Lens 5 / S5].</strong> The schema encodes conventions. It doesn’t encode purpose. An identity.md that answers two questions: what does this system optimize for, and what does it refuse? Not formatting rules. Principles. Without this, the system compounds knowledge in every direction equally. Capable but directionless.</p><p>None of these require rebuilding the wiki. They extend it. The core — substance, relational, informational, S1 through S3 — is already solid. The gaps are at the top of the stack, not the bottom.</p><p>These two frameworks are old. Ontology is ancient. Cybernetics is mid-century. Neither was built for AI. That’s exactly why they work. The things that outlast hype cycles are the things worth building on, and what we can learn from.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a8f98e1ce9d8" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Against Entropy in the AI Era]]></title>
            <link>https://medium.com/@yeesun.chu/against-entropy-in-the-ai-era-15360cf1e59c?source=rss-8151f2116118------2</link>
            <guid isPermaLink="false">https://medium.com/p/15360cf1e59c</guid>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[trading]]></category>
            <dc:creator><![CDATA[Hong Chu]]></dc:creator>
            <pubDate>Sat, 16 May 2026 05:21:20 GMT</pubDate>
            <atom:updated>2026-05-16T05:22:08.308Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/680/1*T2RLdC7tnOfQtrerx-uszQ.gif" /></figure><p>With degrees in both Math and Computer Science, I started as a quant and portfolio manager at hedge funds, and am now a principal data scientist in crypto.</p><p>Sitting at the intersection of four fast-pacing things: AI, Data, Trading, and Crypto, I’ve always been enthusiastic about finding the pattern, the structure underneath the noise, and reducing the entropy as much as possible.</p><p>We are in the fourth Industrial Revolution driven by AI. The relationships between human and human, human and machine, machine and machine are all being remodeled at once. That’s a lot of entropy, and things fall apart unless a robust system holds them together.</p><p>The old rule for anything complex: 2/3 research, 1/3 execution. Considering AI has made execution 10x faster (even with more bugs to be fixed, unfortunately), the ratio now should be even 90% thinking, 10% execution.</p><p><strong>So I’ve been thinking.</strong></p><p>There are new agent tools, new products, and new frameworks every week. Most will be dead in three months. Entropy collects its tax.</p><p>I’ve been studying what’s underneath: Ontology, Cybernetics, the Viable System Model, the architecture of knowledge systems. The structures that were true before LLMs existed and will outlast whatever is exciting today.</p><p>This is what I’ll be writing about: the intersection of deep theory and applied AI, how philosophical frameworks diagnose real system failures, how trading instincts apply to building AI systems, and how to tell what’s actually changing from what’s just noise.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=15360cf1e59c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>