[{"content":" AI isn\u0026rsquo;t just a tool upgrade — it\u0026rsquo;s a new computing platform revolution.\nPart 1: The Cracks Are Already Showing I\u0026rsquo;ve been job hunting recently, and I noticed something interesting: genuine \u0026ldquo;LLM integration developer\u0026rdquo; roles are still surprisingly rare. What\u0026rsquo;s more interesting is that even when companies do post them, most require:\nAI Agent experience LLM project experience RAG experience AI Workflow experience Here\u0026rsquo;s the problem: LLM development has only exploded in the past few years. How many engineers actually have complete AI development experience? Many engineers only started transitioning into LLM development a few months ago.\nIf you keep the bar this rigid and can\u0026rsquo;t hire anyone, those people will get picked up by other companies. In another year or two, you might not be able to hire them at all, even if you want to.\n(So if I\u0026rsquo;m job hunting right now — you could hire me today. Just don\u0026rsquo;t make me do LeetCode.)\nBut the really interesting part isn\u0026rsquo;t the hiring market. It\u0026rsquo;s that most companies, even now, have no idea how to make money with AI. The people who are actually using LLMs to build things are indie developers, small teams, hackers, and solo founders. They don\u0026rsquo;t even know if it will be profitable — but they\u0026rsquo;re running experiments anyway, because \u0026ldquo;this thing is just too cool.\u0026rdquo;\nThat hacker intuition is hard to explain with traditional business logic. Most great tech revolutions didn\u0026rsquo;t start with a clear business model. They started because a group of people thought something was fascinating.\nThat\u0026rsquo;s how the internet started. Personal computers. Smartphones. And now AI.\nThe real danger is that many large companies are still sitting comfortably in their existing lanes, asking:\nCan AI make money? How do we calculate AI ROI? Will AI disrupt our current business? But the question they should actually be asking is:\n\u0026ldquo;Will our company still exist in ten years?\u0026rdquo;\nBecause history has already answered this. Kodak didn\u0026rsquo;t die because its technology was weak. Nokia didn\u0026rsquo;t die because its engineers weren\u0026rsquo;t good enough. They died because when a new computing platform arrived, they were still living in the old era.\nAnd right now, the cracks are already showing.\nThe way I see it, a Niagara Falls is being held back by a thin mud wall — and that wall has started to crack.\nToday, 90% of internet companies are already standing at the edge of a cliff. They just haven\u0026rsquo;t realized it yet. Don\u0026rsquo;t believe me? Let\u0026rsquo;s run a social experiment starting now:\nBuild an AI Skill for Jira Build an AI Skill for productivity tools Build AI-native versions of various Web2.0 apps Watch what happens.\nPart 2: The Web4.0 Architecture \u0026ldquo;Web3.0\u0026rdquo; is a term that\u0026rsquo;s been talked to death. Why? Because it never produced a computing paradigm genuinely capable of restructuring Web2.0.\nBut AI is different.\nI\u0026rsquo;m calling this wave Web4.0, because AI is starting to deeply penetrate software itself. It\u0026rsquo;s no longer just a search bar, a chatbot, or an assistant tool — it\u0026rsquo;s gradually becoming part of the operating logic of software.\nI\u0026rsquo;d even argue this will be the fourth industrial revolution, because for the first time, machines are beginning to participate in producing software themselves.\n1. The Software Interface The software interface of Web4.0 will look very different from today\u0026rsquo;s — but not completely unfamiliar.\nFuture software will most likely split into: software on the left, AI on the right.\nThe left side will still be traditional GUI:\nTask lists Tables Charts Dashboards Status bars Humans still need to see state, so GUI isn\u0026rsquo;t going away.\nBut the right side will become an AI operation layer. Users won\u0026rsquo;t primarily interact through buttons anymore — they\u0026rsquo;ll accomplish most tasks through natural language, conversation, and intent.\nFor example:\n\u0026ldquo;Move this issue to next week and notify the relevant team members.\u0026rdquo;\nAI will:\nUpdate the issue Change the status Send notifications Adjust the timeline The left-side GUI\u0026rsquo;s role shifts to: showing the current state of the system. Users can even watch AI operate within the system and step in manually when needed.\nSoftware will shift from:\n\u0026ldquo;Humans operate software\u0026rdquo;\nto:\n\u0026ldquo;AI operates software. Humans supervise AI.\u0026rdquo;\n2. System Architecture The core shift in Web4.0 is that every frontend will eventually connect to an AI engine.\nWhether it\u0026rsquo;s:\nApp Web Desktop Skill Agent Everything will plug into:\nSLM + RAG\nMany people assume the future will be dominated by ever-larger models, but I don\u0026rsquo;t think so. LLMs are too expensive, enterprise-sensitive data can\u0026rsquo;t leave the building, and no serious company wants its core technology dependent on someone else\u0026rsquo;s API. A truly mature company will never build its core business permanently on external infrastructure.\nSo Web4.0 will inevitably move toward:\nEach company\u0026rsquo;s own SLM (Small Language Model) + proprietary RAG.\nLLMs will be more like early exploration tools, general reasoning engines, and product validation platforms. Mature products will eventually own their own:\nAI Engine Memory Knowledge Base Workflow System The competitive moat for companies will gradually shift away from:\nFrontend pages CRUD systems Database design And toward:\nRAG architecture Workflow orchestration Enterprise knowledge organization Agent collaboration systems 3. The Product Lifecycle The lifecycle of Web4.0 products will also change.\nIn the early stage, most teams will go straight to:\nOpenAI Claude Gemini Combined with:\nMCP RAG Workflow To ship fast — because the cost of experimentation is low, and the product can \u0026ldquo;come alive\u0026rdquo; from day one.\nThis is completely different from before. Products used to require massive amounts of custom logic before they were usable. Now AI already ships with enormous general-purpose capability.\nBut at the mature stage, companies will gradually migrate to:\nSLM + proprietary RAG\nThe reasons are practical:\nReduce costs Control data Reduce API dependency Ensure stability Establish technical sovereignty So the typical Web4.0 product evolution path will likely look like:\nLLM API ↓ RAG ↓ Workflow ↓ SLM ↓ Enterprise AI Engine 4. Customer Support Customer service may be one of the first industries to be fully restructured.\nBut this time, it\u0026rsquo;s real AI support — not the \u0026ldquo;fake AI that makes everyone want to throw their phone\u0026rdquo; from before.\nOld AI customer service:\nCouldn\u0026rsquo;t follow context Couldn\u0026rsquo;t hold a continuous conversation Couldn\u0026rsquo;t read emotions Only matched keywords So users always ended up demanding a human.\nWeb4.0 AI support is different. It will genuinely understand:\nContext Conversation history User sentiment User behavior It can even detect:\n\u0026ldquo;This user is getting frustrated.\u0026rdquo;\nAnd proactively say:\n\u0026ldquo;Let me connect you with a human agent.\u0026rdquo;\nMost companies\u0026rsquo; support operations will become fully AI-manageable. The scenarios that still require humans will shrink to:\nHigh-stakes decisions Emotional de-escalation Edge case handling Another industry, restructured.\n5. Version Iteration This is a more radical idea, but I think it\u0026rsquo;s cool — and the kind of thing that could go viral.\nIt\u0026rsquo;s this:\n\u0026ldquo;What goes into the next version is decided by user vote.\u0026rdquo;\nAI will:\nAnalyze user behavior Summarize user needs Auto-generate candidate features Let users vote And eventually, AI will auto-implement some of those features too.\nThe old software development flow:\nProduct Manager ↓ Requirements ↓ Engineering In the Web4.0 era, it may gradually become:\nUsers ↓ AI Analysis ↓ AI Implementation ↓ User Feedback Software will enter:\n\u0026ldquo;The era of high-velocity self-evolution.\u0026rdquo;\nPart 3: Web4.0 Is Not an Upgrade — It\u0026rsquo;s a Replacement Many companies still think of AI as a plugin, a feature, a chat window, a productivity tool.\nBut what AI is actually changing is the entire software architecture.\nWeb4.0 is not \u0026ldquo;Web2.0 + AI.\u0026rdquo; It\u0026rsquo;s a new computing platform — just like:\nPCs replaced mainframes Smartphones replaced parts of the PC Cloud computing restructured enterprise systems AI will redefine:\nSoftware Workflows Organizational structures Development models User interaction Enterprise architecture Most companies think they\u0026rsquo;re just waiting for AI to mature.\nBut actually:\nAI is waiting to replace them.\nWe may be standing at the single biggest technological inflection point since the invention of the computer. And many companies are already at the edge of the cliff — they just haven\u0026rsquo;t looked down yet.\n","date":"2026-05-06T00:00:00Z","image":"https://pub-deacd49348914a49b1254b01f351ef0d.r2.dev/2026/05/web4-is-coming/en/banner.png","permalink":"/post/web4-is-coming/","title":"Web4.0 Is Coming"},{"content":"LLM-Based AI Agent Architecture: A New Kind of Personal Computer on Your Device For a long time, we\u0026rsquo;ve thought of AI as a \u0026ldquo;chatbot.\u0026rdquo;\nBut if you step back and look from a systems architecture perspective, you\u0026rsquo;ll find that a truly mature AI agent looks more like a new kind of personal computer — one that lives on your device.\nIt has:\nA compute core Memory A file system A software system Input/output devices Long-term storage The difference is:\nIts core isn\u0026rsquo;t a traditional CPU. It\u0026rsquo;s an LLM.\nPart 1: The LLM Engine — A \u0026ldquo;CPU\u0026rdquo; Without Memory The LLM itself has no long-term memory.\nIt\u0026rsquo;s more like an inference engine:\nReceives input Reads context Performs reasoning Produces output Then \u0026ldquo;forgets\u0026rdquo; It cannot natively remember things that happened in the past.\nTherefore:\nThe LLM itself is more like a CPU than a complete agent.\nIt only handles computation.\nWhat makes AI \u0026ldquo;seem like it knows you\u0026rdquo; is the context provided externally.\nPart 2: Context — The AI Agent\u0026rsquo;s Memory If the LLM is the CPU,\nthen Context is the AI\u0026rsquo;s memory.\nAnd this memory should be split into two layers.\n1. Global Context This layer belongs to the entire agent.\nIt records:\nUser preferences Long-term goals Habitual behaviors Persona settings Persistent rules Historical knowledge For example:\n\u0026ldquo;User prefers Markdown\u0026rdquo; \u0026ldquo;User is learning AI Agents\u0026rdquo; \u0026ldquo;User habitually writes in Chinese\u0026rdquo; This information shapes agent behavior over time.\n2. Session Context This layer belongs only to the current conversation.\nFor example:\nThe current topic under discussion The current article structure The most recent rounds of dialogue Temporary reasoning results It\u0026rsquo;s more like temporary memory during program execution.\nThe Context Window Is Essentially a \u0026ldquo;Memory Limit\u0026rdquo; An LLM\u0026rsquo;s Context Window isn\u0026rsquo;t unlimited.\nThis means:\nHistory can\u0026rsquo;t accumulate indefinitely Information gets more expensive as the window fills Past the limit, content must be compressed Therefore:\nAn agent must manage memory like an operating system:\nCompress history Summarize Clear low-priority information Transfer long-term data Dynamically load needed data Therefore:\nThe Context Window is essentially the AI\u0026rsquo;s memory capacity.\nPart 3: Markdown Files — The Agent\u0026rsquo;s Hard Drive Long-term data shouldn\u0026rsquo;t stay in the context window.\nOtherwise:\nCosts keep rising Inference slows down The context balloons rapidly Therefore:\nLong-term memory should live in a file system.\nAnd one very natural form is Markdown files.\nFor example:\nNotes Project materials Journals World-building User profiles Writing material Long-term knowledge bases All of these can be stored as Markdown.\nThis means:\nTraditional Computer AI Agent Hard Drive Markdown File System Markdown has one enormous advantage:\nIt can be read by AI and directly by humans alike.\nTherefore:\nHumans can edit it AI can process it Git can version-control it Files can sync It persists even without AI This creates something like:\n\u0026ldquo;A shared knowledge space between humans and AI.\u0026rdquo;\nPart 4: Skills — Software Installed on AI Future AI agents won\u0026rsquo;t only have \u0026ldquo;knowledge.\u0026rdquo;\nThey\u0026rsquo;ll also have \u0026ldquo;skills.\u0026rdquo;\nFor example:\nWriting Skill Programming Skill Video Editing Skill Data Analysis Skill Project Management Skill These Skills might be composed of:\nPrompts Workflows Python code MCP configurations Tool invocation rules They are like:\nSoftware installed on the AI.\nTherefore:\nTraditional Computer AI Agent Software / App Skill Skills can be:\nInstalled Uninstalled Updated Shared Combined In the future there may even be:\nSkill Stores Skill Marketplaces Open-source Skill communities Part 5: Input/Output — More Than Just Text One of the biggest misconceptions about traditional chatbots is that people think AI only communicates through text.\nIn reality, future AI agents will have a complete multimodal I/O system.\nInput AI can read:\nText Voice Images Video Camera feeds Files Screen content Device state Output AI can generate:\nText Voice Images Video Automated actions Control commands Therefore:\nAn AI agent is fundamentally a new interaction layer.\nThe Complete System: A \u0026ldquo;Von Neumann-style\u0026rdquo; AI Computer When you put the whole architecture together:\nTraditional Computer AI Agent CPU LLM Engine Memory Context Hard Drive Markdown File System Software Skill Input Device Multimodal Input Output Device Multimodal Output You\u0026rsquo;ll find:\nIt increasingly resembles a real computer.\nExcept:\nThis computer isn\u0026rsquo;t built around a GUI.\nIt\u0026rsquo;s built around:\n\u0026ldquo;Language comprehension and reasoning.\u0026rdquo;\nThe Operating System: A Personal AI OS In the future, every person\u0026rsquo;s device may host a persistent AI Agent.\nOne that:\nUnderstands you Remembers you Helps you work Manages your knowledge Schedules your Skills Operates your devices Grows alongside you over time At that point:\nWhat we use might no longer just be:\nWindows macOS Android But rather:\nA new kind of personal AI operating system, with LLM at its core.\nAnd the chat box we use today\nmay only be the earliest prototype of this new era.\nReferences Park, Joon Sung et al.\nMemGPT: Towards LLMs as Operating Systems\narXiv:2310.08560\nhttps://arxiv.org/abs/2310.08560\nWang, Lei et al.\nLLM as OS, Agents as Apps: Envisioning AIOS, Agents and the AIOS-Agent Ecosystem\narXiv:2312.03815\nhttps://arxiv.org/abs/2312.03815\n","date":"2026-05-05T00:00:00Z","image":"https://pub-deacd49348914a49b1254b01f351ef0d.r2.dev/2026/05/llm-agent-architecture-a-new-kind-of-personal-computer/en/banner.png","permalink":"/post/llm-agent-architecture-new-kind-of-personal-computer/","title":"LLM-Based AI Agent Architecture: A New Kind of Personal Computer on Your Device"},{"content":"How AI Cured My Writing Procrastination A Procrastinator\u0026rsquo;s Excuse I haven\u0026rsquo;t written a blog post in years.\nNot because I had nothing to say — I just couldn\u0026rsquo;t bring myself to do it. Writing used to take me at least a week per article: brainstorming, drafting, revising over and over, then hunting for images, drawing flowcharts in draw.io, dragging and dropping for half an hour just to get a box on screen. The whole process was slow and exhausting, and I gradually built up a real resistance to it. Every time I opened my editor, the first thought that crossed my mind wasn\u0026rsquo;t \u0026ldquo;what should I write?\u0026rdquo; — it was \u0026ldquo;forget it, maybe another day.\u0026rdquo;\nAnd another day turned into several years.\nIt was only recently that I realized what had stopped me was never a lack of things to write about. It was that writing itself had become too heavy.\nNow AI has taken away a large part of that weight.\nThe Pain Points Turning ideas into coherent prose is hard: I just need to articulate my thoughts, even roughly, and AI can organize them into flowing paragraphs. I control the direction and the perspective; it polishes the language. This division of labor suits me much better.\nFinding images is time-consuming, and they\u0026rsquo;re often the wrong format or size: Now I describe what I want and AI generates it — style, composition, everything — in seconds.\nDrawing flowcharts is slow: I describe the logic to AI and it generates the diagram directly. I just check if it\u0026rsquo;s right and adjust if needed.\nMy Setup I\u0026rsquo;ve built myself a small writing pipeline, centered around two AI roles.\nThe first is the writer. I throw my ideas, thoughts, and core points at it, and it turns those fragments into a complete first draft.\nOnce the draft is ready, I edit it myself. This step is crucial — don\u0026rsquo;t underestimate your readers. The best articles are as short as possible, and most people these days (myself included) don\u0026rsquo;t have the patience for a long read. Later is never.\nAfter editing, I hand it to the second AI role — the editor. It reviews the article and does another round of polishing. Honestly, I\u0026rsquo;m not sure yet whether I need the editor role. I\u0026rsquo;ll try it this way and see.\nI\u0026rsquo;m not sure how many more articles I\u0026rsquo;ll write after this. But at least today, I opened my editor — and finished this one.\n","date":"2026-05-04T00:00:00Z","image":"https://pub-deacd49348914a49b1254b01f351ef0d.r2.dev/2026/05/ai-cured-writing-procrastination/banner.png","permalink":"/post/how-ai-cured-my-writing-procrastination/","title":"How AI Cured My Writing Procrastination"}]