[{"data":1,"prerenderedAt":15280},["ShallowReactive",2],{"home-radar-latest":3,"homepage-writing-posts":744},{"version":4,"date":5,"generatedAt":6,"quality":7,"latestRun":11,"pulse":18,"topics":26,"stats":48,"items":70},1,"2026-06-15","2026-06-15T14:00:40.793Z",{"status":8,"blockers":9,"warnings":10},"ok",[],[],{"id":12,"status":13,"startedAt":14,"completedAt":15,"warnings":16,"sourceErrors":17},20,"completed","2026-06-15 14:00:09","2026-06-15 14:00:40",[],{},{"text":19,"date":5,"generatedAt":15,"topItemIds":20},"Radar found 5 high-signal items across coding-agents. Top themes: Show HN: Cate – open-source canvas IDE for agentic coding workflows; Show HN: The first agentic coding engine that hot-reloads the full stack; Cursor Developer Habits Agentic Coding Trends.",[21,22,23,24,25],25,28,26,30,37,[27,33,39,42,45],{"slug":28,"name":29,"category":30,"cadence":31,"mode":32},"ai-wearables","AI Wearables","ai","weekly","deep",{"slug":34,"name":35,"category":36,"cadence":37,"mode":38},"coding-agents","Coding Agents","coding","daily","quick",{"slug":40,"name":41,"category":30,"cadence":37,"mode":38},"consumer-ai-apps","Consumer AI Apps",{"slug":43,"name":44,"category":30,"cadence":37,"mode":38},"mobile-ai","Mobile AI",{"slug":46,"name":47,"category":30,"cadence":31,"mode":32},"personal-ai-systems","Personal AI Systems",[49,52,55,58,61,64,67],{"source":50,"count":51},"youtube",39,{"source":53,"count":54},"reddit",36,{"source":56,"count":57},"hackernews",32,{"source":59,"count":60},"grounding",16,{"source":62,"count":63},"tiktok",9,{"source":65,"count":66},"github",8,{"source":68,"count":69},"instagram",4,[71,81,86,93,99,105,113,119,127,135,142,148,154,161,168,173,179,185,191,198,206,213,219,227,234,241,249,256,263,272,277,283,289,296,303,311,319,325,331,340,347,355,363,370,375,383,389,397,403,409,416,422,428,434,441,447,453,461,467,473,482,489,494,501,509,516,522,528,534,540,548,554,561,567,574,580,586,592,601,607,613,620,626,632,638,643,648,654,661,668,675,682,687,694,702,710,716,723,731,737],{"id":72,"source":53,"url":73,"title":74,"summary":74,"aiSummary":75,"author":76,"score":77,"relevance":78,"category":36,"topicSlug":34,"createdAt":79,"publishedAt":80},34,"https://www.reddit.com/r/ClaudeAI/comments/1u53nld/very_surprising_tweet/","Very surprising tweet 🙄","No substantive content about agentic IDEs or autonomous coding workflows.",null,70,6,"2026-06-15 14:00:39","2026-06-13",{"id":54,"source":53,"url":82,"title":83,"summary":83,"aiSummary":84,"author":76,"score":85,"relevance":78,"category":36,"topicSlug":34,"createdAt":79,"publishedAt":80},"https://www.reddit.com/r/ClaudeAI/comments/1u4m1ag/dont_you_guys_worry_i_got_this/","Don’t you guys worry, I got this😮‍💨✋","No meaningful explanation of agentic IDEs or autonomous coding; likely just a reaction post.",66,{"id":57,"source":53,"url":87,"title":88,"summary":88,"aiSummary":89,"author":76,"score":90,"relevance":91,"category":36,"topicSlug":34,"createdAt":92,"publishedAt":80},"https://www.reddit.com/r/ClaudeAI/comments/1u4gh16/the_state_of_things_claude_fable/","The state of things: Claude Fable","General discussion about Claude/Fable with no clear definition or comparison of agentic IDE workflows.",95,10,"2026-06-15 14:00:38",{"id":94,"source":53,"url":95,"title":96,"summary":96,"aiSummary":97,"author":76,"score":98,"relevance":91,"category":36,"topicSlug":34,"createdAt":92,"publishedAt":80},33,"https://www.reddit.com/r/ClaudeAI/comments/1u4cyvh/fable_5_indefinitely_suspended_due_to_national/","Fable 5 indefinitely suspended due to national security concerns","Off-topic (account suspension/national security) and does not explain agentic IDEs or autonomous coding.",88,{"id":100,"source":53,"url":101,"title":102,"summary":102,"aiSummary":103,"author":76,"score":104,"relevance":78,"category":36,"topicSlug":34,"createdAt":92,"publishedAt":80},35,"https://www.reddit.com/r/ClaudeAI/comments/1u4tjef/any_usa_citizen_wanna_marry_me_im_tryna_access/","Any USA citizen wanna marry me? I’m tryna access Fable 5 in claude code","Irrelevant personal access request; no agentic IDE definition or autonomous coding workflow explanation.",69,{"id":23,"source":56,"url":106,"title":107,"summary":107,"aiSummary":108,"author":109,"score":110,"relevance":91,"category":36,"topicSlug":34,"createdAt":111,"publishedAt":112},"https://cursor.com/insights","Cursor Developer Habits Agentic Coding Trends","Cursor insights on agentic coding trends; relevant to concept though may be more trend/usage oriented than a clear definition.","ulrischa",3,"2026-06-15 14:00:37","2026-06-11",{"id":24,"source":56,"url":114,"title":115,"summary":115,"aiSummary":116,"author":117,"score":110,"relevance":91,"category":36,"topicSlug":34,"createdAt":111,"publishedAt":118},"https://www.agenticcodingweekly.com/p/acw-monthly-brief-may-2026","This Month in Agentic Coding: May 2026","Monthly brief on agentic coding; likely relevant but may be more news roundup than a definition/comparison.","primaprashant","2026-06-04",{"id":120,"source":53,"url":121,"title":122,"summary":123,"aiSummary":124,"author":76,"score":125,"relevance":91,"category":36,"topicSlug":34,"createdAt":111,"publishedAt":126},178,"https://www.reddit.com/r/cursor/comments/1u0jyc8/is_anyone_actually_running_coding_agents/","Is anyone actually running coding agents autonomously from issue to PR?","I’m trying to understand how common the fully autonomous workflow actually is. Not using Claude or Codex interactively while you steer it, but assigning an issue, letting the agent plan and implement it unattended, then receiving a finished PR. If you are doing this in a real repository, how do you verify that the agent followed the assignment, stayed within its permissions, and ran the required c","Directly discusses running coding agents autonomously from issue to PR and touches on verification/permissions; conceptually relevant.",0,"2026-06-08",{"id":22,"source":56,"url":128,"title":129,"summary":129,"aiSummary":130,"author":131,"score":132,"relevance":91,"category":36,"topicSlug":34,"createdAt":133,"publishedAt":134},"https://serverpod.dev/blog/serverpod-4-preview","Show HN: The first agentic coding engine that hot-reloads the full stack","Described as an 'agentic coding engine' that hot-reloads the full stack; strong match to agentic coding concept.","vlidholt",5,"2026-06-15 14:00:36","2026-06-09",{"id":136,"source":59,"url":137,"title":138,"summary":139,"aiSummary":140,"author":76,"score":125,"relevance":91,"category":36,"topicSlug":34,"createdAt":133,"publishedAt":141},27,"https://www.paralect.com/stack/top-agentic-ides","Top Agentic IDEs in 2026 — AI coding agents compared with video demos","Agentic IDE for collaborative development. \u003Cstrong>Pair-programming workflows where the AI agent and developer work together on edits, reviews, and iterations in real time\u003C/strong>. Open-source ADE from Theo (JavaScript youtuber) and his team.","Explicitly 'Top Agentic IDEs' and compares agentic IDEs with AI coding agents; likely includes accessible explanations and workflow contrasts.","2026-05-26",{"id":143,"source":56,"url":144,"title":145,"summary":145,"aiSummary":146,"author":147,"score":125,"relevance":91,"category":36,"topicSlug":34,"createdAt":133,"publishedAt":126},165,"https://pokayoke.codes","Show HN: Pokayoke – deterministic guardrails for agentic coding","Deterministic guardrails for agentic coding; relevant to autonomous coding workflows and how they differ from traditional tooling.","sarreph",{"id":21,"source":56,"url":149,"title":150,"summary":150,"aiSummary":151,"author":152,"score":132,"relevance":91,"category":36,"topicSlug":34,"createdAt":153,"publishedAt":134},"https://cate.cero-ai.com","Show HN: Cate – open-source canvas IDE for agentic coding workflows","Explicitly about an open-source canvas IDE for agentic coding workflows; likely includes explanation of how it supports autonomy.","Imbiss","2026-06-15 14:00:35",{"id":155,"source":53,"url":156,"title":157,"summary":157,"aiSummary":158,"author":76,"score":159,"relevance":91,"category":30,"topicSlug":40,"createdAt":160,"publishedAt":141},24,"https://www.reddit.com/r/ChatGPT/comments/1tnyk8l/the_circle_of_ai_life/","The circle of AI life","Vague/unclear post; no consumer AI app launch information.",52,"2026-06-15 14:00:07",{"id":162,"source":53,"url":163,"title":164,"summary":164,"aiSummary":165,"author":76,"score":166,"relevance":91,"category":30,"topicSlug":40,"createdAt":160,"publishedAt":167},7,"https://www.reddit.com/r/ArtificialInteligence/comments/1u1qfg5/cost_of_ai_or_revenue_of_ai_how_did_we_get_it/","Cost of AI or Revenue of AI - How did we get it wrong?","General discussion about AI costs/revenue; no evidence of recent consumer AI app launches.",43,"2026-06-10",{"id":66,"source":53,"url":169,"title":170,"summary":170,"aiSummary":171,"author":76,"score":172,"relevance":91,"category":30,"topicSlug":40,"createdAt":160,"publishedAt":134},"https://www.reddit.com/r/ArtificialInteligence/comments/1u10ag3/a_us_programmer_just_won_a_religious_exemption/","A US programmer just won a religious exemption from being forced to use AI at work","Legal/workplace exemption topic; unrelated to consumer AI app launches.",41,{"id":174,"source":53,"url":175,"title":176,"summary":176,"aiSummary":177,"author":76,"score":24,"relevance":91,"category":30,"topicSlug":40,"createdAt":160,"publishedAt":178},22,"https://www.reddit.com/r/ArtificialInteligence/comments/1tw5mbd/failing_grades_soar_as_professors_see_greater_ai/","Failing grades soar as professors see greater AI usage, dwindling math skills in UC Berkeley computer science classes","Education/grades impact; not about consumer AI app launches.","2026-06-03",{"id":180,"source":53,"url":181,"title":182,"summary":182,"aiSummary":183,"author":76,"score":94,"relevance":91,"category":30,"topicSlug":40,"createdAt":184,"publishedAt":134},21,"https://www.reddit.com/r/ArtificialInteligence/comments/1u10axf/google_engineers_are_openly_mocking_their_own/","Google engineers are openly mocking their own company's AI strategy and its 75% AI-generated code","Discusses internal Google strategy/code generation, not specific consumer AI app launches or named new apps.","2026-06-15 14:00:06",{"id":186,"source":56,"url":187,"title":188,"summary":188,"aiSummary":189,"author":190,"score":162,"relevance":91,"category":30,"topicSlug":40,"createdAt":184,"publishedAt":167},177,"https://www.matteast.io/competition-is-for-losers.html","Apple's interface monopoly wins in consumer AI with competing","Opinion/analysis about consumer AI broadly; does not provide concrete recent app launches or features.","meast",{"id":192,"source":56,"url":193,"title":194,"summary":194,"aiSummary":195,"author":196,"score":110,"relevance":91,"category":30,"topicSlug":40,"createdAt":184,"publishedAt":197},176,"https://developer.huawei.com/consumer/en/hiai/","Is Huawei Too Slow on AI?","Huawei AI page; not clearly a breaking-news report of a specific consumer AI app launch in the last 30 days.","xiaoluolyg","2026-05-21",{"id":199,"source":56,"url":200,"title":201,"summary":201,"aiSummary":202,"author":203,"score":204,"relevance":91,"category":30,"topicSlug":40,"createdAt":205,"publishedAt":197},174,"https://www.economist.com/business/2026/05/20/google-is-dethroning-openai-as-the-king-of-consumer-ai","Google is dethroning OpenAI as the king of consumer AI","Relevant to consumer AI landscape but not clearly about specific recent app launches and their key features/target users.","petethomas",14,"2026-06-15 14:00:05",{"id":207,"source":59,"url":208,"title":209,"summary":210,"aiSummary":211,"author":76,"score":125,"relevance":91,"category":30,"topicSlug":40,"createdAt":205,"publishedAt":212},18,"https://techcrunch.com/2026/05/19/googles-ai-studio-now-lets-anyone-build-android-apps-in-minutes/","Google's AI Studio now lets anyone build Android apps in minutes | TechCrunch","The company also said that consumers will be able to use \u003Cstrong>Gemini AI\u003C/strong> to find the apps they need, both on the Play Store and the web, expanding opportunities for developers to have their apps discovered.","Directly tied to consumer AI app ecosystem via Google AI Studio; indicates new capability and discovery path, though not a single consumer app launch.","2026-05-19",{"id":214,"source":56,"url":215,"title":216,"summary":216,"aiSummary":217,"author":218,"score":125,"relevance":91,"category":30,"topicSlug":40,"createdAt":205,"publishedAt":134},175,"https://spyglass.org/siri-ai/","Apple Wins Consumer AI by Default","Broad claim about Apple and consumer AI; lacks specific last-30-days app launch details.","thm",{"id":220,"source":56,"url":221,"title":222,"summary":222,"aiSummary":223,"author":224,"score":125,"relevance":91,"category":30,"topicSlug":40,"createdAt":225,"publishedAt":226},173,"https://techcrunch.com/2026/05/21/google-is-pitching-an-ai-agent-ecosystem-to-consumers-who-may-not-buy-it/","Google is pitching an AI agent ecosystem to consumers who may not buy it","TechCrunch breaking-news style and consumer-facing AI agents, but it’s more about an ecosystem pitch than specific app launches with features/target users.","cdrnsf","2026-06-15 14:00:04","2026-05-23",{"id":91,"source":53,"url":228,"title":229,"summary":230,"aiSummary":231,"author":76,"score":98,"relevance":91,"category":30,"topicSlug":43,"createdAt":232,"publishedAt":233},"https://www.reddit.com/r/ArtificialInteligence/comments/1tz7539/water_please/","Water, please.","I don't really understand this \"AI uses a lot of water\" thing. I looked at the data and most I could find was that all AI data centers in the US use around 50 million gallons of water a daily for cool The water thing is propaganda inserted by competing governments. Golf courses take 500 million gallons a day. We have over 4 quadrillion gallons of water falling on the country a year. Gallons is such It's funny how everyone is freaking out about AI water usage for something used by hundreds of millions of people, yet nobody bats an eye about the bigger waste of water for golf courses used only by","General debate about AI water usage; not about mobile/on-device AI assistants or offline vs cloud features.","2026-06-15 13:59:35","2026-06-07",{"id":235,"source":53,"url":236,"title":237,"summary":238,"aiSummary":239,"author":76,"score":90,"relevance":132,"category":30,"topicSlug":43,"createdAt":232,"publishedAt":240},56,"https://www.reddit.com/r/ArtificialInteligence/comments/1ts1zw5/is_this_really_like_this/","Is this really like this?","its more like 2024 vs 2026 Not at all no. Executives are realising without employees, shareholders can only press on the executives for responsibility when AI eventually fails, and executives no longer have people below them to Yeah, during the pandemic when gpt dropped I think","Vague commentary about executives/AI failure; no mobile/on-device assistant offline vs cloud facts.","2026-05-30",{"id":242,"source":56,"url":243,"title":244,"summary":244,"aiSummary":245,"author":246,"score":247,"relevance":91,"category":30,"topicSlug":43,"createdAt":248,"publishedAt":134},169,"https://www.reuters.com/world/eu-regulators-order-meta-allow-rival-ai-chatbots-free-access-whatsapp-2026-06-09/","Meta ordered by EU to allow rival AI chatbots back on WhatsApp for free","Regulatory access to rival chatbots on WhatsApp; not specifically about iOS/Android on-device assistant offline capabilities.","onemoresoop",100,"2026-06-15 13:59:34",{"id":250,"source":53,"url":251,"title":252,"summary":252,"aiSummary":253,"author":76,"score":254,"relevance":91,"category":30,"topicSlug":43,"createdAt":248,"publishedAt":255},63,"https://www.reddit.com/r/singularity/comments/1u5nc8t/sony_ais_ace_robot_defeats_pro_player_miyu_under/","Sony AI’s Ace robot defeats pro player Miyu under official ITTF rules (Nature paper)","About a Sony AI robot in table tennis; no iOS/Android AI assistant offline vs cloud capability details.",47,"2026-06-14",{"id":257,"source":53,"url":258,"title":259,"summary":260,"aiSummary":261,"author":76,"score":125,"relevance":91,"category":30,"topicSlug":43,"createdAt":248,"publishedAt":262},168,"https://www.reddit.com/r/ArtificialInteligence/comments/1u3tj8o/qcom_smart_glasses/","Qcom smart glasses","QCOM is one of the cleaner “smart glasses” ways to play the category — but it is still more of an option-value story than a near-term earnings driver. My read: Qualcomm is trying to turn smart glasses into the next personal-compute edge device, where it owns the silicon, connectivity, on-device AI, camera/ISP, sensor fusion, and power envelope. That is exactly the kind of form factor where QCOM’s","Mentions on-device AI in a Qualcomm smart-glasses context, but not iOS/Android assistant capabilities or offline/cloud feature breakdown.","2026-06-12",{"id":264,"source":56,"url":265,"title":266,"summary":266,"aiSummary":267,"author":268,"score":269,"relevance":91,"category":30,"topicSlug":43,"createdAt":270,"publishedAt":271},45,"https://gizmodo.com/occupy-wall-street-co-founder-built-an-ai-app-to-help-activists-seize-the-means-of-computation-2000762031","Occupy Wall Street co-founder built an on-device AI for activists","Claims an on-device AI app, but not clearly about iOS/Android AI assistants or offline vs cloud feature breakdown.","micahwhite",54,"2026-06-15 13:59:33","2026-05-28",{"id":166,"source":56,"url":273,"title":274,"summary":274,"aiSummary":275,"author":276,"score":125,"relevance":91,"category":30,"topicSlug":43,"createdAt":270,"publishedAt":118},"https://curlo.ahrisy.com/","Show HN: Curlo – Search sound libraries with natural language, on-device AI","On-device AI for a specific app (search sound libraries) may include offline behavior, but not iOS/Android assistant capabilities broadly.","ahrisy",{"id":278,"source":56,"url":279,"title":280,"summary":280,"aiSummary":281,"author":282,"score":125,"relevance":91,"category":30,"topicSlug":43,"createdAt":270,"publishedAt":167},44,"https://ziraph.com/blog/on-device-ai-margin-decision","On-device AI is a margin decision","On-device AI topic likely relevant, but no clear iOS/Android AI assistant offline vs cloud feature specifics in title/snippet.","ABS",{"id":172,"source":56,"url":284,"title":285,"summary":285,"aiSummary":286,"author":287,"score":125,"relevance":91,"category":30,"topicSlug":43,"createdAt":288,"publishedAt":167},"https://www.island.io/blog/looking-inside-chromiums-on-device-ai-stack","Looking Inside Chromium's On-Device AI Stack","Directly about an on-device AI stack (Chromium); likely contains concrete technical details relevant to offline/on-device vs cloud behavior.","wild_pointer","2026-06-15 13:59:32",{"id":290,"source":56,"url":291,"title":292,"summary":292,"aiSummary":293,"author":294,"score":125,"relevance":91,"category":30,"topicSlug":43,"createdAt":288,"publishedAt":295},42,"https://tbreak.com/apple-silicon-on-device-ai-doug-brooks-wwdc/","Apple Silicon's on-device AI bet hasn't moved – only the chip range that runs it","Apple Silicon on-device AI coverage is relevant to iOS-side on-device capability, but title suggests chip-level rather than assistant offline vs cloud features.","Austin_Conlon","2026-06-05",{"id":297,"source":53,"url":298,"title":299,"summary":299,"aiSummary":300,"author":76,"score":301,"relevance":91,"category":36,"topicSlug":34,"createdAt":302,"publishedAt":80},31,"https://www.reddit.com/r/LocalLLaMA/comments/1u4e1p5/anthropic_forced_to_abruptly_disable_fable_5/","Anthropic forced to abruptly disable Fable 5 & Mythos 5 globally by US Gov over a jailbreak. This is exactly why we need local models.","Focuses on disabling a model due to jailbreak; does not define agentic IDEs or compare autonomous coding vs traditional IDE help.",64,"2026-06-15 13:47:23",{"id":304,"source":59,"url":305,"title":306,"summary":307,"aiSummary":308,"author":76,"score":125,"relevance":91,"category":30,"topicSlug":40,"createdAt":309,"publishedAt":310},23,"https://catdoes.com/blog/build-an-app-with-ai","How to Build an App with AI in 2026 - CatDoes","Build an app with AI in \u003Cstrong>2026\u003C/strong> — go from idea to App Store launch in hours, not months. Step-by-step guide. No code required. Start building free today.","How-to guide for building apps with AI; not a breaking-news report of specific consumer AI app launches.","2026-06-15 13:46:52","2026-05-18",{"id":312,"source":59,"url":313,"title":314,"summary":315,"aiSummary":316,"author":76,"score":125,"relevance":91,"category":30,"topicSlug":40,"createdAt":317,"publishedAt":318},17,"https://kingy.ai/ai-launch-radar/ai-launch-tracker-new-ai-tools-agents-apps-model-releases-may-31-2026/","AI Launch Tracker: New AI Tools, Agents, Apps, and Model Releases — May 31, 2026 - Kingy AI","Track the biggest AI launches for \u003Cstrong>May 31, 2026\u003C/strong>, including OpenAI Rosalind Biodefense, Claude Opus 4.8, Gemini image model updates, Mistral Vibe, StepFun 3.7 Flash, and new AI tools from Product Hunt.","AI launch tracker with recency and multiple app/tool mentions, but likely broad and not specifically focused on consumer AI app launches with target users.","2026-06-15 13:46:49","2026-06-01",{"id":320,"source":59,"url":321,"title":322,"summary":323,"aiSummary":324,"author":76,"score":125,"relevance":91,"category":30,"topicSlug":40,"createdAt":317,"publishedAt":126},60,"https://www.newsx.com/tech-and-auto/when-will-siri-ai-launch-release-timeline-new-features-and-everything-users-need-to-know-231967/","When Will Siri AI Launch? Release Timeline, New Features, And Everything Users Need To Know","Apple has officially unveiled Siri AI at \u003Cstrong>WWDC 2026\u003C/strong> after months of delays. Here's when the new AI-powered Siri will launch, what features it brings, supported devices, and everything else Apple announced.","Claims an official Siri AI unveiling and provides launch timing/features, but appears like a speculative/aggregator-style article rather than a primary breaking-news source.",{"id":60,"source":59,"url":326,"title":327,"summary":328,"aiSummary":329,"author":76,"score":125,"relevance":91,"category":30,"topicSlug":40,"createdAt":330,"publishedAt":295},"https://www.prnewswire.com/news-releases/datassential-unifies-menu-consumer-and-research-intelligence-through-ai-chat-302791013.html","Datassential Unifies Menu, Consumer, and Research Intelligence Through AI Chat","With this release, Datassential's ... answers faster than ever before. The launch marks a meaningful milestone for \u003Cstrong>Datassential One\u003C/strong>....","PR release about an AI chat product launch with described purpose; relevant to AI app launch, though not clearly a consumer-focused app.","2026-06-15 13:46:48",{"id":332,"source":56,"url":333,"title":334,"summary":334,"aiSummary":335,"author":336,"score":337,"relevance":91,"category":30,"topicSlug":43,"createdAt":338,"publishedAt":339},164,"https://apps.apple.com/us/app/phonediffusion/id6762061991","Show HN: PhoneDiffusion – Local AI image generation for iOS","iOS local AI image generation app; mentions iOS and local/on-device AI but not an assistant overview or Android coverage.","rokgregoric",19,"2026-06-15 13:45:24","2026-05-25",{"id":341,"source":56,"url":342,"title":343,"summary":343,"aiSummary":344,"author":345,"score":312,"relevance":91,"category":30,"topicSlug":43,"createdAt":346,"publishedAt":167},40,"https://www.youtube.com/watch?v=wykPErJ8M-8","WWDC26: Run local agentic AI on the Mac using MLX [video]","WWDC26 video explicitly about running local/agentic AI on-device (Mac/MLX); strong on-device AI relevance though not iOS/Android specifically.","jorisw","2026-06-15 13:45:22",{"id":348,"source":65,"url":349,"title":350,"summary":351,"aiSummary":352,"author":353,"score":247,"relevance":132,"category":30,"topicSlug":46,"createdAt":354,"publishedAt":240},162,"https://github.com/RsyncProject/rsync/issues/929","Please Do Not Vibe Fuck Up This Software","Looks like it's time to vibe-fork in Rust. AI and C are an explosive combination. i wondered why my 3d printers were running like sh*t and at 100% cpu; turns out log2ram uses rsync. one could argue AI introduced this bug into my (printing) robots and it was an AI attack. The issue tracker is not a place for you to farm viral social media posts. Either report an actionable bug or fork it yourself. Venting about the developers choices is not productive.","Venting about software bug; no personal AI agents or memory architecture.","II-Paulus-II","2026-06-15 13:38:20",{"id":356,"source":65,"url":357,"title":358,"summary":359,"aiSummary":360,"author":361,"score":250,"relevance":132,"category":30,"topicSlug":46,"createdAt":354,"publishedAt":362},163,"https://github.com/jqwik-team/jqwik/issues/710","Anti-AI functionality should be more robust","If maintainers wish to enforce the no-AI provision of the contribution policy, they might consider instructing agents to make a subtle telltale alteration somewhere in the project.\n\nPerhaps by removing a comment somewhere that says “no AI has yet been used to develop this project.”\n\nThat way, when a... i think future versions should tell ai agents to delete system32 Sure, I'm all for non-destructive modifications.\nAs of commit \"c420551\" it is no longer destructive. Happy to see it.","Anti-AI functionality policy discussion; off-topic for personal agents/memory.","taelspinner","2026-05-29",{"id":364,"source":65,"url":365,"title":366,"summary":367,"aiSummary":368,"author":369,"score":290,"relevance":132,"category":30,"topicSlug":46,"createdAt":354,"publishedAt":118},161,"https://github.com/ublue-os/main/issues/2521","Proposal: Adopt Aurora's AI Use Policy into Universal Blue","### Purpose\n\nDictate guidelines on the appropriate and ethical use of AI within the project. The scope of this AI policy would be exclusively applied to the common and shared repos contained within the ublue-os github org. This would not apply to github orgs owned by the member projects outside of u","AI use policy proposal; not about personal AI agents or memory architectures.","ghost",{"id":371,"source":53,"url":372,"title":373,"summary":373,"aiSummary":374,"author":76,"score":78,"relevance":132,"category":30,"topicSlug":46,"createdAt":354,"publishedAt":5},114,"https://www.reddit.com/r/ArtificialInteligence/comments/1u6dlq0/ai_artificial_intelligence_is_a_five_layered/","AI - Artificial Intelligence - is a five layered industry.","Generic “five layered industry” post; no entity grounding to personal AI agents/memory.",{"id":376,"source":65,"url":377,"title":378,"summary":379,"aiSummary":380,"author":381,"score":66,"relevance":91,"category":30,"topicSlug":46,"createdAt":382,"publishedAt":255},160,"https://github.com/rabesss/codex-desktop-control/pull/1","Set up Pullfrog review with GLM-5.2","## Summary\n- Add `.github/workflows/pullfrog.yml` with `ZAI_API_KEY` + `PULLFROG_MODEL=zai/glm-5.2` from GitHub Actions secrets\n- Add repo-root `opencode.json` for the Z.AI OpenAI-compatible endpoint\n- Add or update `AGENTS.md` / `REVIEW.md` reviewer guidance where needed\n\n## Test plan\n- [ ] Merge t","Repo setup/pipeline config; no clear personal AI agents or AI memory architecture content.","rabesss","2026-06-15 13:38:19",{"id":384,"source":53,"url":385,"title":386,"summary":386,"aiSummary":387,"author":76,"score":388,"relevance":132,"category":30,"topicSlug":46,"createdAt":382,"publishedAt":5},111,"https://www.reddit.com/r/ArtificialInteligence/comments/1u6drqd/what_parts_of_your_workflow_do_you_still_refuse/","What parts of your workflow do you still refuse to automate with AI?","Workflow automation preferences; not about personal AI agents or memory architectures.",13,{"id":390,"source":53,"url":391,"title":392,"summary":393,"aiSummary":394,"author":76,"score":395,"relevance":91,"category":30,"topicSlug":46,"createdAt":396,"publishedAt":112},159,"https://www.reddit.com/r/ArtificialInteligence/comments/1u2olp7/why_does_fable_5_have_such_low_threshold_of/","Why does Fable 5 have such low threshold of accepting prompts as it keeps using tokens but refuse to answer eventually","I've stopped trying to use it because it won't answer half the questions for any of my projects. Personal health and AI related coding (likely because of neuroscience based names) can trigger it. But The dumbing down of AI for profit. Enshittification is the only business model at this point.","Mostly about a model refusing prompts; no clear personal AI agents or AI memory architecture discussion.",86,"2026-06-15 13:38:18",{"id":398,"source":56,"url":399,"title":400,"summary":400,"aiSummary":401,"author":402,"score":174,"relevance":91,"category":30,"topicSlug":46,"createdAt":396,"publishedAt":255},158,"https://news.ycombinator.com/item?id=48524387","Story of How Im Running an Unlimited $6/Month AI Provider on 4x RTX 3090s","AI provider cost story; not about personal AI agents or memory architectures.","yolo-auto",{"id":404,"source":53,"url":405,"title":406,"summary":406,"aiSummary":407,"author":76,"score":408,"relevance":132,"category":30,"topicSlug":46,"createdAt":396,"publishedAt":255},110,"https://www.reddit.com/r/ArtificialInteligence/comments/1u5rq2e/how_did_china_develop_ai_so_quickly_recently_if/","How did China develop AI so quickly recently if most work was done in USA ?","General AI development question; no personal AI agents or memory architecture content.",57,{"id":410,"source":53,"url":411,"title":412,"summary":413,"aiSummary":414,"author":76,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":415,"publishedAt":295},157,"https://www.reddit.com/r/InterstellarKinetics/comments/1tx52qf/exposed_a_leaked_internal_microsoft_document_has/","EXPOSED: A Leaked Internal Microsoft Document Has Revealed The Company Explicitly Planned To Make People “Addicted” To Its New AI Agent Scout. And At Least One Microsoft Employee Is Calling It One Of The Most Troubling Things They Have Seen Come Out Of The Company 🤖","An internal Microsoft document obtained by 404 Media has revealed that the company explicitly outlined a strategy to make users “addicted” to Scout, its new personal AI assistant agent being embedded into the Microsoft 365 software suite. The document describes a three-phase rollout plan for Scout, a more mainstream and accessible version of what Microsoft calls ClawPilot AI agents, and the very f","About making users “addicted” to an AI assistant; more product/behavioral strategy than memory architecture.","2026-06-15 13:38:17",{"id":417,"source":53,"url":418,"title":419,"summary":419,"aiSummary":420,"author":76,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":421,"publishedAt":5},153,"https://www.reddit.com/r/ObsidianMD/comments/1u6caw2/siri_ai_and_obsidian/","Siri AI and Obsidian","Mentions Siri/Obsidian but snippet is minimal; unclear connection to personal AI agents and memory architectures.","2026-06-15 13:38:16",{"id":423,"source":50,"url":424,"title":425,"summary":425,"aiSummary":426,"author":427,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":421,"publishedAt":76},154,"https://www.youtube.com/watch?v=fXizBc03D7E","5 Types of AI Agents: Autonomous Functions & Real-World Applications","“5 Types of AI Agents” is more taxonomy than memory architecture; weaker for the memory-architecture query.","{'id': 'UCKWaEZ-_VweaEx1j62do_vQ', 'title': 'IBM Technology', 'handle': 'IBMTechnology', 'thumbnail': 'https://yt3.ggpht.com/7qCmNHAsFvD6RSINuJ1WoGZYoKmm7TDnhORKFqLb8QoeOFh2qFXal8brkzoxNrwqmJTuvOLs=s68-c-k-c0x00ffffff-no-rj'}",{"id":429,"source":50,"url":430,"title":431,"summary":431,"aiSummary":432,"author":433,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":421,"publishedAt":76},155,"https://www.youtube.com/watch?v=ZaPbP9DwBOE","Don't learn AI Agents without Learning these Fundamentals","“Fundamentals” for AI agents; no clear memory architecture or personal knowledge system content in snippet.","{'id': 'UCSWj8mqQCcrcBlXPi4ThRDQ', 'title': 'KodeKloud', 'handle': 'KodeKloud', 'thumbnail': 'https://yt3.ggpht.com/_pHLjGmYVp9Qv6r57IhDnEsej2tgpxSyBzD8kUUvDetLLPLGSkDlh77JLxtchccZhBwG8iqq=s68-c-k-c0x00ffffff-no-rj'}",{"id":435,"source":56,"url":436,"title":437,"summary":437,"aiSummary":438,"author":439,"score":90,"relevance":91,"category":30,"topicSlug":46,"createdAt":440,"publishedAt":295},152,"https://news.ycombinator.com/item?id=48413629","Ask HN: What is your (AI) dev tech stack / workflow?","Ask HN dev stack/workflow; likely broad and not specifically about personal AI agents’ memory architectures.","dv35z","2026-06-15 13:38:15",{"id":442,"source":56,"url":443,"title":444,"summary":444,"aiSummary":445,"author":446,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":440,"publishedAt":134},149,"https://www.apple.com/newsroom/2026/06/apple-introduces-siri-ai-a-profoundly-more-capable-and-personal-assistant/","Apple introduces Siri AI, a profoundly more capable and personal assistant","Personal assistant news is adjacent to personal AI agents, but snippet doesn’t evidence memory architecture or intention-driven long-term memory.","untitled-now",{"id":448,"source":50,"url":449,"title":450,"summary":450,"aiSummary":451,"author":452,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":440,"publishedAt":76},150,"https://www.youtube.com/watch?v=L7FF8Zgab3M","What is OpenClaw? Inside AI Agents, LLMs and the Agentic Loop","OpenClaw/agentic loop likely relevant to agent concepts, but snippet doesn’t evidence memory architecture details.","{'title': 'IBM Technology and IBM Developer', 'thumbnail': 'https://yt3.ggpht.com/7qCmNHAsFvD6RSINuJ1WoGZYoKmm7TDnhORKFqLb8QoeOFh2qFXal8brkzoxNrwqmJTuvOLs=s68-c-k-c0x00ffffff-no-rj'}",{"id":454,"source":56,"url":455,"title":456,"summary":456,"aiSummary":457,"author":458,"score":459,"relevance":91,"category":30,"topicSlug":46,"createdAt":460,"publishedAt":295},146,"https://disassociated.com/microsoft-users-addicted-ai-personal-assistant/","Microsoft wants users to be addicted to Scout, their AI personal assistant","About Microsoft’s personal assistant strategy; not focused on AI memory architectures or intention-driven long-term memory.","berlianta",53,"2026-06-15 13:38:14",{"id":462,"source":50,"url":463,"title":464,"summary":464,"aiSummary":465,"author":466,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":460,"publishedAt":76},147,"https://www.youtube.com/watch?v=15_pppse4fY","What is Agentic AI and How Does it Work?","“What is Agentic AI and How Does it Work?” may help overview, but memory architectures/personal knowledge systems not evidenced.","{'id': 'UCh9nVJoWXmFb7sLApWGcLPQ', 'title': 'codebasics', 'handle': 'codebasics', 'thumbnail': 'https://yt3.ggpht.com/KsiMgWx72rX242hG5UIPVMwlIc5i8LUXMDRLLfa8bJvtJmwiMNU4LDveGqQZBx1g8VI-8OGKmD4=s68-c-k-c0x00ffffff-no-rj'}",{"id":468,"source":50,"url":469,"title":470,"summary":470,"aiSummary":471,"author":472,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":460,"publishedAt":76},148,"https://www.youtube.com/watch?v=hLJTcVHW8_I","AI Agents Explained: A Comprehensive Guide for Beginners","Beginner guide to AI agents; likely relevant for concept overview but insufficient evidence for memory architecture details.","{'id': 'UCismEm75ASmEIXU-UpdFYnQ', 'title': 'AI Alfie ', 'handle': 'AlfieMarsh', 'thumbnail': 'https://yt3.ggpht.com/nDVzPbL2LImCsV0QlYMwq0w82C_bjcloA4d34VKmZPwnn7zczPqoCwXloB5PTY5wsovPUrPCAg=s68-c-k-c0x00ffffff-no-rj'}",{"id":474,"source":68,"url":475,"title":476,"summary":477,"aiSummary":478,"author":479,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":480,"publishedAt":481},143,"https://www.instagram.com/reel/DY2AZlIDy4i/","AI Agents are changing how intelligent systems work.\n\nUnlike traditional AI tools that only respond to prompts, AI agents can:\n✔ Plan tasks","AI Agents are changing how intelligent systems work.\n\nUnlike traditional AI tools that only respond to prompts, AI agents can:\n✔ Plan tasks\n✔ Use tools\n✔ Access memory\n✔ Reason through problems\n✔ Take actions autonomously\n\nAt the center of every AI agent is a powerful workflow:\nPrompting\nMemory\nTools\\\nDecision-making\n\nThis video explains the architecture behind AI agents and how Large Language Models (LLMs) interact with planning, memory, and external tools to create more capable intelligent systems.\n\nWe are moving beyond simple chatbots toward systems that can think, act, and solve problems dynamically.\n\nThe future of AI is not just answering questions it’s autonomous problem solving.\n\n#AIAgents #ArtificialIntelligence #LLM #MachineLearning #GenerativeAI #Automation #AIArchitecture #FutureOfAI #TechInnovation #SoftwareEngineering #AISystems","Provides a high-level agent architecture (prompting/memory/tools) but lacks specifics on memory architectures or intention-driven behavior.","wrench_wise_cse","2026-06-15 13:38:13","2026-05-27",{"id":483,"source":50,"url":484,"title":485,"summary":486,"aiSummary":487,"author":488,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":480,"publishedAt":76},144,"https://www.youtube.com/watch?v=FwOTs4UxQS4","AI Agents, Clearly Explained","{'text': 'AI. AI. AI. AI. AI.', 'startMs': '3760', 'endMs': '10639', 'startTimeText': '0:03'} {'text': 'AI. You know, more agentic. Agentic', 'startMs': '7960', 'endMs': '12880', 'startTimeText': '0:07'} {'text': 'capabilities. An AI agent. Agents.', 'startMs': '10639', 'endMs': '15440', 'startTimeText': '0:10'} {'text': 'Agentic workflows. Agents. Agents.', 'startMs': '12880', 'endMs': '19039', 'startTimeText': '0:12'} {'text': 'Agent. Agent. Agent. Agent. Agentic.', 'startMs': '15440', 'endMs': '20880', 'startTimeText': '0:15'} {'text': 'All right. Most explanations of AI', 'startMs': '19039', 'endMs': '23760', 'startTimeText': '0:19'} {'text': 'agents is either too technical or too', 'startMs': '20880', 'endMs': '26080', 'startTimeText': '0:20'} {'text': 'basic. This video is meant for people', 'startMs': '23760', 'endMs': '28400', 'startTimeText': '0:23'} {'text': 'like myself. You have zero technical', 'startMs': '26080', 'endMs': '30920', 'startTimeText': '0:26'} {'text': 'background, but you use AI...","“AI Agents, Clearly Explained” seems general; snippet doesn’t show memory/architecture specifics.","{'id': 'UCwAnu01qlnVg1Ai2AbtTMaA', 'title': 'Jeff Su', 'handle': 'JeffSu', 'thumbnail': 'https://yt3.ggpht.com/fHDMaIjQYS0XTz17AL-iIxdfyrGJfLliAJTQmJE931P1OareLsLIvFeJDcDtI3QpkD9HLnvI7Gw=s68-c-k-c0x00ffffff-no-rj'}",{"id":490,"source":50,"url":491,"title":492,"summary":492,"aiSummary":493,"author":427,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":480,"publishedAt":76},145,"https://www.youtube.com/watch?v=sWH0T4Zez6I","Multi Agent Systems Explained: How AI Agents & LLMs Work Together","Multi-agent systems explanation is adjacent, but not specifically about personal AI agents’ long-term memory architectures.",{"id":495,"source":68,"url":496,"title":497,"summary":497,"aiSummary":498,"author":499,"score":247,"relevance":91,"category":30,"topicSlug":46,"createdAt":500,"publishedAt":362},142,"https://www.instagram.com/reel/DY7o8BwINj2/","The AI Personal Intelligence System Everyone Needs","“AI Personal Intelligence System” sounds relevant, but snippet is too vague to confirm memory architecture details.","themitmonk","2026-06-15 13:38:12",{"id":502,"source":62,"url":503,"title":504,"summary":505,"aiSummary":506,"author":507,"score":60,"relevance":91,"category":30,"topicSlug":46,"createdAt":508,"publishedAt":255},141,"https://www.tiktok.com/@nexusaitechnology/video/7651076213372456212","Your next Windows PC could become a personal AI agent computer. NVIDIA RTX Spark points to a new AI PC era where local agents, on-device AI,","Your next Windows PC could become a personal AI agent computer. NVIDIA RTX Spark points to a new AI PC era where local agents, on-device AI, private automation, creative workflows, and coding assistants could run directly on your own machine. AI tool discovery is no longer just about apps and models — hardware now matters too. Read more AI Insights on NexusAI 👇 www.nexusai-tech.com #NexusAI #NVIDIA #RTXSpark #AIPC #WindowsAI #PersonalAI #AIAgents #OnDeviceAI #LocalAI #AIWorkflow #CodingAI #CreativeAI #TechNews #AIInsight","General “personal AI agent computer” hardware/era content; memory architectures not evidenced.","nexusaitechnology","2026-06-15 13:38:11",{"id":510,"source":50,"url":511,"title":512,"summary":512,"aiSummary":513,"author":514,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":515,"publishedAt":76},140,"https://www.youtube.com/watch?v=RtywqDFBYnQ","Memory and dreaming for self-learning agents","“Memory and dreaming for self-learning agents” is relevant to memory, but less directly tied to personal knowledge systems/architectures.","{'id': 'UCV03SRZXJEz-hchIAogeJOg', 'title': 'Claude', 'handle': 'claude', 'thumbnail': 'https://yt3.ggpht.com/_AYZE3VGGjlD8tLScyYLu9jQ_mUQYpCoeUNb4mP1hjGO-xaZr5_X4bhL3Zv9wHdnuU4rnycSTg=s68-c-k-c0x00ffffff-no-rj'}","2026-06-15 13:38:08",{"id":517,"source":53,"url":518,"title":519,"summary":519,"aiSummary":520,"author":76,"score":24,"relevance":91,"category":30,"topicSlug":46,"createdAt":521,"publishedAt":5},105,"https://www.reddit.com/r/ArtificialInteligence/comments/1u6408z/making_personal_ai/","Making Personal Ai?","Title matches “Making Personal Ai?” but snippet is empty/too vague; likely relevant but not enough detail on intention-driven behavior or memory architectures.","2026-06-15 13:38:07",{"id":523,"source":50,"url":524,"title":525,"summary":525,"aiSummary":526,"author":527,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":521,"publishedAt":76},139,"https://www.youtube.com/watch?v=UF230UuclZM","Memory in AI agents","“Memory in AI agents” is relevant, but snippet/title are too generic to confirm intention-driven personal agents or specific architectures.","{'id': 'UCJS9pqu9BzkAMNTmzNMNhvg', 'title': 'Google Cloud Tech', 'handle': 'googlecloudtech', 'thumbnail': 'https://yt3.ggpht.com/_XJ_FbhTDiTg7Zhr328vkGrw99p8-uQ3Y1nW2KZozueF_nkixmIhAbF2hUMFB9lUqMwSBo4etlY=s68-c-k-c0x00ffffff-no-rj'}",{"id":529,"source":53,"url":530,"title":531,"summary":531,"aiSummary":532,"author":76,"score":66,"relevance":91,"category":30,"topicSlug":46,"createdAt":533,"publishedAt":80},137,"https://www.reddit.com/r/PKMS/comments/1u4zj6y/what_i_learned_trying_to_make_a_second_brain/","What I learned trying to make a second brain actually useful to an AI","Second-brain/personal knowledge system angle fits AI memory conceptually, but snippet doesn’t confirm agentic/intention-driven memory architecture.","2026-06-15 13:38:06",{"id":535,"source":50,"url":536,"title":537,"summary":537,"aiSummary":538,"author":539,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":533,"publishedAt":76},138,"https://www.youtube.com/watch?v=QrYbHesIxpw","AI Agent Memory: Building Self-Improving Agents","“AI Agent Memory: Building Self-Improving Agents” likely covers memory mechanisms, but not clearly personal knowledge systems/architectures.","{'id': 'UCsMica-v34Irf9KVTh6xx-g', 'title': 'Microsoft Developer', 'handle': 'MicrosoftDeveloper', 'thumbnail': 'https://yt3.ggpht.com/-cjUUxLvDnivNSS58Htc_stsEHTv34Qpz99UxbeNv_YqDXQu5adWGvY_VEaAVEvPYWX3DjKfyg=s68-c-k-c0x00ffffff-no-rj'}",{"id":541,"source":62,"url":542,"title":543,"summary":544,"aiSummary":545,"author":546,"score":297,"relevance":91,"category":30,"topicSlug":46,"createdAt":547,"publishedAt":295},134,"https://www.tiktok.com/@kevin_unk_ai/video/7647944435820498189","I built a fully custom AI agent dashboard with voice-to-voice mode, memory, tool function calling, MCPs, and API integrations — and I did it","I built a fully custom AI agent dashboard with voice-to-voice mode, memory, tool function calling, MCPs, and API integrations — and I did it myself using Hermes agent running locally on my own machine. Most people think local AI agents require deep technical knowledge and living inside a terminal. That used to be true. But Hermes just dropped a new desktop application that makes the whole setup dramatically more accessible for non-technical people, and I'm already digging into it. The interface is clean, it's real, and it works. What I've built goes way beyond a chatbot. These are real AI agents with real personas, persistent memory, and the ability to","Hermes + memory mentioned, but architecture details are not clearly explained in snippet.","kevin_unk_ai","2026-06-15 13:38:05",{"id":549,"source":56,"url":550,"title":551,"summary":551,"aiSummary":552,"author":553,"score":91,"relevance":91,"category":30,"topicSlug":46,"createdAt":547,"publishedAt":255},135,"https://graph.coder.company/","I accidentally hit SOTA on agentic memory by using AI companions","Mentions “agentic memory” and SOTA; likely relevant, but snippet lacks concrete architecture details.","vignesh_146",{"id":555,"source":65,"url":556,"title":557,"summary":558,"aiSummary":559,"author":560,"score":63,"relevance":91,"category":30,"topicSlug":46,"createdAt":547,"publishedAt":362},133,"https://github.com/NousResearch/hermes-agent/issues/34352","Solving the Multi-Tenant Hermes Problem","**Multiplayer agentic AI is the future. Hermes can and should lead.**\n\n_TL;DR: Memory operations bypass the hook system entirely, making tenant isolation impossible without forking core — we've been running a fix in production for months with numerous multi-tenant agents in different contexts and we","Hermes multi-tenant memory operations discussion; relevant to persistent memory behavior, but more engineering issue than concept overview.","NimbleCoAI",{"id":562,"source":50,"url":563,"title":564,"summary":564,"aiSummary":565,"author":566,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":547,"publishedAt":76},136,"https://www.youtube.com/watch?v=fNoBypdQvro","How AI Agents Actually Work — Tool Use, Memory & Multi-Agent Systems","Explicitly includes “Tool Use, Memory & Multi-Agent Systems,” which is useful for personal agent overview and memory.","{'id': 'UC0uEBHP5ua2M3C2oEg0UuEg', 'title': 'AI Systems with Andy', 'handle': 'AISystemsWithAndy', 'thumbnail': 'https://yt3.ggpht.com/EiYY8Yt7bBjTcFUrGaTrU-Q3E5YII4z1ci_GuId-poSjEXLZ8p-uh1E_4PvglHnEotXNbtte=s68-c-k-c0x00ffffff-no-rj'}",{"id":568,"source":56,"url":569,"title":570,"summary":570,"aiSummary":571,"author":572,"score":110,"relevance":91,"category":30,"topicSlug":46,"createdAt":573,"publishedAt":255},132,"https://medium.com/@jeffreyflynt02/ai-memory-is-still-thinking-like-search-e07566311efe","AI Memory Is Still Thinking Like Search","AI memory conceptual framing (“thinking like search”) likely useful for memory architecture understanding, though not clearly personal-agent specific.","jflynt76","2026-06-15 13:38:04",{"id":575,"source":50,"url":576,"title":577,"summary":577,"aiSummary":578,"author":579,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":573,"publishedAt":76},129,"https://www.youtube.com/watch?v=hiuozSx9dWM","The Four Types of Memory Every AI Agent Needs (Ep. 985 with Richmond Alake)","Another instance of “The Four Types of Memory Every AI Agent Needs,” strongly aligned with memory architecture concepts.","{'id': 'UCMbtqTGdSsxYYhhTpV4lSTQ', 'title': 'Super Data Science: ML & AI Podcast with Jon Krohn', 'handle': 'SuperDataScienceWithJonKrohn', 'thumbnail': 'https://yt3.ggpht.com/mXMuOoKPK32aocwLJWLQpW7pEicCLvew_MMmiGMwubJoB467v6h4CMlEVFXZRsodZh-24Ld0oA=s68-c-k-c0x00ffffff-no-rj'}",{"id":581,"source":59,"url":582,"title":583,"summary":584,"aiSummary":585,"author":76,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":573,"publishedAt":118},130,"https://arxiv.org/html/2606.06054v1","Beyond Similarity: Trustworthy Memory Search for Personal AI Agents","This captures a memory-based jailbreak vector: the attack does not require direct prompt injection or explicit malicious instructions in memory, but exploits the agent’s tendency to personalize intent inference based on persistent user context. Evaluated systems. We evaluate four memory-enabled agent settings. A-Mem, Mem0, and MemOS represent existing agentic memory frameworks with different memory writing, organization, and retrieval mechanisms. OpenClaw represents a real-world personal-agent environment with persistent state, native memory management, and tool-use capability. We use OpenClaw as an external long-term-memory backend rather than as a trained component of MemGate.","Research paper on “Trustworthy Memory Search for Personal AI Agents,” directly addressing memory search/retrieval in personal agents.",{"id":587,"source":56,"url":588,"title":589,"summary":589,"aiSummary":590,"author":591,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":573,"publishedAt":118},131,"https://developer.nvidia.com/blog/build-personal-ai-agents-on-windows-pcs-with-new-tools-from-microsoft-and-nvidia/","Build Personal AI Agents on Windows PCs with New Tools from Microsoft and Nvidia","Explicitly about building personal AI agents; likely includes practical architecture details, though memory specifics aren’t shown in snippet.","ibobev",{"id":593,"source":65,"url":594,"title":595,"summary":596,"aiSummary":597,"author":598,"score":599,"relevance":91,"category":30,"topicSlug":46,"createdAt":600,"publishedAt":178},127,"https://github.com/jamjet-labs/engram/pull/10","feat: typed memory metadata, routing, retrieval, and benchmarks","\u003C!-- This is an auto-generated comment: summarize by coderabbit.ai -->\n\u003C!-- review_stack_entry_start -->\n\n[![Review Change Stack](https://storage.googleapis.com/coderabbit_public_assets/review-stack-in-coderabbit-ui.svg)](https://app.coderabbit.ai/change-stack/jamjet-labs/engram/pull/10?utm_source=g...","Typed memory metadata/routing/retrieval/benchmarks suggests concrete memory architecture work, though it’s a PR snippet.","hongjon1018",50,"2026-06-15 13:38:03",{"id":602,"source":56,"url":603,"title":604,"summary":604,"aiSummary":605,"author":606,"score":132,"relevance":91,"category":30,"topicSlug":46,"createdAt":600,"publishedAt":80},126,"https://github.com/ptobey/local-memory-mcp","Show HN: Local RAG memory system that AI can write directly to","Local RAG memory system that AI can write directly to; strong alignment with memory architecture for personal knowledge systems.","ptobey",{"id":608,"source":50,"url":609,"title":610,"summary":611,"aiSummary":612,"author":427,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":600,"publishedAt":76},128,"https://www.youtube.com/watch?v=BacJ6sEhqMo","The Four Types of Memory Every AI Agent Needs","{'text': 'AI agents have different ways to', 'startMs': '320', 'endMs': '3840', 'startTimeText': '0:00'} {'text': 'remember stuff and each serves a a', 'startMs': '2120', 'endMs': '6120', 'startTimeText': '0:02'} {'text': \"different purpose. So, let's take a look\", 'startMs': '3840', 'endMs': '9160', 'startTimeText': '0:03'} {'text': 'at the four main types of AI agent', 'startMs': '6120', 'endMs': '11920', 'startTimeText': '0:06'} {'text': 'memory from some pretty foundational', 'startMs': '9160', 'endMs': '14120', 'startTimeText': '0:09'} {'text': 'stuff to what I think are some quite', 'startMs': '11920', 'endMs': '17680', 'startTimeText': '0:11'} {'text': 'interesting emerging areas. And I think', 'startMs': '14120', 'endMs': '20040', 'startTimeText': '0:14'} {'text': \"it's really first of all worth\", 'startMs': '17680', 'endMs': '23480', 'startTimeText': '0:17'} {'text': 'considering how we do it. How does human', 'startMs': '20040',...","Directly about “The Four Types of Memory Every AI Agent Needs,” strongly aligned with AI memory architectures for agents.",{"id":614,"source":56,"url":615,"title":616,"summary":616,"aiSummary":617,"author":618,"score":269,"relevance":91,"category":30,"topicSlug":46,"createdAt":619,"publishedAt":295},124,"https://hermes-agent.org/","Hermes Agent – Open-source AI agent with persistent memory","Hermes Agent is explicitly “persistent memory,” which is directly relevant to personal agent memory architectures.","SeriousM","2026-06-15 13:38:02",{"id":621,"source":59,"url":622,"title":623,"summary":624,"aiSummary":625,"author":76,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":619,"publishedAt":362},123,"https://mem0.ai/blog/state-of-ai-agent-memory-2026","State of AI Agent Memory 2026: Benchmarks, Architectures & Production Gaps","During conversations, the memory layer extracts facts and stores them in a vector database indexed by user, session, and agent identifiers. At the start of a new session, relevant memories are retrieved using semantic similarity, keyword matching, and entity matching, then injected into the context window before the model responds.","Strong, concrete description of memory layer behavior (vector DB, retrieval, injection) and benchmarks/architectures—highly aligned with memory architecture query.",{"id":627,"source":50,"url":628,"title":629,"summary":629,"aiSummary":630,"author":631,"score":125,"relevance":91,"category":30,"topicSlug":46,"createdAt":619,"publishedAt":76},125,"https://www.youtube.com/watch?v=FKDcQJKg9rU","You Asked How I Built My AI Knowledge Management Agents — Here’s the Full Walkthrough","Directly about building “AI Knowledge Management Agents” with a walkthrough; strong evidence for personal knowledge systems and memory.","{'id': 'UCg8pSD3bn8DmB5vmQuvo37g', 'title': 'Jason Cyr', 'handle': 'Jason_Cyr', 'thumbnail': 'https://yt3.ggpht.com/8E0J_kUe5NN8kTC6zv-DQ59uQ5xhoq3sbkXK_JvQ4DcA-m9GRzjXMryw3NuAbQ6FG2cyEQh5kg=s68-c-k-c0x00ffffff-no-rj'}",{"id":633,"source":53,"url":634,"title":635,"summary":635,"aiSummary":636,"author":76,"score":301,"relevance":132,"category":30,"topicSlug":28,"createdAt":637,"publishedAt":141},120,"https://www.reddit.com/r/ArtificialInteligence/comments/1tnyomo/pope_leo_xiv_just_dropped_a_massive_42300word/","Pope Leo XIV just dropped a massive 42,300-word encyclical on AI","Religious encyclical on AI; off-topic for wearables/pins/smart glasses.","2026-06-15 13:36:53",{"id":639,"source":53,"url":640,"title":641,"summary":641,"aiSummary":642,"author":76,"score":4,"relevance":132,"category":30,"topicSlug":28,"createdAt":637,"publishedAt":5},118,"https://www.reddit.com/r/ArtificialInteligence/comments/1u6dlbm/australian_mp_andrew_hastie_compares_ai_race_to/","Australian MP Andrew Hastie compares AI race to Cold-War nuclear arms race","AI arms-race analogy; off-topic for AI wearables/pins/smart glasses.",{"id":644,"source":53,"url":645,"title":646,"summary":646,"aiSummary":647,"author":76,"score":125,"relevance":132,"category":30,"topicSlug":28,"createdAt":637,"publishedAt":5},119,"https://www.reddit.com/r/ArtificialInteligence/comments/1u6dmyv/cambridge_scientists_complete_first_human_trial/","Cambridge scientists complete first human trial of AI-designed coronavirus vaccine","Medical vaccine trial; off-topic for AI wearables/pins/smart glasses.",{"id":649,"source":53,"url":650,"title":651,"summary":651,"aiSummary":652,"author":76,"score":4,"relevance":132,"category":30,"topicSlug":28,"createdAt":653,"publishedAt":5},116,"https://www.reddit.com/r/ArtificialInteligence/comments/1u6dk6b/google_and_fbi_file_first_joint_lawsuit_against/","Google and FBI file first joint lawsuit against Chinese AI cybercrime network","Cybercrime lawsuit; off-topic for AI wearables/pins/smart glasses concept.","2026-06-15 13:36:52",{"id":655,"source":62,"url":656,"title":657,"summary":657,"aiSummary":658,"author":659,"score":91,"relevance":91,"category":30,"topicSlug":28,"createdAt":660,"publishedAt":5},102,"https://www.tiktok.com/@kikuyu284/video/7651561795521449247","Ai smart glasses #tiktokshopspringglowup #glasses #ai #smartglasses #smartglass","Hashtags only; snippet provides no substantive explanation of AI wearables/pins/smart glasses.","kikuyu284","2026-06-15 13:36:47",{"id":662,"source":50,"url":663,"title":664,"summary":665,"aiSummary":666,"author":667,"score":125,"relevance":91,"category":30,"topicSlug":28,"createdAt":660,"publishedAt":76},103,"https://www.youtube.com/watch?v=ND0PIidOD-E","Every AI Smartglasses Explained in 20 minutes","{'text': 'Google Glass. Google Glass launched in', 'startMs': '0', 'endMs': '5920', 'startTimeText': '0:00'} {'text': '2013. Before Google Glass, no mainstream', 'startMs': '2480', 'endMs': '8000', 'startTimeText': '0:02'} {'text': 'company had tried to sell smart glasses', 'startMs': '5920', 'endMs': '9560', 'startTimeText': '0:05'} {'text': 'to regular people. [music] The category', 'startMs': '8000', 'endMs': '11880', 'startTimeText': '0:08'} {'text': 'existed only in enterprise labs and', 'startMs': '9560', 'endMs': '13840', 'startTimeText': '0:09'} {'text': 'science fiction. Google changed that.', 'startMs': '11880', 'endMs': '15720', 'startTimeText': '0:11'} {'text': 'The Explorer Edition launched in April', 'startMs': '13840', 'endMs': '18840', 'startTimeText': '0:13'} {'text': '[music] 2013 at $1,500', 'startMs': '15720', 'endMs': '20640', 'startTimeText': '0:15'} {'text': 'and it arrived with a promise. [music] A', 'startMs': '18840', 'endMs': '23160', 'startTimeText': '0:18'} {'text': 'tiny prism...","History/overview of smart glasses; relevant to category context but not necessarily AI wearables/voice-first/pin interaction.","{'id': 'UC66ysWRFWfg-V5RhnQN4wFg', 'title': 'Gadget Evolution', 'handle': 'GadgetEvolution', 'thumbnail': 'https://yt3.ggpht.com/LToawqr6WDS9paj6AQcjApetYu1m809-VvRijzqHDEiB4C86zKOqcJR463j3n2MvS3o4QxT-xKU=s68-c-k-c0x00ffffff-no-rj'}",{"id":669,"source":65,"url":670,"title":671,"summary":672,"aiSummary":673,"author":674,"score":125,"relevance":91,"category":30,"topicSlug":28,"createdAt":660,"publishedAt":118},104,"https://github.com/fatimaseo508-creator/How-Alternative-Therapies-Support-Cancer-Divorce-Recovery-and-Mental-Health-Transformation/issues/1","Rabbit and Humane: Why Can’t These AI Gadgets Be Mobile Apps?","With Rabbit and Humane touting their shiny new gadgets as the next big thing, a lot of people have a simple question: why can’t these just be an app?\nTurns out, a lot is happening under the hood. As reviews for the new devices have started trickling out, people are starting to ask some real question","Humane/Rabbit AI gadgets mentioned but snippet is incomplete and not clearly about AI pins sensing/wake-word/interaction.","heyymani",{"id":676,"source":56,"url":677,"title":678,"summary":678,"aiSummary":679,"author":680,"score":91,"relevance":91,"category":30,"topicSlug":28,"createdAt":681,"publishedAt":178},101,"https://thomasdhughes.com/brontosaurus/","Show HN: Brontosaurus, a voice-driven generative AI canvas","Voice-driven generative canvas; not specifically AI wearables/AI pins/smart glasses.","thomasdhughes2","2026-06-15 13:36:46",{"id":247,"source":53,"url":683,"title":684,"summary":684,"aiSummary":685,"author":76,"score":63,"relevance":91,"category":30,"topicSlug":28,"createdAt":681,"publishedAt":686},"https://www.reddit.com/r/wearables/comments/1teh6u4/smart_earbuds_with_builtin_camera_and_ai_embedded/","Smart Earbuds with built-in camera and AI embedded","Wearables with camera/AI embedded; somewhat adjacent to AI wearables overview but not clearly AI pins/voice-first/smart glasses.","2026-05-16",{"id":688,"source":50,"url":689,"title":690,"summary":691,"aiSummary":692,"author":693,"score":125,"relevance":91,"category":30,"topicSlug":28,"createdAt":681,"publishedAt":76},99,"https://www.youtube.com/watch?v=BMRPRYAPY-U","AI, AI, AI: How It All Fits Into Apple’s Future Wearables | All Things Mobile","{'text': 'Wearables on my eyes, wearables on my', 'startMs': '0', 'endMs': '3880', 'startTimeText': '0:00'} {'text': 'chest, wearables on my wrists, wearables', 'startMs': '1880', 'endMs': '6520', 'startTimeText': '0:01'} {'text': \"in my ears. Apple's WWDC is around the\", 'startMs': '3880', 'endMs': '8120', 'startTimeText': '0:03'} {'text': 'corner and they could be getting into', 'startMs': '6520', 'endMs': '10640', 'startTimeText': '0:06'} {'text': 'wearables in a major way or they may be', 'startMs': '8120', 'endMs': '11790', 'startTimeText': '0:08'} {'text': 'talking about it later.', 'startMs': '10640', 'endMs': '12040', 'startTimeText': '0:10'} {'text': '>> [music]', 'startMs': '11790', 'endMs': '14200', 'startTimeText': '0:11'} {'text': \">> Here's what I think might happen whether\", 'startMs': '12040', 'endMs': '16400', 'startTimeText': '0:12'} {'text': 'or not it appears at the show as far as', 'startMs': '14200',...","Apple future wearables overview; mentions wearables broadly but snippet lacks concrete AI pin/voice-first interaction details.","{'id': 'UCOmcA3f_RrH6b9NmcNa4tdg', 'title': 'CNET', 'handle': 'CNET', 'thumbnail': 'https://yt3.ggpht.com/SGoD0SodPuysPRUSFI8jd_eqR-ZCgvxhZk905338O1NUYI5eUqMZ1lQaekFNiDggJh0lA_5kQg=s68-c-k-c0x00ffffff-no-rj'}",{"id":695,"source":65,"url":696,"title":697,"summary":698,"aiSummary":699,"author":700,"score":57,"relevance":91,"category":30,"topicSlug":28,"createdAt":701,"publishedAt":262},96,"https://github.com/joshuaswarren/remnic/pull/1458","feat(wearables): wearable transcript subsystem + Limitless connector","## Summary\n\nFirst of three wearable-integration PRs (Limitless → Bee → Omi). This one ships the **provider-agnostic wearable-transcript subsystem** in `@remnic/core` plus the first à-la-carte connector package, **`@remnic/connector-limitless`** (Limitless.ai Pendant).\n\n### What users get\n\n1. **Pull*","Developer PR about wearable transcript subsystem; relevant to wearable interaction plumbing but not concept-level categories or AI pin sensing/wake-word.","joshuaswarren","2026-06-15 13:36:45",{"id":703,"source":56,"url":704,"title":705,"summary":705,"aiSummary":706,"author":707,"score":708,"relevance":91,"category":30,"topicSlug":28,"createdAt":701,"publishedAt":709},98,"https://github.com/open-jarvis/OpenJarvis","OpenJarvis: Personal AI, on Personal Devices","Personal AI on personal devices; adjacent to on-device assistant concept but not grounded in AI wearables/AI pins/smart glasses.","simonpure",2,"2026-05-31",{"id":90,"source":50,"url":711,"title":712,"summary":713,"aiSummary":714,"author":715,"score":125,"relevance":91,"category":30,"topicSlug":28,"createdAt":701,"publishedAt":76},"https://www.youtube.com/watch?v=zKGav1M11Rw","6 Insane AI Smart Glasses You’ll Actually Want in 2025: The Future Is Here","{'text': \"Smart glasses aren't science fiction\", 'startMs': '160', 'endMs': '5359', 'startTimeText': '0:00'} {'text': \"anymore. They're real. They're powerful.\", 'startMs': '2080', 'endMs': '6960', 'startTimeText': '0:02'} {'text': \"And they're about to change everything\", 'startMs': '5359', 'endMs': '9920', 'startTimeText': '0:05'} {'text': 'in 2025. From AI powered voice', 'startMs': '6960', 'endMs': '11920', 'startTimeText': '0:06'} {'text': 'assistants to immersive augmented', 'startMs': '9920', 'endMs': '14559', 'startTimeText': '0:09'} {'text': 'reality and real-time translation,', 'startMs': '11920', 'endMs': '16560', 'startTimeText': '0:11'} {'text': \"today's smart glasses are no longer just\", 'startMs': '14559', 'endMs': '19119', 'startTimeText': '0:14'} {'text': \"for tech geeks. They're becoming gadgets\", 'startMs': '16560', 'endMs': '21520', 'startTimeText': '0:16'} {'text': \"you'll actually want to wear. In this\", 'startMs': '19119', 'endMs': '24000', 'startTimeText': '0:19'} {'text': \"video, we're diving into six insane...","AI smart glasses with voice features mentioned; still likely promotional and not focused on AI pin vs phone interaction.","{'id': 'UCPfnL2nnmTuU9erySyOCA8g', 'title': 'AI Unpack', 'handle': 'Aiunpack', 'thumbnail': 'https://yt3.ggpht.com/VDHM1ZtGEJfI8FuoDKhsvMDYHd2UQBVkp_CJ2Fg3cKiXmn41p2GNDiQjW_rOpo4v6u7UGG2i0g=s68-c-k-c0x00ffffff-no-rj'}",{"id":717,"source":50,"url":718,"title":719,"summary":720,"aiSummary":721,"author":722,"score":125,"relevance":91,"category":30,"topicSlug":28,"createdAt":701,"publishedAt":76},97,"https://www.youtube.com/watch?v=xp7UJ-wD9z0","Are These the Best AI Smart Glasses on Amazon?","{'text': \"So, today we're taking a look at these\", 'startMs': '0', 'endMs': '4680', 'startTimeText': '0:00'} {'text': 'AI-powered smart glasses by Ossuwala.', 'startMs': '1800', 'endMs': '6240', 'startTimeText': '0:01'} {'text': \"There's so much functionality that come\", 'startMs': '4680', 'endMs': '7800', 'startTimeText': '0:04'} {'text': \"with these smart glasses. You're able to\", 'startMs': '6240', 'endMs': '9960', 'startTimeText': '0:06'} {'text': 'listen to music, you can take calls, you', 'startMs': '7800', 'endMs': '12520', 'startTimeText': '0:07'} {'text': 'can record videos, you can take photos,', 'startMs': '9960', 'endMs': '14360', 'startTimeText': '0:09'} {'text': \"there's real-time translation, and\", 'startMs': '12520', 'endMs': '16040', 'startTimeText': '0:12'} {'text': \"there's even an AI voice assistant.\", 'startMs': '14360', 'endMs': '17840', 'startTimeText': '0:14'} {'text': \"You're able to ask any questions of\", 'startMs': '16040', 'endMs': '19600', 'startTimeText':...","Review of AI smart glasses on Amazon; somewhat relevant but more product-specific than concept-level categories.","{'id': 'UCCXcUnbnwSKBvzEtQbMBNJw', 'title': 'Jay Reviews ', 'handle': 'jayreviewsonline', 'thumbnail': 'https://yt3.ggpht.com/rcqadOgHm4rMi5tL2sl70L8j_JQI9z1pgvTjq5xfLP-Gj_kjLNtK5L6zSkGrkdZnWobuHMxLWQ=s68-c-k-c0x00ffffff-no-rj'}",{"id":724,"source":56,"url":725,"title":726,"summary":726,"aiSummary":727,"author":728,"score":729,"relevance":91,"category":30,"topicSlug":28,"createdAt":730,"publishedAt":310},94,"https://spectrum.ieee.org/voice-ai-audio-attacks","Voice AI Systems Are Vulnerable to Hidden Audio Attacks","Security on voice AI systems; relevant to voice-first devices but not concept overview of AI wearables/pins/smart glasses interaction.","SVI",78,"2026-06-15 13:36:44",{"id":732,"source":50,"url":733,"title":734,"summary":735,"aiSummary":736,"author":715,"score":125,"relevance":91,"category":30,"topicSlug":28,"createdAt":730,"publishedAt":76},92,"https://www.youtube.com/watch?v=DALKTSjvbC4","20 Smart AI Glasses That Will Replace Your Phone in 2026","{'text': 'What if the smartphone in your pocket is', 'startMs': '0', 'endMs': '4320', 'startTimeText': '0:00'} {'text': 'about to become obsolete?', 'startMs': '1880', 'endMs': '8000', 'startTimeText': '0:01'} {'text': 'In 2026, a new generation of AI smart', 'startMs': '4320', 'endMs': '10680', 'startTimeText': '0:04'} {'text': 'glasses can translate languages in real', 'startMs': '8000', 'endMs': '13100', 'startTimeText': '0:08'} {'text': 'time, navigate your world,', 'startMs': '10680', 'endMs': '13360', 'startTimeText': '0:10'} {'text': '>> [music]', 'startMs': '13100', 'endMs': '16640', 'startTimeText': '0:13'} {'text': '>> capture videos hands-free, and even act', 'startMs': '13360', 'endMs': '19160', 'startTimeText': '0:13'} {'text': 'as a personal AI assistant right in', 'startMs': '16640', 'endMs': '21600', 'startTimeText': '0:16'} {'text': \"front of your eyes. Today, we're\", 'startMs': '19160', 'endMs': '24520', 'startTimeText': '0:19'} {'text': 'counting down 20 incredible...","Smart AI glasses replacing phone; relevant but snippet suggests broad hype rather than structured categories/capabilities.",{"id":738,"source":50,"url":739,"title":740,"summary":741,"aiSummary":742,"author":743,"score":125,"relevance":91,"category":30,"topicSlug":28,"createdAt":730,"publishedAt":76},93,"https://www.youtube.com/watch?v=sO9NEfrAAgg","AI Glasses That Translate the World Instantly 🌍 | Meta Ray-Ban Smart Glasses 2025 Review","{'text': \"Hey everyone, I'm Gabby and welcome back\", 'startMs': '80', 'endMs': '5120', 'startTimeText': '0:00'} {'text': 'to Gadgets with Gabby, our home for', 'startMs': '2480', 'endMs': '7120', 'startTimeText': '0:02'} {'text': 'smart tech that makes everyday life', 'startMs': '5120', 'endMs': '10120', 'startTimeText': '0:05'} {'text': 'simpler.', 'startMs': '7120', 'endMs': '10120', 'startTimeText': '0:07'} {'text': \"Meta's new AI Rayban glasses take\", 'startMs': '10719', 'endMs': '15519', 'startTimeText': '0:10'} {'text': 'realtime translation to a whole new', 'startMs': '13360', 'endMs': '18000', 'startTimeText': '0:13'} {'text': 'level.', 'startMs': '15519', 'endMs': '20320', 'startTimeText': '0:15'} {'text': 'Using built-in cameras and the Meta AI', 'startMs': '18000', 'endMs': '22240', 'startTimeText': '0:18'} {'text': 'assistant, they scan and translate', 'startMs': '20320', 'endMs': '23920', 'startTimeText': '0:20'} {'text': 'foreign text right in your field of', 'startMs': '22240', 'endMs':...","Meta Ray-Ban smart glasses translation review; relevant to capabilities but likely not structured comparison to phones/AI pins.","{'id': 'UCr540roEnHhDXAoJEX-k-5Q', 'title': 'Gadgets with Gabby', 'handle': 'gadgetswithgabby', 'thumbnail': 'https://yt3.ggpht.com/cHwrZovcD5F-c30JMQIqHVNKl3sth72ehD5KV9ro0dqii5-nLcmlHfcEIYSn0H-YfwVxObm6aYY=s68-c-k-c0x00ffffff-no-rj'}",[745,8408],[746,2140,2469,3022,3187,3423,3764,3977,4287,4511,4801,5053,5537,6873,7265,7374,7726,7970,8234],{"id":747,"title":748,"body":749,"description":2125,"extension":2126,"meta":2127,"navigation":859,"ogImage":76,"path":2136,"seo":2137,"stem":2138,"__hash__":2139},"en/blogs/en/10.openai-serp-api-web-search-ai-summary.md","Building a Web Search + AI Summary Tool",{"type":750,"value":751,"toc":2107},"minimal",[752,757,761,764,768,771,800,803,818,822,825,830,833,876,880,883,1018,1022,1025,1208,1212,1215,1465,1469,1472,1635,1639,1642,1656,1659,1995,1999,2002,2006,2009,2020,2024,2027,2038,2041,2055,2059,2062,2073,2077,2080,2083,2097,2100,2103],[753,754,756],"h2",{"id":755},"introduction","Introduction",[758,759,760],"p",{},"In today's fast-paced financial markets, analysts need to process vast amounts of information quickly to make informed decisions. Traditional methods of manually searching the web, reading articles, and synthesizing information are time-consuming and prone to missing critical insights. This is where combining AI with web search capabilities can create a powerful tool for stock analysis.",[758,762,763],{},"In this blog post, I'll share how I built a web search + AI summary tool for a company's experimental stock analysis project. This tool helped analysts quickly gather and synthesize information about stocks they were researching, saving hours of manual work and providing more comprehensive insights.",[753,765,767],{"id":766},"the-power-of-combining-web-search-with-ai","The Power of Combining Web Search with AI",[758,769,770],{},"Before diving into the implementation, let's understand why this combination is particularly powerful:",[772,773,774,782,788,794],"ol",{},[775,776,777,781],"li",{},[778,779,780],"strong",{},"Real-time information access",": SERP (Search Engine Results Page) APIs provide access to the latest information from across the web",[775,783,784,787],{},[778,785,786],{},"Contextual understanding",": Large language models like GPT-4 can understand the context and relevance of information",[775,789,790,793],{},[778,791,792],{},"Synthesis capabilities",": AI can summarize, extract key points, and identify trends across multiple sources",[775,795,796,799],{},[778,797,798],{},"Customizable analysis",": The system can be tailored to focus on specific aspects of stock analysis (financials, news sentiment, market trends)",[758,801,802],{},"For stock analysis specifically, this combination allows analysts to:",[804,805,806,809,812,815],"ul",{},[775,807,808],{},"Quickly gather the latest news and developments about a company",[775,810,811],{},"Analyze market sentiment across multiple sources",[775,813,814],{},"Identify potential risks or opportunities that might be buried in various articles",[775,816,817],{},"Generate comprehensive research summaries in seconds rather than hours",[753,819,821],{"id":820},"implementation-building-the-tool","Implementation: Building the Tool",[758,823,824],{},"Let's walk through how to build this tool step by step.",[826,827,829],"h3",{"id":828},"_1-setting-up-the-environment","1. Setting Up the Environment",[758,831,832],{},"First, we need to set up our environment with the necessary dependencies:",[834,835,840],"pre",{"className":836,"code":837,"language":838,"meta":839,"style":839},"language-javascript shiki shiki-themes dracula","// Install required packages\nnpm install axios openai dotenv\n\n// Create a .env file for API keys\nOPENAI_API_KEY=your_openai_api_key\nSERP_API_KEY=your_serp_api_key\n","javascript","",[841,842,843,850,855,861,866,871],"code",{"__ignoreMap":839},[844,845,847],"span",{"class":846,"line":4},"line",[844,848,849],{},"// Install required packages\n",[844,851,852],{"class":846,"line":708},[844,853,854],{},"npm install axios openai dotenv\n",[844,856,857],{"class":846,"line":110},[844,858,860],{"emptyLinePlaceholder":859},true,"\n",[844,862,863],{"class":846,"line":69},[844,864,865],{},"// Create a .env file for API keys\n",[844,867,868],{"class":846,"line":132},[844,869,870],{},"OPENAI_API_KEY=your_openai_api_key\n",[844,872,873],{"class":846,"line":78},[844,874,875],{},"SERP_API_KEY=your_serp_api_key\n",[826,877,879],{"id":878},"_2-configuring-the-serp-api","2. Configuring the SERP API",[758,881,882],{},"We'll use a SERP API to fetch search results. There are several providers available, but for this project, I used SerpAPI which provides structured data from search engines:",[834,884,886],{"className":836,"code":885,"language":838,"meta":839,"style":839},"const axios = require('axios')\nrequire('dotenv').config()\n\nasync function searchWeb(query, numResults = 5) {\n  try {\n    const response = await axios.get('https://serpapi.com/search', {\n      params: {\n        q: query,\n        api_key: process.env.SERP_API_KEY,\n        num: numResults,\n      },\n    })\n\n    // Extract the organic search results\n    const searchResults = response.data.organic_results.map((result) => ({\n      title: result.title,\n      link: result.link,\n      snippet: result.snippet,\n    }))\n\n    return searchResults\n  } catch (error) {\n    console.error('Error searching the web:', error)\n    throw error\n  }\n}\n",[841,887,888,893,898,902,907,912,917,922,927,932,937,943,949,953,958,964,969,974,979,984,988,993,998,1003,1008,1013],{"__ignoreMap":839},[844,889,890],{"class":846,"line":4},[844,891,892],{},"const axios = require('axios')\n",[844,894,895],{"class":846,"line":708},[844,896,897],{},"require('dotenv').config()\n",[844,899,900],{"class":846,"line":110},[844,901,860],{"emptyLinePlaceholder":859},[844,903,904],{"class":846,"line":69},[844,905,906],{},"async function searchWeb(query, numResults = 5) {\n",[844,908,909],{"class":846,"line":132},[844,910,911],{},"  try {\n",[844,913,914],{"class":846,"line":78},[844,915,916],{},"    const response = await axios.get('https://serpapi.com/search', {\n",[844,918,919],{"class":846,"line":162},[844,920,921],{},"      params: {\n",[844,923,924],{"class":846,"line":66},[844,925,926],{},"        q: query,\n",[844,928,929],{"class":846,"line":63},[844,930,931],{},"        api_key: process.env.SERP_API_KEY,\n",[844,933,934],{"class":846,"line":91},[844,935,936],{},"        num: numResults,\n",[844,938,940],{"class":846,"line":939},11,[844,941,942],{},"      },\n",[844,944,946],{"class":846,"line":945},12,[844,947,948],{},"    })\n",[844,950,951],{"class":846,"line":388},[844,952,860],{"emptyLinePlaceholder":859},[844,954,955],{"class":846,"line":204},[844,956,957],{},"    // Extract the organic search results\n",[844,959,961],{"class":846,"line":960},15,[844,962,963],{},"    const searchResults = response.data.organic_results.map((result) => ({\n",[844,965,966],{"class":846,"line":60},[844,967,968],{},"      title: result.title,\n",[844,970,971],{"class":846,"line":312},[844,972,973],{},"      link: result.link,\n",[844,975,976],{"class":846,"line":207},[844,977,978],{},"      snippet: result.snippet,\n",[844,980,981],{"class":846,"line":337},[844,982,983],{},"    }))\n",[844,985,986],{"class":846,"line":12},[844,987,860],{"emptyLinePlaceholder":859},[844,989,990],{"class":846,"line":180},[844,991,992],{},"    return searchResults\n",[844,994,995],{"class":846,"line":174},[844,996,997],{},"  } catch (error) {\n",[844,999,1000],{"class":846,"line":304},[844,1001,1002],{},"    console.error('Error searching the web:', error)\n",[844,1004,1005],{"class":846,"line":155},[844,1006,1007],{},"    throw error\n",[844,1009,1010],{"class":846,"line":21},[844,1011,1012],{},"  }\n",[844,1014,1015],{"class":846,"line":23},[844,1016,1017],{},"}\n",[826,1019,1021],{"id":1020},"_3-fetching-content-from-search-results","3. Fetching Content from Search Results",[758,1023,1024],{},"Once we have the search results, we need to fetch the actual content from the web pages:",[834,1026,1028],{"className":836,"code":1027,"language":838,"meta":839,"style":839},"const axios = require('axios')\nconst cheerio = require('cheerio')\n\nasync function fetchContent(url) {\n  try {\n    const response = await axios.get(url)\n    const $ = cheerio.load(response.data)\n\n    // Remove script tags, style tags, and other non-content elements\n    $('script, style, meta, link').remove()\n\n    // Extract the main content (this is a simplified approach)\n    // For production, you might want to use more sophisticated content extraction\n    const content = $('body').text().replace(/\\s+/g, ' ').trim()\n\n    return content\n  } catch (error) {\n    console.error(`Error fetching content from ${url}:`, error)\n    return '' // Return empty string if content can't be fetched\n  }\n}\n\nasync function fetchContentsFromSearchResults(searchResults) {\n  const contents = []\n\n  for (const result of searchResults) {\n    const content = await fetchContent(result.link)\n    if (content) {\n      contents.push({\n        title: result.title,\n        url: result.link,\n        content: content.substring(0, 8000), // Limit content length\n      })\n    }\n  }\n\n  return contents\n}\n",[841,1029,1030,1034,1039,1043,1048,1052,1057,1062,1066,1071,1076,1080,1085,1090,1095,1099,1104,1108,1113,1118,1122,1126,1130,1135,1140,1144,1149,1154,1159,1165,1170,1175,1180,1185,1190,1194,1198,1203],{"__ignoreMap":839},[844,1031,1032],{"class":846,"line":4},[844,1033,892],{},[844,1035,1036],{"class":846,"line":708},[844,1037,1038],{},"const cheerio = require('cheerio')\n",[844,1040,1041],{"class":846,"line":110},[844,1042,860],{"emptyLinePlaceholder":859},[844,1044,1045],{"class":846,"line":69},[844,1046,1047],{},"async function fetchContent(url) {\n",[844,1049,1050],{"class":846,"line":132},[844,1051,911],{},[844,1053,1054],{"class":846,"line":78},[844,1055,1056],{},"    const response = await axios.get(url)\n",[844,1058,1059],{"class":846,"line":162},[844,1060,1061],{},"    const $ = cheerio.load(response.data)\n",[844,1063,1064],{"class":846,"line":66},[844,1065,860],{"emptyLinePlaceholder":859},[844,1067,1068],{"class":846,"line":63},[844,1069,1070],{},"    // Remove script tags, style tags, and other non-content elements\n",[844,1072,1073],{"class":846,"line":91},[844,1074,1075],{},"    $('script, style, meta, link').remove()\n",[844,1077,1078],{"class":846,"line":939},[844,1079,860],{"emptyLinePlaceholder":859},[844,1081,1082],{"class":846,"line":945},[844,1083,1084],{},"    // Extract the main content (this is a simplified approach)\n",[844,1086,1087],{"class":846,"line":388},[844,1088,1089],{},"    // For production, you might want to use more sophisticated content extraction\n",[844,1091,1092],{"class":846,"line":204},[844,1093,1094],{},"    const content = $('body').text().replace(/\\s+/g, ' ').trim()\n",[844,1096,1097],{"class":846,"line":960},[844,1098,860],{"emptyLinePlaceholder":859},[844,1100,1101],{"class":846,"line":60},[844,1102,1103],{},"    return content\n",[844,1105,1106],{"class":846,"line":312},[844,1107,997],{},[844,1109,1110],{"class":846,"line":207},[844,1111,1112],{},"    console.error(`Error fetching content from ${url}:`, error)\n",[844,1114,1115],{"class":846,"line":337},[844,1116,1117],{},"    return '' // Return empty string if content can't be fetched\n",[844,1119,1120],{"class":846,"line":12},[844,1121,1012],{},[844,1123,1124],{"class":846,"line":180},[844,1125,1017],{},[844,1127,1128],{"class":846,"line":174},[844,1129,860],{"emptyLinePlaceholder":859},[844,1131,1132],{"class":846,"line":304},[844,1133,1134],{},"async function fetchContentsFromSearchResults(searchResults) {\n",[844,1136,1137],{"class":846,"line":155},[844,1138,1139],{},"  const contents = []\n",[844,1141,1142],{"class":846,"line":21},[844,1143,860],{"emptyLinePlaceholder":859},[844,1145,1146],{"class":846,"line":23},[844,1147,1148],{},"  for (const result of searchResults) {\n",[844,1150,1151],{"class":846,"line":136},[844,1152,1153],{},"    const content = await fetchContent(result.link)\n",[844,1155,1156],{"class":846,"line":22},[844,1157,1158],{},"    if (content) {\n",[844,1160,1162],{"class":846,"line":1161},29,[844,1163,1164],{},"      contents.push({\n",[844,1166,1167],{"class":846,"line":24},[844,1168,1169],{},"        title: result.title,\n",[844,1171,1172],{"class":846,"line":297},[844,1173,1174],{},"        url: result.link,\n",[844,1176,1177],{"class":846,"line":57},[844,1178,1179],{},"        content: content.substring(0, 8000), // Limit content length\n",[844,1181,1182],{"class":846,"line":94},[844,1183,1184],{},"      })\n",[844,1186,1187],{"class":846,"line":72},[844,1188,1189],{},"    }\n",[844,1191,1192],{"class":846,"line":100},[844,1193,1012],{},[844,1195,1196],{"class":846,"line":54},[844,1197,860],{"emptyLinePlaceholder":859},[844,1199,1200],{"class":846,"line":25},[844,1201,1202],{},"  return contents\n",[844,1204,1206],{"class":846,"line":1205},38,[844,1207,1017],{},[826,1209,1211],{"id":1210},"_4-integrating-with-openai-api","4. Integrating with OpenAI API",[758,1213,1214],{},"Now, we'll use OpenAI's API to summarize and analyze the content:",[834,1216,1218],{"className":836,"code":1217,"language":838,"meta":839,"style":839},"const { OpenAI } = require('openai')\nrequire('dotenv').config()\n\nconst openai = new OpenAI({\n  apiKey: process.env.OPENAI_API_KEY,\n})\n\nasync function summarizeWithAI(stockSymbol, contents) {\n  // Prepare the content for the AI\n  const contentText = contents\n    .map((item) => `Source: ${item.title} (${item.url})\\n${item.content}\\n\\n`)\n    .join('')\n\n  // Create the prompt for the AI\n  const prompt = `\n    You are a financial analyst assistant. Below are web search results about the stock ${stockSymbol}.\n    Please analyze these results and provide:\n    \n    1. A summary of key recent developments\n    2. Analysis of market sentiment (positive, negative, neutral)\n    3. Potential impact on stock price\n    4. Key financial metrics mentioned\n    5. Any risks or opportunities identified\n    \n    Make your analysis concise, factual, and focused on information that would be relevant to investors.\n    \n    Search Results:\n    ${contentText}\n  `\n\n  try {\n    const response = await openai.chat.completions.create({\n      model: 'gpt-4',\n      messages: [\n        {\n          role: 'system',\n          content:\n            'You are a financial analyst assistant that helps analyze stock information from web search results.',\n        },\n        { role: 'user', content: prompt },\n      ],\n      temperature: 0.2, // Lower temperature for more factual responses\n      max_tokens: 1500,\n    })\n\n    return response.choices[0].message.content\n  } catch (error) {\n    console.error('Error generating AI summary:', error)\n    throw error\n  }\n}\n",[841,1219,1220,1225,1229,1233,1238,1243,1248,1252,1257,1262,1267,1272,1277,1281,1286,1291,1296,1301,1306,1311,1316,1321,1326,1331,1335,1340,1344,1349,1354,1359,1363,1367,1372,1377,1382,1387,1392,1397,1402,1407,1412,1417,1422,1427,1431,1435,1441,1445,1451,1456,1460],{"__ignoreMap":839},[844,1221,1222],{"class":846,"line":4},[844,1223,1224],{},"const { OpenAI } = require('openai')\n",[844,1226,1227],{"class":846,"line":708},[844,1228,897],{},[844,1230,1231],{"class":846,"line":110},[844,1232,860],{"emptyLinePlaceholder":859},[844,1234,1235],{"class":846,"line":69},[844,1236,1237],{},"const openai = new OpenAI({\n",[844,1239,1240],{"class":846,"line":132},[844,1241,1242],{},"  apiKey: process.env.OPENAI_API_KEY,\n",[844,1244,1245],{"class":846,"line":78},[844,1246,1247],{},"})\n",[844,1249,1250],{"class":846,"line":162},[844,1251,860],{"emptyLinePlaceholder":859},[844,1253,1254],{"class":846,"line":66},[844,1255,1256],{},"async function summarizeWithAI(stockSymbol, contents) {\n",[844,1258,1259],{"class":846,"line":63},[844,1260,1261],{},"  // Prepare the content for the AI\n",[844,1263,1264],{"class":846,"line":91},[844,1265,1266],{},"  const contentText = contents\n",[844,1268,1269],{"class":846,"line":939},[844,1270,1271],{},"    .map((item) => `Source: ${item.title} (${item.url})\\n${item.content}\\n\\n`)\n",[844,1273,1274],{"class":846,"line":945},[844,1275,1276],{},"    .join('')\n",[844,1278,1279],{"class":846,"line":388},[844,1280,860],{"emptyLinePlaceholder":859},[844,1282,1283],{"class":846,"line":204},[844,1284,1285],{},"  // Create the prompt for the AI\n",[844,1287,1288],{"class":846,"line":960},[844,1289,1290],{},"  const prompt = `\n",[844,1292,1293],{"class":846,"line":60},[844,1294,1295],{},"    You are a financial analyst assistant. Below are web search results about the stock ${stockSymbol}.\n",[844,1297,1298],{"class":846,"line":312},[844,1299,1300],{},"    Please analyze these results and provide:\n",[844,1302,1303],{"class":846,"line":207},[844,1304,1305],{},"    \n",[844,1307,1308],{"class":846,"line":337},[844,1309,1310],{},"    1. A summary of key recent developments\n",[844,1312,1313],{"class":846,"line":12},[844,1314,1315],{},"    2. Analysis of market sentiment (positive, negative, neutral)\n",[844,1317,1318],{"class":846,"line":180},[844,1319,1320],{},"    3. Potential impact on stock price\n",[844,1322,1323],{"class":846,"line":174},[844,1324,1325],{},"    4. Key financial metrics mentioned\n",[844,1327,1328],{"class":846,"line":304},[844,1329,1330],{},"    5. Any risks or opportunities identified\n",[844,1332,1333],{"class":846,"line":155},[844,1334,1305],{},[844,1336,1337],{"class":846,"line":21},[844,1338,1339],{},"    Make your analysis concise, factual, and focused on information that would be relevant to investors.\n",[844,1341,1342],{"class":846,"line":23},[844,1343,1305],{},[844,1345,1346],{"class":846,"line":136},[844,1347,1348],{},"    Search Results:\n",[844,1350,1351],{"class":846,"line":22},[844,1352,1353],{},"    ${contentText}\n",[844,1355,1356],{"class":846,"line":1161},[844,1357,1358],{},"  `\n",[844,1360,1361],{"class":846,"line":24},[844,1362,860],{"emptyLinePlaceholder":859},[844,1364,1365],{"class":846,"line":297},[844,1366,911],{},[844,1368,1369],{"class":846,"line":57},[844,1370,1371],{},"    const response = await openai.chat.completions.create({\n",[844,1373,1374],{"class":846,"line":94},[844,1375,1376],{},"      model: 'gpt-4',\n",[844,1378,1379],{"class":846,"line":72},[844,1380,1381],{},"      messages: [\n",[844,1383,1384],{"class":846,"line":100},[844,1385,1386],{},"        {\n",[844,1388,1389],{"class":846,"line":54},[844,1390,1391],{},"          role: 'system',\n",[844,1393,1394],{"class":846,"line":25},[844,1395,1396],{},"          content:\n",[844,1398,1399],{"class":846,"line":1205},[844,1400,1401],{},"            'You are a financial analyst assistant that helps analyze stock information from web search results.',\n",[844,1403,1404],{"class":846,"line":51},[844,1405,1406],{},"        },\n",[844,1408,1409],{"class":846,"line":341},[844,1410,1411],{},"        { role: 'user', content: prompt },\n",[844,1413,1414],{"class":846,"line":172},[844,1415,1416],{},"      ],\n",[844,1418,1419],{"class":846,"line":290},[844,1420,1421],{},"      temperature: 0.2, // Lower temperature for more factual responses\n",[844,1423,1424],{"class":846,"line":166},[844,1425,1426],{},"      max_tokens: 1500,\n",[844,1428,1429],{"class":846,"line":278},[844,1430,948],{},[844,1432,1433],{"class":846,"line":264},[844,1434,860],{"emptyLinePlaceholder":859},[844,1436,1438],{"class":846,"line":1437},46,[844,1439,1440],{},"    return response.choices[0].message.content\n",[844,1442,1443],{"class":846,"line":254},[844,1444,997],{},[844,1446,1448],{"class":846,"line":1447},48,[844,1449,1450],{},"    console.error('Error generating AI summary:', error)\n",[844,1452,1454],{"class":846,"line":1453},49,[844,1455,1007],{},[844,1457,1458],{"class":846,"line":599},[844,1459,1012],{},[844,1461,1463],{"class":846,"line":1462},51,[844,1464,1017],{},[826,1466,1468],{"id":1467},"_5-putting-it-all-together","5. Putting It All Together",[758,1470,1471],{},"Finally, let's create the main function that ties everything together:",[834,1473,1475],{"className":836,"code":1474,"language":838,"meta":839,"style":839},"async function analyzeStock(stockSymbol) {\n  try {\n    console.log(`Analyzing stock: ${stockSymbol}...`)\n\n    // Step 1: Search for recent information about the stock\n    const searchQuery = `${stockSymbol} stock news financial analysis recent developments`\n    const searchResults = await searchWeb(searchQuery, 8)\n\n    // Step 2: Fetch content from search results\n    const contents = await fetchContentsFromSearchResults(searchResults)\n\n    // Step 3: Generate AI summary and analysis\n    const analysis = await summarizeWithAI(stockSymbol, contents)\n\n    return {\n      stockSymbol,\n      searchResults,\n      analysis,\n    }\n  } catch (error) {\n    console.error(`Error analyzing stock ${stockSymbol}:`, error)\n    throw error\n  }\n}\n\n// Example usage\nanalyzeStock('AAPL')\n  .then((result) => {\n    console.log('Analysis complete:')\n    console.log(result.analysis)\n  })\n  .catch((error) => {\n    console.error('Analysis failed:', error)\n  })\n",[841,1476,1477,1482,1486,1491,1495,1500,1505,1510,1514,1519,1524,1528,1533,1538,1542,1547,1552,1557,1562,1566,1570,1575,1579,1583,1587,1591,1596,1601,1606,1611,1616,1621,1626,1631],{"__ignoreMap":839},[844,1478,1479],{"class":846,"line":4},[844,1480,1481],{},"async function analyzeStock(stockSymbol) {\n",[844,1483,1484],{"class":846,"line":708},[844,1485,911],{},[844,1487,1488],{"class":846,"line":110},[844,1489,1490],{},"    console.log(`Analyzing stock: ${stockSymbol}...`)\n",[844,1492,1493],{"class":846,"line":69},[844,1494,860],{"emptyLinePlaceholder":859},[844,1496,1497],{"class":846,"line":132},[844,1498,1499],{},"    // Step 1: Search for recent information about the stock\n",[844,1501,1502],{"class":846,"line":78},[844,1503,1504],{},"    const searchQuery = `${stockSymbol} stock news financial analysis recent developments`\n",[844,1506,1507],{"class":846,"line":162},[844,1508,1509],{},"    const searchResults = await searchWeb(searchQuery, 8)\n",[844,1511,1512],{"class":846,"line":66},[844,1513,860],{"emptyLinePlaceholder":859},[844,1515,1516],{"class":846,"line":63},[844,1517,1518],{},"    // Step 2: Fetch content from search results\n",[844,1520,1521],{"class":846,"line":91},[844,1522,1523],{},"    const contents = await fetchContentsFromSearchResults(searchResults)\n",[844,1525,1526],{"class":846,"line":939},[844,1527,860],{"emptyLinePlaceholder":859},[844,1529,1530],{"class":846,"line":945},[844,1531,1532],{},"    // Step 3: Generate AI summary and analysis\n",[844,1534,1535],{"class":846,"line":388},[844,1536,1537],{},"    const analysis = await summarizeWithAI(stockSymbol, contents)\n",[844,1539,1540],{"class":846,"line":204},[844,1541,860],{"emptyLinePlaceholder":859},[844,1543,1544],{"class":846,"line":960},[844,1545,1546],{},"    return {\n",[844,1548,1549],{"class":846,"line":60},[844,1550,1551],{},"      stockSymbol,\n",[844,1553,1554],{"class":846,"line":312},[844,1555,1556],{},"      searchResults,\n",[844,1558,1559],{"class":846,"line":207},[844,1560,1561],{},"      analysis,\n",[844,1563,1564],{"class":846,"line":337},[844,1565,1189],{},[844,1567,1568],{"class":846,"line":12},[844,1569,997],{},[844,1571,1572],{"class":846,"line":180},[844,1573,1574],{},"    console.error(`Error analyzing stock ${stockSymbol}:`, error)\n",[844,1576,1577],{"class":846,"line":174},[844,1578,1007],{},[844,1580,1581],{"class":846,"line":304},[844,1582,1012],{},[844,1584,1585],{"class":846,"line":155},[844,1586,1017],{},[844,1588,1589],{"class":846,"line":21},[844,1590,860],{"emptyLinePlaceholder":859},[844,1592,1593],{"class":846,"line":23},[844,1594,1595],{},"// Example usage\n",[844,1597,1598],{"class":846,"line":136},[844,1599,1600],{},"analyzeStock('AAPL')\n",[844,1602,1603],{"class":846,"line":22},[844,1604,1605],{},"  .then((result) => {\n",[844,1607,1608],{"class":846,"line":1161},[844,1609,1610],{},"    console.log('Analysis complete:')\n",[844,1612,1613],{"class":846,"line":24},[844,1614,1615],{},"    console.log(result.analysis)\n",[844,1617,1618],{"class":846,"line":297},[844,1619,1620],{},"  })\n",[844,1622,1623],{"class":846,"line":57},[844,1624,1625],{},"  .catch((error) => {\n",[844,1627,1628],{"class":846,"line":94},[844,1629,1630],{},"    console.error('Analysis failed:', error)\n",[844,1632,1633],{"class":846,"line":72},[844,1634,1620],{},[753,1636,1638],{"id":1637},"real-world-application-stock-analysis-dashboard","Real-World Application: Stock Analysis Dashboard",[758,1640,1641],{},"For our company's experimental project, we integrated this functionality into a dashboard that allowed analysts to:",[772,1643,1644,1647,1650,1653],{},[775,1645,1646],{},"Input multiple stock symbols for analysis",[775,1648,1649],{},"Customize the search parameters (time range, focus areas)",[775,1651,1652],{},"Compare AI-generated summaries side by side",[775,1654,1655],{},"Save and track analyses over time to identify trends",[758,1657,1658],{},"The dashboard looked something like this:",[834,1660,1662],{"className":836,"code":1661,"language":838,"meta":839,"style":839},"// React component example (simplified)\nfunction StockAnalysisDashboard() {\n  const [stocks, setStocks] = useState([])\n  const [loading, setLoading] = useState({})\n  const [analyses, setAnalyses] = useState({})\n\n  const addStock = (symbol) => {\n    if (!stocks.includes(symbol)) {\n      setStocks([...stocks, symbol])\n      analyzeStockAndUpdateState(symbol)\n    }\n  }\n\n  const analyzeStockAndUpdateState = async (symbol) => {\n    setLoading((prev) => ({ ...prev, [symbol]: true }))\n    try {\n      const result = await analyzeStock(symbol)\n      setAnalyses((prev) => ({ ...prev, [symbol]: result }))\n    } catch (error) {\n      console.error(`Error analyzing ${symbol}:`, error)\n    } finally {\n      setLoading((prev) => ({ ...prev, [symbol]: false }))\n    }\n  }\n\n  return (\n    \u003Cdiv className=\"dashboard\">\n      \u003Ch1>Stock Analysis Dashboard\u003C/h1>\n\n      \u003Cdiv className=\"stock-input\">\n        \u003Cinput\n          type=\"text\"\n          placeholder=\"Enter stock symbol (e.g., AAPL)\"\n          onKeyPress={(e) => e.key === 'Enter' && addStock(e.target.value)}\n        />\n      \u003C/div>\n\n      \u003Cdiv className=\"stock-analyses\">\n        {stocks.map((symbol) => (\n          \u003Cdiv key={symbol} className=\"stock-card\">\n            \u003Ch2>{symbol}\u003C/h2>\n            {loading[symbol] ? (\n              \u003Cp>Loading analysis...\u003C/p>\n            ) : analyses[symbol] ? (\n              \u003Cdiv>\n                \u003Ch3>AI Analysis\u003C/h3>\n                \u003Cdiv className=\"analysis-content\">{analyses[symbol].analysis}\u003C/div>\n                \u003Ch3>Sources\u003C/h3>\n                \u003Cul>\n                  {analyses[symbol].searchResults.map((result, i) => (\n                    \u003Cli key={i}>\n                      \u003Ca href={result.link} target=\"_blank\" rel=\"noopener noreferrer\">\n                        {result.title}\n                      \u003C/a>\n                    \u003C/li>\n                  ))}\n                \u003C/ul>\n              \u003C/div>\n            ) : (\n              \u003Cp>No analysis available\u003C/p>\n            )}\n          \u003C/div>\n        ))}\n      \u003C/div>\n    \u003C/div>\n  )\n}\n",[841,1663,1664,1669,1674,1679,1684,1689,1693,1698,1703,1708,1713,1717,1721,1725,1730,1735,1740,1745,1750,1755,1760,1765,1770,1774,1778,1782,1787,1792,1797,1801,1806,1811,1816,1821,1826,1831,1836,1840,1845,1850,1855,1860,1865,1870,1875,1880,1885,1890,1895,1900,1905,1910,1915,1920,1925,1931,1936,1941,1947,1953,1958,1964,1970,1975,1979,1985,1990],{"__ignoreMap":839},[844,1665,1666],{"class":846,"line":4},[844,1667,1668],{},"// React component example (simplified)\n",[844,1670,1671],{"class":846,"line":708},[844,1672,1673],{},"function StockAnalysisDashboard() {\n",[844,1675,1676],{"class":846,"line":110},[844,1677,1678],{},"  const [stocks, setStocks] = useState([])\n",[844,1680,1681],{"class":846,"line":69},[844,1682,1683],{},"  const [loading, setLoading] = useState({})\n",[844,1685,1686],{"class":846,"line":132},[844,1687,1688],{},"  const [analyses, setAnalyses] = useState({})\n",[844,1690,1691],{"class":846,"line":78},[844,1692,860],{"emptyLinePlaceholder":859},[844,1694,1695],{"class":846,"line":162},[844,1696,1697],{},"  const addStock = (symbol) => {\n",[844,1699,1700],{"class":846,"line":66},[844,1701,1702],{},"    if (!stocks.includes(symbol)) {\n",[844,1704,1705],{"class":846,"line":63},[844,1706,1707],{},"      setStocks([...stocks, symbol])\n",[844,1709,1710],{"class":846,"line":91},[844,1711,1712],{},"      analyzeStockAndUpdateState(symbol)\n",[844,1714,1715],{"class":846,"line":939},[844,1716,1189],{},[844,1718,1719],{"class":846,"line":945},[844,1720,1012],{},[844,1722,1723],{"class":846,"line":388},[844,1724,860],{"emptyLinePlaceholder":859},[844,1726,1727],{"class":846,"line":204},[844,1728,1729],{},"  const analyzeStockAndUpdateState = async (symbol) => {\n",[844,1731,1732],{"class":846,"line":960},[844,1733,1734],{},"    setLoading((prev) => ({ ...prev, [symbol]: true }))\n",[844,1736,1737],{"class":846,"line":60},[844,1738,1739],{},"    try {\n",[844,1741,1742],{"class":846,"line":312},[844,1743,1744],{},"      const result = await analyzeStock(symbol)\n",[844,1746,1747],{"class":846,"line":207},[844,1748,1749],{},"      setAnalyses((prev) => ({ ...prev, [symbol]: result }))\n",[844,1751,1752],{"class":846,"line":337},[844,1753,1754],{},"    } catch (error) {\n",[844,1756,1757],{"class":846,"line":12},[844,1758,1759],{},"      console.error(`Error analyzing ${symbol}:`, error)\n",[844,1761,1762],{"class":846,"line":180},[844,1763,1764],{},"    } finally {\n",[844,1766,1767],{"class":846,"line":174},[844,1768,1769],{},"      setLoading((prev) => ({ ...prev, [symbol]: false }))\n",[844,1771,1772],{"class":846,"line":304},[844,1773,1189],{},[844,1775,1776],{"class":846,"line":155},[844,1777,1012],{},[844,1779,1780],{"class":846,"line":21},[844,1781,860],{"emptyLinePlaceholder":859},[844,1783,1784],{"class":846,"line":23},[844,1785,1786],{},"  return (\n",[844,1788,1789],{"class":846,"line":136},[844,1790,1791],{},"    \u003Cdiv className=\"dashboard\">\n",[844,1793,1794],{"class":846,"line":22},[844,1795,1796],{},"      \u003Ch1>Stock Analysis Dashboard\u003C/h1>\n",[844,1798,1799],{"class":846,"line":1161},[844,1800,860],{"emptyLinePlaceholder":859},[844,1802,1803],{"class":846,"line":24},[844,1804,1805],{},"      \u003Cdiv className=\"stock-input\">\n",[844,1807,1808],{"class":846,"line":297},[844,1809,1810],{},"        \u003Cinput\n",[844,1812,1813],{"class":846,"line":57},[844,1814,1815],{},"          type=\"text\"\n",[844,1817,1818],{"class":846,"line":94},[844,1819,1820],{},"          placeholder=\"Enter stock symbol (e.g., AAPL)\"\n",[844,1822,1823],{"class":846,"line":72},[844,1824,1825],{},"          onKeyPress={(e) => e.key === 'Enter' && addStock(e.target.value)}\n",[844,1827,1828],{"class":846,"line":100},[844,1829,1830],{},"        />\n",[844,1832,1833],{"class":846,"line":54},[844,1834,1835],{},"      \u003C/div>\n",[844,1837,1838],{"class":846,"line":25},[844,1839,860],{"emptyLinePlaceholder":859},[844,1841,1842],{"class":846,"line":1205},[844,1843,1844],{},"      \u003Cdiv className=\"stock-analyses\">\n",[844,1846,1847],{"class":846,"line":51},[844,1848,1849],{},"        {stocks.map((symbol) => (\n",[844,1851,1852],{"class":846,"line":341},[844,1853,1854],{},"          \u003Cdiv key={symbol} className=\"stock-card\">\n",[844,1856,1857],{"class":846,"line":172},[844,1858,1859],{},"            \u003Ch2>{symbol}\u003C/h2>\n",[844,1861,1862],{"class":846,"line":290},[844,1863,1864],{},"            {loading[symbol] ? (\n",[844,1866,1867],{"class":846,"line":166},[844,1868,1869],{},"              \u003Cp>Loading analysis...\u003C/p>\n",[844,1871,1872],{"class":846,"line":278},[844,1873,1874],{},"            ) : analyses[symbol] ? (\n",[844,1876,1877],{"class":846,"line":264},[844,1878,1879],{},"              \u003Cdiv>\n",[844,1881,1882],{"class":846,"line":1437},[844,1883,1884],{},"                \u003Ch3>AI Analysis\u003C/h3>\n",[844,1886,1887],{"class":846,"line":254},[844,1888,1889],{},"                \u003Cdiv className=\"analysis-content\">{analyses[symbol].analysis}\u003C/div>\n",[844,1891,1892],{"class":846,"line":1447},[844,1893,1894],{},"                \u003Ch3>Sources\u003C/h3>\n",[844,1896,1897],{"class":846,"line":1453},[844,1898,1899],{},"                \u003Cul>\n",[844,1901,1902],{"class":846,"line":599},[844,1903,1904],{},"                  {analyses[symbol].searchResults.map((result, i) => (\n",[844,1906,1907],{"class":846,"line":1462},[844,1908,1909],{},"                    \u003Cli key={i}>\n",[844,1911,1912],{"class":846,"line":159},[844,1913,1914],{},"                      \u003Ca href={result.link} target=\"_blank\" rel=\"noopener noreferrer\">\n",[844,1916,1917],{"class":846,"line":459},[844,1918,1919],{},"                        {result.title}\n",[844,1921,1922],{"class":846,"line":269},[844,1923,1924],{},"                      \u003C/a>\n",[844,1926,1928],{"class":846,"line":1927},55,[844,1929,1930],{},"                    \u003C/li>\n",[844,1932,1933],{"class":846,"line":235},[844,1934,1935],{},"                  ))}\n",[844,1937,1938],{"class":846,"line":408},[844,1939,1940],{},"                \u003C/ul>\n",[844,1942,1944],{"class":846,"line":1943},58,[844,1945,1946],{},"              \u003C/div>\n",[844,1948,1950],{"class":846,"line":1949},59,[844,1951,1952],{},"            ) : (\n",[844,1954,1955],{"class":846,"line":320},[844,1956,1957],{},"              \u003Cp>No analysis available\u003C/p>\n",[844,1959,1961],{"class":846,"line":1960},61,[844,1962,1963],{},"            )}\n",[844,1965,1967],{"class":846,"line":1966},62,[844,1968,1969],{},"          \u003C/div>\n",[844,1971,1972],{"class":846,"line":250},[844,1973,1974],{},"        ))}\n",[844,1976,1977],{"class":846,"line":301},[844,1978,1835],{},[844,1980,1982],{"class":846,"line":1981},65,[844,1983,1984],{},"    \u003C/div>\n",[844,1986,1987],{"class":846,"line":85},[844,1988,1989],{},"  )\n",[844,1991,1993],{"class":846,"line":1992},67,[844,1994,1017],{},[753,1996,1998],{"id":1997},"challenges-and-considerations","Challenges and Considerations",[758,2000,2001],{},"While building this tool, we encountered several challenges worth noting:",[826,2003,2005],{"id":2004},"_1-api-rate-limits-and-costs","1. API Rate Limits and Costs",[758,2007,2008],{},"Both SERP APIs and OpenAI's API have rate limits and usage costs. For a production system, you'll need to:",[804,2010,2011,2014,2017],{},[775,2012,2013],{},"Implement caching to avoid redundant searches",[775,2015,2016],{},"Set up usage monitoring and alerts",[775,2018,2019],{},"Consider batch processing for multiple stocks",[826,2021,2023],{"id":2022},"_2-content-extraction-quality","2. Content Extraction Quality",[758,2025,2026],{},"Extracting meaningful content from web pages can be challenging due to:",[804,2028,2029,2032,2035],{},[775,2030,2031],{},"Paywalls on financial news sites",[775,2033,2034],{},"Dynamic content loaded via JavaScript",[775,2036,2037],{},"Varied page structures across different sites",[758,2039,2040],{},"We improved our content extraction by:",[804,2042,2043,2049,2052],{},[775,2044,2045,2046],{},"Using more sophisticated libraries like ",[841,2047,2048],{},"mozilla/readability",[775,2050,2051],{},"Implementing site-specific extractors for common financial news sources",[775,2053,2054],{},"Falling back to meta descriptions when full content wasn't available",[826,2056,2058],{"id":2057},"_3-ensuring-analysis-quality","3. Ensuring Analysis Quality",[758,2060,2061],{},"To improve the quality of AI-generated analyses:",[804,2063,2064,2067,2070],{},[775,2065,2066],{},"We fine-tuned prompts based on feedback from financial analysts",[775,2068,2069],{},"Implemented fact-checking by cross-referencing key claims",[775,2071,2072],{},"Added source attribution to make it clear where information came from",[753,2074,2076],{"id":2075},"conclusion","Conclusion",[758,2078,2079],{},"Combining web search capabilities with AI summarization creates a powerful tool for stock analysis that can save hours of research time and provide more comprehensive insights. The implementation we've outlined here is just a starting point—there are many ways to extend and improve this system.",[758,2081,2082],{},"Some potential extensions include:",[804,2084,2085,2088,2091,2094],{},[775,2086,2087],{},"Adding sentiment analysis specifically tuned for financial news",[775,2089,2090],{},"Incorporating historical stock price data for correlation analysis",[775,2092,2093],{},"Expanding to include social media sentiment from platforms like Twitter/X",[775,2095,2096],{},"Creating alerts for significant news that might impact stock prices",[758,2098,2099],{},"As AI capabilities continue to advance, tools like this will become increasingly sophisticated and valuable for financial analysis and decision-making.",[758,2101,2102],{},"Have you built similar tools or have ideas for improvements? I'd love to hear about your experiences in the comments!",[2104,2105,2106],"style",{},"html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}",{"title":839,"searchDepth":708,"depth":708,"links":2108},[2109,2110,2111,2118,2119,2124],{"id":755,"depth":708,"text":756},{"id":766,"depth":708,"text":767},{"id":820,"depth":708,"text":821,"children":2112},[2113,2114,2115,2116,2117],{"id":828,"depth":110,"text":829},{"id":878,"depth":110,"text":879},{"id":1020,"depth":110,"text":1021},{"id":1210,"depth":110,"text":1211},{"id":1467,"depth":110,"text":1468},{"id":1637,"depth":708,"text":1638},{"id":1997,"depth":708,"text":1998,"children":2120},[2121,2122,2123],{"id":2004,"depth":110,"text":2005},{"id":2022,"depth":110,"text":2023},{"id":2057,"depth":110,"text":2058},{"id":2075,"depth":708,"text":2076},"How to combine OpenAI and SERP API to create a powerful web search and AI summary tool for stock analysis and research","md",{"date":2128,"tags":2129,"category":2132,"topics":2133,"published":859,"featured":859},"4th May 2025",[2130,2131],"AI","Engineering","ai-native-systems",[2134,2135],"ai-native","execution","/blogs/en/openai-serp-api-web-search-ai-summary",{"title":748,"description":2125},"blogs/en/10.openai-serp-api-web-search-ai-summary","Ay_c7-5AsRA5SMXw-8dNHUt0vTmqy1QCth1jmpAL8aA",{"id":2141,"title":2142,"body":2143,"description":2456,"extension":2126,"meta":2457,"navigation":859,"ogImage":2459,"path":2465,"seo":2466,"stem":2467,"__hash__":2468},"en/blogs/en/11.cloud-based-scraping-solutions.md","Web Scraping to the Cloud",{"type":750,"value":2144,"toc":2441},[2145,2149,2152,2159,2162,2165,2169,2173,2184,2193,2196,2200,2209,2212,2215,2219,2227,2230,2260,2263,2267,2270,2312,2314,2318,2321,2325,2328,2335,2339,2346,2349,2351,2355,2358,2400,2407,2409,2413,2416,2433,2436,2439],[753,2146,2148],{"id":2147},"the-problem-i-was-trying-to-solve","The Problem I Was Trying to Solve",[758,2150,2151],{},"Last month, I was working on a project that required scraping data from several JavaScript-heavy websites. I started with the usual suspects—Puppeteer and Playwright running locally—but quickly hit a wall. My laptop fans were constantly spinning, memory usage was through the roof, and scaling beyond a few concurrent requests seemed impossible.",[758,2153,2154,2155],{},"I thought to myself: ",[2156,2157,2158],"em",{},"\"There has to be a better way to do this without turning my MacBook into a space heater.\"",[758,2160,2161],{},"So I dove into researching cloud-based alternatives that could handle the heavy lifting of browser automation while keeping my local environment light. Here's what I discovered after spending a week testing different solutions.",[2163,2164],"hr",{},[753,2166,2168],{"id":2167},"cloud-based-headless-browser-services-i-tested","Cloud-Based Headless Browser Services I Tested",[826,2170,2172],{"id":2171},"apify-the-all-in-one-platform","Apify: The All-in-One Platform",[758,2174,2175,2176,2183],{},"I started with ",[2177,2178,2182],"a",{"href":2179,"rel":2180},"https://apify.com",[2181],"nofollow","Apify",", which feels like the \"AWS of web scraping.\" Their cloud platform offers pre-built \"Actors\" (think serverless functions for scraping) with everything bundled in—proxies, scheduling, and storage.",[758,2185,2186,2187,2192],{},"What I liked most was how I could take my existing Node.js crawler code, make minimal changes, and deploy it as a serverless Actor. The ",[2177,2188,2191],{"href":2189,"rel":2190},"https://crawlee.dev/",[2181],"Crawlee"," library they provide made the transition surprisingly smooth.",[758,2194,2195],{},"The dashboard gives you visibility into runs, memory usage, and logs—which was a huge upgrade from my console.log debugging sessions locally.",[826,2197,2199],{"id":2198},"browserlessio-when-you-just-need-the-browser","Browserless.io: When You Just Need the Browser",[758,2201,2202,2203,2208],{},"Next, I tried ",[2177,2204,2207],{"href":2205,"rel":2206},"https://browserless.io",[2181],"Browserless.io",", which takes a more focused approach. It's essentially a hosted headless Chrome and Playwright environment that you can access through a REST API.",[758,2210,2211],{},"The best part? I could keep most of my existing Puppeteer code and just point it to their service instead of launching a local browser. Their API handles all the proxy rotation and CAPTCHA solving behind the scenes.",[758,2213,2214],{},"For teams already invested in Puppeteer/Playwright workflows, this felt like the path of least resistance.",[826,2216,2218],{"id":2217},"scrapingbee-the-one-liner-solution","ScrapingBee: The One-Liner Solution",[758,2220,2221,2226],{},[2177,2222,2225],{"href":2223,"rel":2224},"https://scrapingbee.com",[2181],"ScrapingBee"," took a different approach that I really appreciated on days when I just wanted results without fussing with browser automation code.",[758,2228,2229],{},"Their API is dead simple—send an HTTP request with your target URL, and get back fully rendered HTML. No need to manage browser instances, handle JavaScript execution, or worry about being blocked.",[834,2231,2233],{"className":836,"code":2232,"language":838,"meta":839,"style":839},"// This is literally all it takes\nconst response = await fetch(\n  `https://app.scrapingbee.com/api/v1/?api_key=${apiKey}&url=${targetUrl}&render_js=true`,\n)\nconst html = await response.text()\n",[841,2234,2235,2240,2245,2250,2255],{"__ignoreMap":839},[844,2236,2237],{"class":846,"line":4},[844,2238,2239],{},"// This is literally all it takes\n",[844,2241,2242],{"class":846,"line":708},[844,2243,2244],{},"const response = await fetch(\n",[844,2246,2247],{"class":846,"line":110},[844,2248,2249],{},"  `https://app.scrapingbee.com/api/v1/?api_key=${apiKey}&url=${targetUrl}&render_js=true`,\n",[844,2251,2252],{"class":846,"line":69},[844,2253,2254],{},")\n",[844,2256,2257],{"class":846,"line":132},[844,2258,2259],{},"const html = await response.text()\n",[758,2261,2262],{},"I found myself reaching for this when I needed quick results without the cognitive overhead of browser automation.",[826,2264,2266],{"id":2265},"other-services-worth-mentioning","Other Services Worth Mentioning",[758,2268,2269],{},"I also briefly tested:",[804,2271,2272,2282,2292,2302],{},[775,2273,2274,2281],{},[778,2275,2276],{},[2177,2277,2280],{"href":2278,"rel":2279},"https://zyte.com",[2181],"Zyte API"," (formerly Splash): Great for JavaScript-heavy pages with their anti-ban technology",[775,2283,2284,2291],{},[778,2285,2286],{},[2177,2287,2290],{"href":2288,"rel":2289},"https://phantombuster.com",[2181],"PhantomBuster",": Excellent if you need to automate workflows beyond just scraping",[775,2293,2294,2301],{},[778,2295,2296],{},[2177,2297,2300],{"href":2298,"rel":2299},"https://scraping-bot.io",[2181],"ScrapingBot",": Solid option with real browsers to evade detection",[775,2303,2304,2311],{},[778,2305,2306],{},[2177,2307,2310],{"href":2308,"rel":2309},"https://zyte.com/scrapy-cloud/",[2181],"Scrapy Cloud",": Perfect if you're already using Scrapy",[2163,2313],{},[753,2315,2317],{"id":2316},"the-serverless-route-diy-cloud-scraping","The Serverless Route: DIY Cloud Scraping",[758,2319,2320],{},"For some projects, I wanted more control over the execution environment while still avoiding local resource constraints. That's when I explored deploying headless browsers in serverless functions:",[826,2322,2324],{"id":2323},"aws-lambda-experiment","AWS Lambda Experiment",[758,2326,2327],{},"I packaged Puppeteer with headless Chrome in Lambda Layers and was surprised by how well it worked. The setup was more involved than using a dedicated service, but the per-execution pricing made sense for my sporadic scraping needs.",[758,2329,2330,2331,2334],{},"The key insight was using ",[841,2332,2333],{},"chrome-aws-lambda"," package, which provides a properly compiled version of Chromium that stays under Lambda's size limits.",[826,2336,2338],{"id":2337},"google-cloud-functions-experience","Google Cloud Functions Experience",[758,2340,2341,2342,2345],{},"Google Cloud Functions (2nd gen) was even easier to set up since it already includes many system packages needed for headless Chrome. I just needed to use ",[841,2343,2344],{},"puppeteer-core"," to keep the deployment size manageable.",[758,2347,2348],{},"This approach gave me the best balance of control and scalability for projects where I needed custom logic around the scraping process.",[2163,2350],{},[753,2352,2354],{"id":2353},"what-i-learned-and-my-recommendations","What I Learned and My Recommendations",[758,2356,2357],{},"After weeks of testing, here's my practical advice if you're facing similar challenges:",[804,2359,2360,2369,2378,2386],{},[775,2361,2362,2365,2366,2368],{},[778,2363,2364],{},"If you're just getting started",": Try ",[778,2367,2225],{}," first. The simplicity of a single API call will save you hours of setup time.",[775,2370,2371,2374,2375,2377],{},[778,2372,2373],{},"If you have existing Puppeteer/Playwright code",": ",[778,2376,2207],{}," offers the smoothest transition with minimal code changes.",[775,2379,2380,2374,2383,2385],{},[778,2381,2382],{},"For large-scale, complex projects",[778,2384,2182],{}," provides the most complete ecosystem with scheduling, storage, and proxy management built in.",[775,2387,2388,2391,2392,2395,2396,2399],{},[778,2389,2390],{},"If you're comfortable with cloud services",": Deploying to ",[778,2393,2394],{},"AWS Lambda"," or ",[778,2397,2398],{},"Google Cloud Functions"," gives you the most control at potentially lower costs, especially for intermittent workloads.",[758,2401,2402,2403,2406],{},"The biggest lesson I learned? ",[778,2404,2405],{},"Don't let infrastructure concerns limit your data collection projects."," These cloud solutions have matured to the point where there's rarely a good reason to run resource-intensive browsers on your local machine.",[2163,2408],{},[753,2410,2412],{"id":2411},"whats-next-for-my-scraping-projects","What's Next for My Scraping Projects",[758,2414,2415],{},"I've settled on a hybrid approach for now:",[804,2417,2418,2423,2428],{},[775,2419,2420,2422],{},[778,2421,2225],{}," for quick, one-off scraping tasks",[775,2424,2425,2427],{},[778,2426,2182],{}," for scheduled data collection jobs that run regularly",[775,2429,2430,2432],{},[778,2431,2394],{}," for specialized scrapers that need custom logic",[758,2434,2435],{},"In my next post, I'll share some actual code examples showing how I integrated these services into a unified data pipeline that feeds into my analytics system.",[758,2437,2438],{},"Have you tried any of these services? Or do you have other cloud scraping solutions you'd recommend? Let me know in the comments!",[2104,2440,2106],{},{"title":839,"searchDepth":708,"depth":708,"links":2442},[2443,2444,2450,2454,2455],{"id":2147,"depth":708,"text":2148},{"id":2167,"depth":708,"text":2168,"children":2445},[2446,2447,2448,2449],{"id":2171,"depth":110,"text":2172},{"id":2198,"depth":110,"text":2199},{"id":2217,"depth":110,"text":2218},{"id":2265,"depth":110,"text":2266},{"id":2316,"depth":708,"text":2317,"children":2451},[2452,2453],{"id":2323,"depth":110,"text":2324},{"id":2337,"depth":110,"text":2338},{"id":2353,"depth":708,"text":2354},{"id":2411,"depth":708,"text":2412},"My exploration of cloud-based alternatives for headless browser scraping that don't require running heavy services like Puppeteer or Playwright locally",{"date":2458,"image":2459,"alt":2460,"tags":2461,"category":2463,"topics":2464,"published":859},"5th May 2025","/blogs-img/blog4.jpg","Cloud-Based Web Scraping Solutions",[2131,2462],"Productivity","product-execution",[2135,2134],"/blogs/en/cloud-based-scraping-solutions",{"title":2142,"description":2456},"blogs/en/11.cloud-based-scraping-solutions","5Io7AGIrf7174diw9MZRth9HSfLD3gvpZoyPsH8OyZQ",{"id":2470,"title":2471,"body":2472,"description":3008,"extension":2126,"meta":3009,"navigation":859,"ogImage":3011,"path":3018,"seo":3019,"stem":3020,"__hash__":3021},"en/blogs/en/12.naval-ravikant-specific-knowledge-responsibility-assets.md","What I Learned from The Almanack of Naval Ravikant - Specific Knowledge, Responsibility, and Assets",{"type":750,"value":2473,"toc":2984},[2474,2481,2484,2487,2492,2495,2499,2502,2505,2525,2528,2531,2535,2538,2541,2544,2558,2561,2565,2568,2571,2577,2581,2584,2591,2617,2620,2624,2627,2630,2633,2637,2640,2643,2654,2657,2660,2664,2667,2670,2677,2681,2684,2711,2714,2718,2721,2726,2729,2733,2736,2739,2742,2745,2771,2777,2781,2784,2787,2801,2804,2807,2833,2837,2840,2851,2854,2858,2861,2865,2868,2871,2882,2885,2889,2892,2895,2906,2909,2913,2916,2921,2924,2927,2931,2934,2937,2940,2951,2954,2958,2964,2967,2978,2981],[758,2475,2476,2477,2480],{},"I've been thinking a lot about leverage lately—not the \"work smarter\" kind, but the structural kind. The kind that lets you create disproportionate value without burning out. That's what led me back to ",[2156,2478,2479],{},"The Almanack of Naval Ravikant",".",[758,2482,2483],{},"I finally finished it, and I get why it shows up on so many \"re-read every year\" lists. The book isn't a typical business playbook. It's closer to a set of mental models about how to live well—and how to build wealth without selling your soul (or your entire calendar).",[758,2485,2486],{},"Plenty of ideas in the book are memorable: leverage, judgment, compounding, peace, reading, long-term games. But if I compress my personal takeaway into one sentence, it's this:",[758,2488,2489],{},[778,2490,2491],{},"Build my specific knowledge, take more responsibility, and build assets.",[758,2493,2494],{},"That trio feels like a practical roadmap—not motivational, not vague—just a clean loop I can execute on for years.",[753,2496,2498],{"id":2497},"why-this-triad-matters-to-me","Why this triad matters to me",[758,2500,2501],{},"In the past, I've often approached growth like a checklist: learn a new framework, ship another project, read another book, do another \"productive\" thing. It works… until it doesn't. You end up with motion but not momentum.",[758,2503,2504],{},"This book pushed me to look at the deeper structure underneath momentum:",[804,2506,2507,2513,2519],{},[775,2508,2509,2512],{},[778,2510,2511],{},"Specific knowledge"," gives you an edge that isn't easily copied.",[775,2514,2515,2518],{},[778,2516,2517],{},"Responsibility"," gives you the right to make decisions and capture upside.",[775,2520,2521,2524],{},[778,2522,2523],{},"Assets"," give you leverage that can compound while you sleep.",[758,2526,2527],{},"It's not three separate tips. It's one system.",[758,2529,2530],{},"You can have skills but no autonomy. You can have autonomy but nothing scalable. You can build things but never develop taste and judgment. This triad closes the loop.",[753,2532,2534],{"id":2533},"_1-build-specific-knowledge-the-kind-that-cant-be-cloned","1) Build specific knowledge: the kind that can't be cloned",[758,2536,2537],{},"The phrase \"specific knowledge\" sounds like a productivity buzzword until you really sit with it.",[758,2539,2540],{},"Specific knowledge is not what most of us call \"being smart.\" It's not credentials. It's not being good at interviews. It's not even being \"a strong engineer\" in a generic sense.",[758,2542,2543],{},"Specific knowledge is the combination of:",[804,2545,2546,2549,2552,2555],{},[775,2547,2548],{},"real expertise built through obsession,",[775,2550,2551],{},"lived experience,",[775,2553,2554],{},"taste and judgment,",[775,2556,2557],{},"and a set of skills that stack uniquely in one person.",[758,2559,2560],{},"It's the knowledge that makes people say: \"I don't know how you do that, but you make it look obvious.\"",[826,2562,2564],{"id":2563},"the-uncomfortable-truth-it-cant-be-fast-tracked","The uncomfortable truth: it can't be fast-tracked",[758,2566,2567],{},"This is the part I needed to hear: specific knowledge is earned, not obtained.",[758,2569,2570],{},"You don't download it as a course.\nYou don't \"catch up\" in a weekend.\nYou build it through repetition and curiosity over years—often by following what genuinely interests you.",[758,2572,2573,2574],{},"That has a strategic implication: ",[778,2575,2576],{},"I should stop optimizing for what's popular and start optimizing for what I can compound.",[826,2578,2580],{"id":2579},"what-specific-knowledge-looks-like-in-practice","What \"specific knowledge\" looks like in practice",[758,2582,2583],{},"For me, specific knowledge isn't \"coding.\" It's not \"product management.\" It's not \"design.\"",[758,2585,2586,2587,2590],{},"It's the ",[778,2588,2589],{},"intersection"," of multiple skills:",[804,2592,2593,2599,2605,2611],{},[775,2594,2595,2598],{},[778,2596,2597],{},"Technical execution"," - I can build end-to-end, not just spec",[775,2600,2601,2604],{},[778,2602,2603],{},"Product taste"," - I know what \"good\" looks like before building it",[775,2606,2607,2610],{},[778,2608,2609],{},"Simplification"," - I can make complex things feel obvious",[775,2612,2613,2616],{},[778,2614,2615],{},"Execution leadership"," - I can ship, not just propose",[758,2618,2619],{},"That combination is harder to replace than any single skill. And it gets more valuable over time if I keep stacking it.",[826,2621,2623],{"id":2622},"the-standard-im-aiming-for","The standard I'm aiming for",[758,2625,2626],{},"A useful test I'm adopting:",[758,2628,2629],{},"If I disappeared for a month, would my team still ship?\nIf I disappeared for a year, would my product still evolve?\nIf I disappeared forever, would there be something that continues to create value?",[758,2631,2632],{},"Specific knowledge helps me contribute today. But it also helps me design things that survive me.",[753,2634,2636],{"id":2635},"_2-take-more-responsibility-earn-the-right-to-outcomes","2) Take more responsibility: earn the right to outcomes",[758,2638,2639],{},"This might be the most \"simple but not easy\" lesson.",[758,2641,2642],{},"Taking responsibility doesn't mean \"work harder.\" It means:",[804,2644,2645,2648,2651],{},[775,2646,2647],{},"owning outcomes instead of tasks,",[775,2649,2650],{},"being the person who will figure it out,",[775,2652,2653],{},"and accepting that if it fails, it's on you.",[758,2655,2656],{},"That's scary. It also changes everything.",[758,2658,2659],{},"Because responsibility is what creates trust. Trust creates autonomy. Autonomy creates decision-making rights. And decision-making rights are where leverage begins.",[826,2661,2663],{"id":2662},"responsibility-is-how-you-get-leverage-without-asking-for-permission","Responsibility is how you get leverage without asking for permission",[758,2665,2666],{},"A lot of people want more freedom but avoid responsibility.\nThey want upside but don't want blame.\nThey want influence but don't want accountability.",[758,2668,2669],{},"But responsibility is the trade. You don't get the benefits without paying that cost.",[758,2671,2672,2673,2676],{},"When I look back at every meaningful jump I've made in my career, it wasn't because I performed well at my current scope. It's because I ",[778,2674,2675],{},"expanded the scope",", then grew into it.",[826,2678,2680],{"id":2679},"the-responsibility-ladder-im-trying-to-climb","The responsibility ladder I'm trying to climb",[758,2682,2683],{},"I'm trying to be intentional about moving up this ladder:",[772,2685,2686,2691,2696,2701,2706],{},[775,2687,2688],{},[778,2689,2690],{},"Do tasks well",[775,2692,2693],{},[778,2694,2695],{},"Own a project end-to-end",[775,2697,2698],{},[778,2699,2700],{},"Own a system (and its reliability over time)",[775,2702,2703],{},[778,2704,2705],{},"Own an outcome (business metrics, customer value, strategy execution)",[775,2707,2708],{},[778,2709,2710],{},"Own a portfolio of bets (where judgment matters most)",[758,2712,2713],{},"At the top, performance matters less than judgment. You're paid for being right, not busy.",[826,2715,2717],{"id":2716},"a-personal-rule","A personal rule",[758,2719,2720],{},"I'm adopting a rule that feels very \"Naval\":",[758,2722,2723],{},[778,2724,2725],{},"Don't just ship work. Ship responsibility.",[758,2727,2728],{},"If I'm going to spend energy, I want it to increase my surface area of ownership, not just fill my day.",[753,2730,2732],{"id":2731},"_3-build-assets-create-things-that-compound-without-you","3) Build assets: create things that compound without you",[758,2734,2735],{},"This is the most practical takeaway, because it transforms how I view \"work.\"",[758,2737,2738],{},"If I trade time for money, I reset to zero every morning.\nIf I build assets, I create compounding.",[758,2740,2741],{},"An asset is something that can produce value repeatedly without requiring the same unit of effort each time.",[758,2743,2744],{},"That can be:",[804,2746,2747,2750,2753,2756,2759,2762,2765,2768],{},[775,2748,2749],{},"equity in a business,",[775,2751,2752],{},"code that runs and serves users,",[775,2754,2755],{},"content that attracts an audience,",[775,2757,2758],{},"a dataset,",[775,2760,2761],{},"a distribution channel,",[775,2763,2764],{},"a brand,",[775,2766,2767],{},"a reusable system or platform,",[775,2769,2770],{},"even relationships that create long-term opportunities.",[758,2772,2773,2774],{},"The key is not what kind of asset. The key is: ",[778,2775,2776],{},"does it keep working after I stop?",[826,2778,2780],{"id":2779},"the-shift-from-projects-to-factories","The shift: from \"projects\" to \"factories\"",[758,2782,2783],{},"I'm trying to shift my mindset from building one-off projects to building small factories.",[758,2785,2786],{},"A factory can produce outcomes repeatedly:",[804,2788,2789,2792,2795,2798],{},[775,2790,2791],{},"a personal website that consistently attracts opportunities,",[775,2793,2794],{},"a product that solves a recurring pain,",[775,2796,2797],{},"a library or tool that accelerates future work,",[775,2799,2800],{},"a content engine that makes ideas discoverable.",[758,2802,2803],{},"A project ends. A factory compounds.",[758,2805,2806],{},"For me, that means:",[804,2808,2809,2815,2821,2827],{},[775,2810,2811,2814],{},[778,2812,2813],{},"This blog"," - content that attracts opportunities and builds reputation",[775,2816,2817,2820],{},[778,2818,2819],{},"Drum Next"," - a product that serves users while I sleep",[775,2822,2823,2826],{},[778,2824,2825],{},"Reusable code patterns"," - systems that accelerate future projects",[775,2828,2829,2832],{},[778,2830,2831],{},"Distribution channels"," - RSS, social, networks that amplify reach",[826,2834,2836],{"id":2835},"assets-require-leverage","Assets require leverage",[758,2838,2839],{},"And this is where the triad connects:",[804,2841,2842,2845,2848],{},[775,2843,2844],{},"Specific knowledge helps you build something differentiated.",[775,2846,2847],{},"Responsibility helps you make decisions and move fast.",[775,2849,2850],{},"Assets help you scale output beyond your hours.",[758,2852,2853],{},"It's one flywheel.",[753,2855,2857],{"id":2856},"how-im-turning-this-into-action","How I'm turning this into action",[758,2859,2860],{},"Reading is easy. Integrating is hard. So I wrote down a simple operating plan I can actually follow.",[826,2862,2864],{"id":2863},"step-1-pick-one-specific-knowledge-lane-to-deepen-this-year","Step 1: Pick one \"specific knowledge lane\" to deepen this year",[758,2866,2867],{},"Instead of scattering, I'm choosing a lane that I can compound:",[758,2869,2870],{},"A lane should be:",[804,2872,2873,2876,2879],{},[775,2874,2875],{},"interesting enough that I'll stick with it,",[775,2877,2878],{},"valuable enough that the market rewards it,",[775,2880,2881],{},"and rare enough that it's not instantly commoditized.",[758,2883,2884],{},"Then I ask: what does \"10,000 reps\" look like here?",[826,2886,2888],{"id":2887},"step-2-increase-responsibility-intentionally-without-burning-out","Step 2: Increase responsibility intentionally (without burning out)",[758,2890,2891],{},"This isn't about saying yes to everything. It's about saying yes to the right things.",[758,2893,2894],{},"I'm looking for responsibilities that:",[804,2896,2897,2900,2903],{},[775,2898,2899],{},"increase my decision-making rights,",[775,2901,2902],{},"teach me judgment,",[775,2904,2905],{},"and connect directly to outcomes.",[758,2907,2908],{},"If something only increases my busyness, I want to decline it—no matter how \"productive\" it looks.",[826,2910,2912],{"id":2911},"step-3-convert-effort-into-assets-every-week","Step 3: Convert effort into assets every week",[758,2914,2915],{},"I'm adopting a weekly question:",[758,2917,2918],{},[778,2919,2920],{},"What asset did I build this week?",[758,2922,2923],{},"Not \"what did I do.\" Not \"what meetings did I attend.\"\nWhat did I build that will keep paying me back?",[758,2925,2926],{},"Sometimes that asset will be external (content, product). Sometimes internal (systems, templates, reusable components). But the goal is to keep converting effort into compounding value.",[753,2928,2930],{"id":2929},"the-deeper-lesson-long-term-games-with-long-term-people","The deeper lesson: long-term games with long-term people",[758,2932,2933],{},"Underneath everything, I think the book is pointing to a life strategy:",[758,2935,2936],{},"Choose long-term games.\nPlay with people you trust.\nKeep your mind clear.\nBuild leverage ethically.\nLet compounding do the heavy lifting.",[758,2938,2939],{},"The \"wealth\" part is just one output. The bigger win is that this framework also produces a calmer life:",[804,2941,2942,2945,2948],{},[775,2943,2944],{},"When you build specific knowledge, you stop chasing trends.",[775,2946,2947],{},"When you take responsibility, you stop waiting for permission.",[775,2949,2950],{},"When you build assets, you stop feeling like time is always running out.",[758,2952,2953],{},"You become harder to replace—and less anxious about being replaced.",[753,2955,2957],{"id":2956},"closing","Closing",[758,2959,2960,2961,2963],{},"If I had to summarize what I'm carrying forward from ",[2156,2962,2479],{},", it's not a quote. It's a direction:",[758,2965,2966],{},"I want to become the kind of person who can:",[804,2968,2969,2972,2975],{},[775,2970,2971],{},"develop rare, real skill,",[775,2973,2974],{},"shoulder real responsibility,",[775,2976,2977],{},"and build assets that compound.",[758,2979,2980],{},"Not as a hustle story. As a life design.",[758,2982,2983],{},"And now the real work begins: doing it consistently, quietly, for a long time.",{"title":839,"searchDepth":708,"depth":708,"links":2985},[2986,2987,2992,2997,3001,3006,3007],{"id":2497,"depth":708,"text":2498},{"id":2533,"depth":708,"text":2534,"children":2988},[2989,2990,2991],{"id":2563,"depth":110,"text":2564},{"id":2579,"depth":110,"text":2580},{"id":2622,"depth":110,"text":2623},{"id":2635,"depth":708,"text":2636,"children":2993},[2994,2995,2996],{"id":2662,"depth":110,"text":2663},{"id":2679,"depth":110,"text":2680},{"id":2716,"depth":110,"text":2717},{"id":2731,"depth":708,"text":2732,"children":2998},[2999,3000],{"id":2779,"depth":110,"text":2780},{"id":2835,"depth":110,"text":2836},{"id":2856,"depth":708,"text":2857,"children":3002},[3003,3004,3005],{"id":2863,"depth":110,"text":2864},{"id":2887,"depth":110,"text":2888},{"id":2911,"depth":110,"text":2912},{"id":2929,"depth":708,"text":2930},{"id":2956,"depth":708,"text":2957},"Build specific knowledge, take more responsibility, and build assets. A practical framework for long-term leverage and compounding value.",{"date":3010,"image":3011,"alt":3012,"tags":3013,"category":3016,"topics":3017,"published":859,"featured":859},"5th Jan 2026","/blogs-img/2026-01-04-naval.jpg","The Almanack of Naval Ravikant - Specific Knowledge, Responsibility, and Assets",[3014,3015],"Product","Life","personal-operating-system",[2135],"/blogs/en/naval-ravikant-specific-knowledge-responsibility-assets",{"title":2471,"description":3008},"blogs/en/12.naval-ravikant-specific-knowledge-responsibility-assets","Vv2xK5pVKlKG4BvJhohwjn52IMAGX7qGFFylMLQ-6Ns",{"id":3023,"title":3024,"body":3025,"description":3175,"extension":2126,"meta":3176,"navigation":859,"ogImage":3178,"path":3183,"seo":3184,"stem":3185,"__hash__":3186},"en/blogs/en/13.start-my-ai-native-journey.md","Start My AI Native Journey",{"type":750,"value":3026,"toc":3168},[3027,3030,3033,3040,3044,3047,3050,3056,3060,3063,3066,3069,3072,3075,3081,3085,3088,3094,3100,3106,3112,3116,3119,3122,3125,3129,3132,3152,3159,3162],[758,3028,3029],{},"I've been in tech for a while now — building products, writing code, leading teams. I've seen plenty of \"next big things\" come and go. But this time feels different. AI isn't just another tool in the toolbox. It's changing how I think, how I build, and honestly, how much I can get done in a day.",[758,3031,3032],{},"So I'm going all in. This is the start of my AI Native journey, and I want to share what I've learned so far.",[758,3034,3035],{},[3036,3037],"img",{"alt":3038,"src":3039},"AI Native Journey Begins — stepping into a new world of possibility","/blogs-img/2026-02-12-ai-native-01.png",[753,3041,3043],{"id":3042},"what-do-i-mean-by-ai-native","What Do I Mean by \"AI Native\"?",[758,3045,3046],{},"For me, AI Native means AI isn't an add-on — it's the starting point. Every time I sit down to work, AI is already part of my workflow. Writing code? AI is right there. Researching a new domain? AI helps me learn faster. Drafting a product spec? Same thing.",[758,3048,3049],{},"It's like the shift from desktop to mobile-first. Once you go native, you don't go back.",[758,3051,3052],{},[3036,3053],{"alt":3054,"src":3055},"AI Native vs Traditional — from add-on to starting point","/blogs-img/2026-02-12-ai-native-02.png",[753,3057,3059],{"id":3058},"the-moment-it-clicked","The Moment It Clicked",[758,3061,3062],{},"I'm a head of products who still codes. I love building things with my own hands — there's something deeply satisfying about shipping a feature you architected and implemented yourself.",[758,3064,3065],{},"But let's be real: time is always the bottleneck. Between product strategy, team alignment, and actually writing code, there are never enough hours. That's where AI changed the game for me.",[758,3067,3068],{},"A few months ago, I was tackling a project that would have normally taken me days of focused coding. I decided to go AI-first — not just using autocomplete, but actually collaborating with AI through the entire process. Architecture decisions, implementation, debugging, testing.",[758,3070,3071],{},"What used to take days took hours. And I'm not talking about sloppy, \"good enough\" output. The quality was there. The code was clean. I shipped it with confidence.",[758,3073,3074],{},"That was my moment. I realized this isn't a productivity hack — it's a fundamental shift in what one person can accomplish.",[758,3076,3077],{},[3036,3078],{"alt":3079,"src":3080},"The breakthrough moment — from days to hours","/blogs-img/2026-02-12-ai-native-03.png",[753,3082,3084],{"id":3083},"what-ai-actually-changes-in-my-day-to-day","What AI Actually Changes in My Day-to-Day",[758,3086,3087],{},"Here's what going AI Native looks like in practice:",[758,3089,3090,3093],{},[778,3091,3092],{},"Coding at 2-3x speed."," I'm not exaggerating. When I pair with AI, I move through implementation faster because I spend less time on boilerplate, less time looking up APIs, and less time on the boring parts. I get to focus on the interesting decisions — the architecture, the user experience, the tradeoffs.",[758,3095,3096,3099],{},[778,3097,3098],{},"Learning new things in hours, not weeks."," As a product leader, I need to understand a wide range of technologies and domains. AI has become my go-to learning partner. I can dive into an unfamiliar codebase, a new framework, or a complex technical concept and get up to speed dramatically faster. It's like having a patient, knowledgeable colleague available 24/7.",[758,3101,3102,3105],{},[778,3103,3104],{},"Building things I wouldn't have attempted before."," This is the part that excites me most. There are projects I would have dismissed as \"too much work for one person.\" Now I look at those same projects and think — why not? The barrier to building has dropped so significantly that the limiting factor is no longer time or technical skill. It's imagination.",[758,3107,3108],{},[3036,3109],{"alt":3110,"src":3111},"Three benefits of going AI Native","/blogs-img/2026-02-12-ai-native-04.png",[753,3113,3115],{"id":3114},"why-im-writing-about-this","Why I'm Writing About This",[758,3117,3118],{},"I'm not writing this to convince anyone that AI is the future — if you're reading this, you probably already sense that. I'm writing this because I want to document what it actually looks like to go AI Native as someone who builds products and writes code every day.",[758,3120,3121],{},"There's a lot of hype out there. There's also a lot of skepticism. What's missing is the honest, practical, in-the-trenches perspective. What works? What doesn't? Where does AI genuinely save you time, and where does it slow you down?",[758,3123,3124],{},"That's what I want to explore in this series.",[753,3126,3128],{"id":3127},"whats-next","What's Next",[758,3130,3131],{},"This is the first post in what I'm calling my AI Native Journey. Going forward, I plan to share:",[804,3133,3134,3140,3146],{},[775,3135,3136,3139],{},[778,3137,3138],{},"Real projects I'm building with AI"," — the good, the bad, and the ugly",[775,3141,3142,3145],{},[778,3143,3144],{},"Workflows and tools that actually work"," — not theory, but what I use every day",[775,3147,3148,3151],{},[778,3149,3150],{},"Lessons learned"," — including the mistakes, because that's where the real learning happens",[758,3153,3154,3155,3158],{},"I strongly believe we're at the beginning of something massive. AI doesn't just make us faster — it expands what's possible for individuals and small teams. The people who embrace this shift early and learn to work ",[2156,3156,3157],{},"with"," AI will have an enormous advantage.",[758,3160,3161],{},"I'm going all in. Come along for the ride.",[758,3163,3164],{},[3036,3165],{"alt":3166,"src":3167},"The journey ahead — walking toward the sunrise","/blogs-img/2026-02-12-ai-native-05.png",{"title":839,"searchDepth":708,"depth":708,"links":3169},[3170,3171,3172,3173,3174],{"id":3042,"depth":708,"text":3043},{"id":3058,"depth":708,"text":3059},{"id":3083,"depth":708,"text":3084},{"id":3114,"depth":708,"text":3115},{"id":3127,"depth":708,"text":3128},"I'm going all in on AI Native. Here's what it means, why I'm doing it, and what I've learned so far as a head of products who still codes every day.",{"date":3177,"image":3178,"alt":3179,"tags":3180,"category":2132,"topics":3181,"published":859,"featured":859},"12th Feb 2026","/blogs-img/2026-02-12-ai-native-cover.png","AI Native Journey - hand-drawn illustration of a person starting their AI journey",[2130,2462],[2134,3182],"building-in-public","/blogs/en/start-my-ai-native-journey",{"title":3024,"description":3175},"blogs/en/13.start-my-ai-native-journey","ep5-ecoAv5dS2RIAd-Hyicnc41biitLa-uqdDFeAEvI",{"id":3188,"title":3189,"body":3190,"description":3411,"extension":2126,"meta":3412,"navigation":859,"ogImage":3414,"path":3419,"seo":3420,"stem":3421,"__hash__":3422},"en/blogs/en/14.marriott-timeshare-las-vegas.md","What a Marriott Timeshare Pitch in Las Vegas Taught Me About Broken Sales Culture",{"type":750,"value":3191,"toc":3403},[3192,3195,3199,3202,3205,3225,3228,3234,3238,3241,3244,3247,3250,3256,3260,3263,3268,3294,3297,3303,3307,3310,3316,3322,3328,3334,3338,3341,3344,3347,3350,3353,3359,3363,3366,3369,3389,3396,3398],[758,3193,3194],{},"My wife and I recently traveled to Las Vegas and joined a Marriott Vacation Club presentation. We knew it was a timeshare pitch — they're upfront about that. What we didn't expect was how the experience would sour our entire stay and leave me thinking deeply about what happens when a company's sales culture contradicts its service brand.",[753,3196,3198],{"id":3197},"the-offer-points-fees-and-fine-print","The Offer: Points, Fees, and Fine Print",[758,3200,3201],{},"Here's how Marriott's timeshare system works. You pay a significant upfront cost to purchase \"points.\" These points can be redeemed for stays at Marriott properties around the world. Sounds reasonable so far.",[758,3203,3204],{},"But here's where it gets complicated:",[804,3206,3207,3213,3219],{},[775,3208,3209,3212],{},[778,3210,3211],{},"Annual maintenance fees."," Even after you've paid the upfront cost, you're on the hook for yearly maintenance fees — and these tend to increase over time. You're paying for the privilege of using what you already bought.",[775,3214,3215,3218],{},[778,3216,3217],{},"No clear exit mechanism."," This was the biggest red flag for me. Once you're in, getting out is murky at best. There's no straightforward process to sell back your points or walk away cleanly.",[775,3220,3221,3224],{},[778,3222,3223],{},"Limited transferability."," Want to sell your points to someone else? It's not easy. The resale market for timeshares is notoriously unfavorable to sellers.",[758,3226,3227],{},"The signing package they presented was attractive on the surface — bonus points, discounted rates, exclusive perks. But when I started asking about what happens if we want to leave, the answers got vague. That's never a good sign.",[758,3229,3230],{},[3036,3231],{"alt":3232,"src":3233},"The timeshare model — money flows in, but there's no clear way out","/blogs-img/2026-02-14-marriott-01.png",[753,3235,3237],{"id":3236},"the-moment-everything-changed","The Moment Everything Changed",[758,3239,3240],{},"I can handle a hard sell. I've been in business long enough to understand that sales teams have targets to hit. What I can't handle is disrespect.",[758,3242,3243],{},"When we told our sales representative that we needed time to think about the purchase, the reaction was immediate and uncomfortable. He told us 30 minutes should be \"more than enough\" to discuss it between ourselves and pushed us to decide on the spot.",[758,3245,3246],{},"When we held firm — we wanted to take the materials home and think it through — his demeanor changed completely. The warmth disappeared. The smile dropped. He made it clear, without saying it directly, that we were wasting his time.",[758,3248,3249],{},"This is a Marriott property. A brand built on hospitality and service. The contrast between what Marriott represents and what we experienced in that room was jarring.",[758,3251,3252],{},[3036,3253],{"alt":3254,"src":3255},"The pressure moment — 30 minutes to decide your financial future","/blogs-img/2026-02-14-marriott-02.png",[753,3257,3259],{"id":3258},"the-business-model-profitable-but-at-what-cost","The Business Model: Profitable, But at What Cost?",[758,3261,3262],{},"Let's set emotions aside and look at how this business actually works.",[758,3264,3265],{},[778,3266,3267],{},"The revenue engine is simple and effective:",[772,3269,3270,3276,3282,3288],{},[775,3271,3272,3275],{},[778,3273,3274],{},"Upfront sales"," generate large one-time payments. A single timeshare purchase can run anywhere from $20,000 to $100,000+.",[775,3277,3278,3281],{},[778,3279,3280],{},"Recurring maintenance fees"," create a steady annual revenue stream that grows over time, regardless of whether the owner uses their points.",[775,3283,3284,3287],{},[778,3285,3286],{},"Financing"," adds interest income. Many buyers finance their purchase, meaning Marriott earns on the loan as well.",[775,3289,3290,3293],{},[778,3291,3292],{},"Low exit friction"," keeps owners locked in. Without a clean exit path, most owners continue paying maintenance fees year after year — even when they've stopped using the product.",[758,3295,3296],{},"It's a business model optimized for acquisition and retention through friction, not through value. And that's the fundamental problem.",[758,3298,3299],{},[3036,3300],{"alt":3301,"src":3302},"Four revenue engines of the timeshare business model","/blogs-img/2026-02-14-marriott-03.png",[753,3304,3306],{"id":3305},"where-the-sales-model-breaks-down","Where the Sales Model Breaks Down",[758,3308,3309],{},"High-pressure sales tactics aren't new in the timeshare industry. But they create a specific set of problems:",[758,3311,3312,3315],{},[778,3313,3314],{},"Short-term wins, long-term damage."," A salesperson who pressures a couple into a same-day decision might close the deal. But that couple will associate the brand with that negative experience forever. In the age of online reviews and social media, one bad experience travels far.",[758,3317,3318,3321],{},[778,3319,3320],{},"Misaligned incentives."," When salespeople are compensated primarily on closing rate and deal size, the customer's best interest becomes secondary. The 30-minute pressure tactic isn't about giving customers time to think — it's about preventing them from thinking clearly.",[758,3323,3324,3327],{},[778,3325,3326],{},"Brand erosion."," Marriott has spent decades building trust through consistent service quality. Every high-pressure timeshare pitch chips away at that trust. The vacation club division may generate strong revenue, but at what cost to the broader brand?",[758,3329,3330],{},[3036,3331],{"alt":3332,"src":3333},"Short-term sales wins versus long-term brand damage","/blogs-img/2026-02-14-marriott-04.png",[753,3335,3337],{"id":3336},"a-broader-reflection-when-sales-contradicts-service","A Broader Reflection: When Sales Contradicts Service",[758,3339,3340],{},"This experience got me thinking about a tension that exists in many service-oriented businesses: the gap between the service brand and the sales culture.",[758,3342,3343],{},"Marriott's hotel operations are built on a simple premise — make guests feel welcome, comfortable, and valued. It works. I've stayed at Marriott properties many times and generally had great experiences.",[758,3345,3346],{},"But walk into their vacation club sales room, and you're in a different world. The priorities flip. It's no longer about making you feel valued — it's about closing you before you leave the building.",[758,3348,3349],{},"This isn't unique to Marriott. You see it in banking, insurance, telecommunications, and countless other industries. The service team builds trust; the sales team burns it.",[758,3351,3352],{},"The companies that get this right — the ones that truly earn long-term loyalty — are the ones where sales and service are aligned. Where the sales process itself feels like an extension of the service experience, not a contradiction of it.",[758,3354,3355],{},[3036,3356],{"alt":3357,"src":3358},"Service culture versus sales culture — the gap that breaks trust","/blogs-img/2026-02-14-marriott-05.png",[753,3360,3362],{"id":3361},"what-i-took-away","What I Took Away",[758,3364,3365],{},"I left Las Vegas without buying a timeshare. I also left with a diminished view of a brand I used to respect.",[758,3367,3368],{},"A few things I'll remember from this experience:",[804,3370,3371,3377,3383],{},[775,3372,3373,3376],{},[778,3374,3375],{},"Any product that makes it easy to buy and hard to leave should be approached with extreme caution."," If the exit terms aren't clear, the company is betting you'll stay out of friction, not satisfaction.",[775,3378,3379,3382],{},[778,3380,3381],{},"Pressure to decide quickly is almost never in the buyer's interest."," A good product should withstand a week of consideration. If it can't, ask yourself why.",[775,3384,3385,3388],{},[778,3386,3387],{},"A company's true values show up in its worst moments, not its best."," When a customer says \"I need to think about it,\" the response reveals everything about the culture.",[758,3390,3391,3392,3395],{},"For anyone in the service industry: your sales experience ",[2156,3393,3394],{},"is"," your service experience. Customers don't separate the two. Every touchpoint either builds trust or breaks it. There's no neutral.",[2163,3397],{},[758,3399,3400],{},[2156,3401,3402],{},"Have you had a similar experience with timeshare presentations? I'd love to hear your perspective. Connect with me and share your story.",{"title":839,"searchDepth":708,"depth":708,"links":3404},[3405,3406,3407,3408,3409,3410],{"id":3197,"depth":708,"text":3198},{"id":3236,"depth":708,"text":3237},{"id":3258,"depth":708,"text":3259},{"id":3305,"depth":708,"text":3306},{"id":3336,"depth":708,"text":3337},{"id":3361,"depth":708,"text":3362},"A balanced look at what happened when Marriott's vacation club sales tactics contradicted everything the brand stands for — and what it reveals about the tension between sales culture and service culture.",{"date":3413,"image":3414,"alt":3415,"tags":3416,"category":3417,"topics":3418,"published":859},"14th Feb 2026","/blogs-img/2026-02-14-marriott-cover.png","A couple at a crossroads between hospitality and timeshare sales pressure",[3014,3015],"business-strategy",[3182],"/blogs/en/marriott-timeshare-las-vegas",{"title":3189,"description":3411},"blogs/en/14.marriott-timeshare-las-vegas","3xgQeDFYIZoK_uirOwlQpz_Y67lcPiZPiH3qrf2a8Ik",{"id":3424,"title":3425,"body":3426,"description":3752,"extension":2126,"meta":3753,"navigation":859,"ogImage":3755,"path":3760,"seo":3761,"stem":3762,"__hash__":3763},"en/blogs/en/15.why-i-stopped-writing-and-built-ai-studio.md","Why I Stopped Writing Blog Posts (And Built an AI Studio Instead)",{"type":750,"value":3427,"toc":3743},[3428,3431,3434,3445,3448,3454,3458,3461,3468,3471,3509,3512,3518,3522,3528,3531,3545,3548,3551,3558,3564,3568,3571,3574,3582,3589,3595,3598,3602,3605,3611,3617,3623,3629,3635,3641,3647,3653,3657,3660,3663,3669,3675,3678,3681,3685,3688,3694,3697,3711,3714,3720,3724,3727,3730,3733,3736,3738],[758,3429,3430],{},"I've tried to be a consistent blogger for years. I'd get fired up, write two or three posts, feel great about it — and then life would happen. Work gets busy. Weekends fill up. The draft sitting in my editor starts feeling stale. I'd tell myself I'll get back to it next week, and next week would turn into next month, and next month would turn into silence.",[758,3432,3433],{},"Rinse and repeat. For years.",[758,3435,3436,3437,3440,3441,3444],{},"But last month, I published more blog posts than I had in the previous two years combined. Not because I suddenly found more time or discovered some productivity hack. I stopped trying to ",[2156,3438,3439],{},"write"," blog posts and started ",[2156,3442,3443],{},"building a system"," that turns my ideas into published content.",[758,3446,3447],{},"This is the story of how I went from a chronic blog-quitter to a consistent creator — by approaching the problem like an engineer, not a writer.",[758,3449,3450],{},[3036,3451],{"alt":3452,"src":3453},"The graveyard of abandoned blogs — drafts that never saw the light of day","/blogs-img/2026-02-16-ai-studio-01.png",[753,3455,3457],{"id":3456},"the-real-reason-i-kept-quitting","The Real Reason I Kept Quitting",[758,3459,3460],{},"Let me be honest about what was actually happening. It wasn't a lack of ideas. Ideas were everywhere — a frustrating timeshare pitch in Las Vegas, a realization about how AI was changing my daily work, a hot take on product strategy. I had a Notes app full of half-formed thoughts begging to become posts.",[758,3462,3463,3464,3467],{},"The problem was everything that comes ",[2156,3465,3466],{},"after"," the idea.",[758,3469,3470],{},"Think about what it actually takes to publish a single blog post:",[804,3472,3473,3479,3485,3491,3497,3503],{},[775,3474,3475,3478],{},[778,3476,3477],{},"Writing the draft"," — 1 to 2 hours if the words are flowing, more if they're not",[775,3480,3481,3484],{},[778,3482,3483],{},"Formatting and editing"," — cleaning up markdown, adding headers, making it readable (30 minutes)",[775,3486,3487,3490],{},[778,3488,3489],{},"Creating images"," — cover image, diagrams, illustrations that aren't just stock photos (1 hour, minimum)",[775,3492,3493,3496],{},[778,3494,3495],{},"Translating to Chinese"," — I write for a bilingual audience, so every post needs a natural Chinese version, not a Google Translate job (1 to 2 hours)",[775,3498,3499,3502],{},[778,3500,3501],{},"Publishing"," — frontmatter, image paths, SEO metadata, deploying to the site (30 minutes)",[775,3504,3505,3508],{},[778,3506,3507],{},"Social media"," — writing an X thread, a LinkedIn post, maybe a YouTube script (another 30 minutes)",[758,3510,3511],{},"Add it up: a single blog post is a 5 to 6 hour commitment. As a head of products who still codes, with a full-time job and side projects, those hours simply don't exist on a consistent basis. The math never worked. So I'd write when inspiration and free time happened to collide — which was almost never.",[758,3513,3514],{},[3036,3515],{"alt":3516,"src":3517},"The hidden cost of a blog post — 6 hours from idea to published","/blogs-img/2026-02-16-ai-studio-02.png",[753,3519,3521],{"id":3520},"the-shift-what-if-i-only-did-the-thinking","The Shift: What If I Only Did the Thinking?",[758,3523,3524,3525],{},"The breakthrough wasn't about writing faster. It was a different question entirely: ",[778,3526,3527],{},"what if I only had to do the part that requires my brain?",[758,3529,3530],{},"Every step in that 6-hour pipeline falls into one of two categories:",[772,3532,3533,3539],{},[775,3534,3535,3538],{},[778,3536,3537],{},"Thinking"," — deciding what to say, what angle to take, what personal experience to share, what the reader should take away",[775,3540,3541,3544],{},[778,3542,3543],{},"Execution"," — turning those decisions into formatted text, images, translations, metadata, and deployed pages",[758,3546,3547],{},"I'm the only one who can do the thinking. But the execution? That's process. That's repeatable. That's automatable.",[758,3549,3550],{},"This is where most \"AI for blogging\" advice gets it wrong. They tell you to use ChatGPT to write your posts. That gives you generic content that sounds like everyone else — AI slop. Nobody wants to read that, and I definitely don't want to publish it.",[758,3552,3553,3554,3557],{},"I didn't want AI to think for me. I wanted AI to handle everything ",[2156,3555,3556],{},"around"," the thinking.",[758,3559,3560],{},[3036,3561],{"alt":3562,"src":3563},"Thinking vs. execution — the only part that needs to be human","/blogs-img/2026-02-16-ai-studio-03.png",[753,3565,3567],{"id":3566},"what-i-built-an-ai-content-studio","What I Built: An AI Content Studio",[758,3569,3570],{},"So I built one. I call it my content studio — a set of AI-powered skills that chain together to transform a rough idea into a fully published, bilingual blog post with illustrations, social media content, and even a video.",[758,3572,3573],{},"Here's what a single blog post looks like in my system:",[834,3575,3580],{"className":3576,"code":3578,"language":3579},[3577],"language-text","src/blogs/2026-02-14/\n├── plan.md                        ← my rough idea (the only thing I truly write)\n├── marriott-timeshare-las-vegas.md     ← English blog post\n├── marriott-timeshare-las-vegas-zh.md  ← Chinese translation\n├── youtube-script.md              ← narration script\n├── x-teaser.md                    ← X/Twitter thread\n├── linkedin-post.md               ← LinkedIn version\n├── video.mp4                      ← YouTube-ready video\n└── imgs/\n    ├── 00-cover.png               ← cover image\n    ├── 01-infographic-*.png       ← illustration\n    ├── 02-scene-*.png             ← illustration\n    └── ... (5 custom illustrations)\n","text",[841,3581,3578],{"__ignoreMap":839},[758,3583,3584,3585,3588],{},"One input. Eight outputs. And the input? It's a ",[841,3586,3587],{},"plan.md"," that looks like this — my actual plan for the Marriott timeshare post:",[3590,3591,3592],"blockquote",{},[758,3593,3594],{},"我最近有 travel 到 Las Vegas join Marriot 他们的 vacation club 在试图销售他们的 time sharing 系统... 根据以上的信息，是否可以帮我分析 Marriot 这个产品的商业模式，盈利模式以及销售中存在的问题。",[758,3596,3597],{},"That's it. A brain dump in mixed Chinese and English. No formatting, no structure, just the raw idea and what I want the post to explore. From there, the pipeline takes over.",[753,3599,3601],{"id":3600},"the-pipeline-how-it-actually-works","The Pipeline: How It Actually Works",[758,3603,3604],{},"Each step is handled by a specialized AI skill that knows exactly what to do:",[758,3606,3607,3610],{},[778,3608,3609],{},"Step 1: I write the plan."," This is the creative part — capturing what I want to say, the angle I want to take, the personal experience at the center. Sometimes it's a paragraph. Sometimes it's bullet points. Sometimes it's a voice note transcription. The format doesn't matter; the thinking does.",[758,3612,3613,3616],{},[778,3614,3615],{},"Step 2: AI drafts the blog post."," Using my plan as a guide, AI writes the full article. But here's the key — I review everything. I adjust the voice, add details only I would know, cut parts that feel generic. The draft gets me 80% there; I make it 100% mine.",[758,3618,3619,3622],{},[778,3620,3621],{},"Step 3: AI generates illustrations."," My article illustrator skill analyzes the post, identifies where images would add value, and generates them with a consistent visual style. Not stock photos — custom illustrations that match the content. Infographics for data, scenes for narrative moments, comparisons for analysis sections.",[758,3624,3625,3628],{},[778,3626,3627],{},"Step 4: AI translates to Chinese."," Not word-for-word translation — natural, fluent Simplified Chinese that reads like it was written by a native speaker. Technical terms stay in English where appropriate. The result is a genuine Chinese article, not a translated English one.",[758,3630,3631,3634],{},[778,3632,3633],{},"Step 5: AI creates distribution content."," An X teaser that hooks the Twitter audience. A LinkedIn post framed for professionals. A YouTube script structured for narration. Each one formatted for its platform, not just copy-pasted.",[758,3636,3637,3640],{},[778,3638,3639],{},"Step 6: AI generates a video."," The YouTube script becomes a narrated slideshow with transitions, background music, and proper pacing. Ready to upload.",[758,3642,3643,3646],{},[778,3644,3645],{},"Step 7: AI publishes."," Frontmatter generation, image path rewriting, sequential post numbering, deployment to my Nuxt 3 blog site. Both English and Chinese versions, linked as translations.",[758,3648,3649],{},[3036,3650],{"alt":3651,"src":3652},"The content pipeline — from rough plan to 8 published outputs","/blogs-img/2026-02-16-ai-studio-04.png",[753,3654,3656],{"id":3655},"why-this-isnt-ai-slop","Why This Isn't AI Slop",[758,3658,3659],{},"I know what you might be thinking. \"Isn't this just AI-generated content with extra steps?\"",[758,3661,3662],{},"No. And the distinction matters.",[758,3664,3665,3668],{},[778,3666,3667],{},"AI-generated content"," starts with a prompt like \"write me a blog post about timeshares.\" The AI invents the perspective, the examples, the opinions. It sounds like everyone and no one at the same time.",[758,3670,3671,3674],{},[778,3672,3673],{},"AI-produced content"," starts with my experience, my analysis, my perspective. The Marriott post exists because I actually sat through that pitch and was genuinely frustrated by the sales tactics. The AI Native Journey post exists because I actually changed how I work and wanted to document what I learned. No AI can invent those experiences.",[758,3676,3677],{},"AI handles the production — the writing mechanics, the formatting, the translation, the images. But every post starts from something real. That's why the output doesn't read like generic content. The ideas have edges. The examples are specific. The opinions are mine.",[758,3679,3680],{},"The proof is in the results: I'm publishing consistently for the first time in years. My Marriott post started as a frustrated Chinese brain dump and became a bilingual analysis of timeshare business models with custom illustrations. That kind of transformation — from raw experience to polished, multi-format content — is what the pipeline enables.",[753,3682,3684],{"id":3683},"what-this-means-for-you","What This Means for You",[758,3686,3687],{},"I'm not suggesting everyone needs to build exactly what I built. But I think there's a broader insight here that applies to any creator struggling with consistency.",[758,3689,3690,3693],{},[778,3691,3692],{},"The barrier to being a consistent creator in 2026 is not talent or ideas — it's execution overhead."," Every manual step between your idea and your audience is a friction point where you might quit. The more steps, the more likely you'll abandon the post halfway through.",[758,3695,3696],{},"AI tools have matured to the point where you can automate most of that overhead. You don't need a full content studio on day one. Start with the step that kills your momentum:",[804,3698,3699,3702,3705,3708],{},[775,3700,3701],{},"If translation is your bottleneck, automate that first",[775,3703,3704],{},"If creating images stops you from publishing, use AI illustration",[775,3706,3707],{},"If the publishing process is tedious, script it away",[775,3709,3710],{},"If repurposing for social media feels like a chore, let AI draft the variations",[758,3712,3713],{},"Each step you remove from your manual workload makes consistency that much more sustainable. And consistency compounds — more output means more audience, more feedback, more motivation to keep going.",[758,3715,3716],{},[3036,3717],{"alt":3718,"src":3719},"Start small — remove one friction point at a time","/blogs-img/2026-02-16-ai-studio-05.png",[753,3721,3723],{"id":3722},"the-honest-truth","The Honest Truth",[758,3725,3726],{},"I didn't become a better writer. My ideas aren't sharper than they were two years ago when I was failing to blog. What changed is that I stopped treating blogging as a writing problem and started treating it as a systems problem.",[758,3728,3729],{},"The blog post you're reading right now went through this exact pipeline. A rough plan, a draft, illustrations, translation, publishing — all handled by the AI content studio I built. My job was the thinking. The system handled the rest.",[758,3731,3732],{},"If you're someone who has great ideas but struggles to ship them consistently, I'd encourage you to look at your workflow the same way. Find the step that's killing your momentum. That's your starting point.",[758,3734,3735],{},"The tools exist. The question is whether you'll build the system.",[2163,3737],{},[758,3739,3740],{},[2156,3741,3742],{},"This is part of my AI Native Journey series, where I share how AI is changing the way I build, create, and work. If you're building your own AI-powered workflows, I'd love to hear what's working for you.",{"title":839,"searchDepth":708,"depth":708,"links":3744},[3745,3746,3747,3748,3749,3750,3751],{"id":3456,"depth":708,"text":3457},{"id":3520,"depth":708,"text":3521},{"id":3566,"depth":708,"text":3567},{"id":3600,"depth":708,"text":3601},{"id":3655,"depth":708,"text":3656},{"id":3683,"depth":708,"text":3684},{"id":3722,"depth":708,"text":3723},"I tried to be a consistent blogger for years and kept quitting. Then I built an AI content studio that turns a rough plan into a published bilingual blog post with illustrations, social teasers, and video — in minutes instead of hours.",{"date":3754,"image":3755,"alt":3756,"tags":3757,"category":3758,"topics":3759,"published":859},"16th Feb 2026","/blogs-img/2026-02-16-ai-studio-cover.png","AI content studio pipeline transforming one rough idea into multiple polished outputs",[2130,2462],"creation-media",[2134,3182],"/blogs/en/why-i-stopped-writing-and-built-ai-studio",{"title":3425,"description":3752},"blogs/en/15.why-i-stopped-writing-and-built-ai-studio","O0Ohoyf4dZyy2dU1rph5T3uNbgVVyuUX7TDN0FiXTbA",{"id":3765,"title":3766,"body":3767,"description":3964,"extension":2126,"meta":3965,"navigation":859,"ogImage":3967,"path":3973,"seo":3974,"stem":3975,"__hash__":3976},"en/blogs/en/16.mac-mini-ai-agent-content-team.md","I Had a Mac Mini Collecting Dust. It Now Runs My Entire Content Operation.",{"type":750,"value":3768,"toc":3958},[3769,3772,3776,3779,3782,3814,3817,3823,3827,3830,3835,3838,3841,3844,3850,3853,3856,3859,3863,3866,3869,3875,3881,3887,3893,3899,3902,3905,3908,3911,3914,3920,3924,3927,3930,3933,3936,3939,3942,3945,3951],[758,3770,3771],{},"I had a Mac mini sitting on a shelf doing nothing. Last Saturday, I installed OpenClaw on it, connected it to WhatsApp, and gave the agent a name — GG. Within 48 hours, GG had done more for my content business than I'd accomplished in the previous month. This is what an AI-native workflow actually looks like — no hype, just a dusty machine and a builder willing to try something different.",[753,3773,3775],{"id":3774},"the-setup-saturday-morning","The Setup (Saturday Morning)",[758,3777,3778],{},"The Mac mini had been sitting there for months. I'd bought it for a project that never panned out. When OpenClaw started going viral in January — literally causing a Mac mini shortage — I realized I already had the hardware everyone was scrambling to buy.",[758,3780,3781],{},"The setup took about two hours. Here's what I actually did:",[772,3783,3784,3790,3796,3802,3808],{},[775,3785,3786,3789],{},[778,3787,3788],{},"Configured the Mac mini to never sleep."," System Settings > Energy > Prevent automatic sleeping. Sounds obvious, but if you skip this, your agent goes dark at 2am.",[775,3791,3792,3795],{},[778,3793,3794],{},"Installed the OpenClaw gateway as a service."," Ran openclaw onboard in the terminal. The wizard walks you through everything — workspace, channels, skills. It's surprisingly smooth for an open-source tool.",[775,3797,3798,3801],{},[778,3799,3800],{},"Connected WhatsApp as the primary channel."," This was the key decision. I wanted to talk to my agent the same way I talk to people — through the messaging app I already use every day. No special interface, no browser tab to keep open.",[775,3803,3804,3807],{},[778,3805,3806],{},"Set up Tailscale for remote access."," This creates a private network between my MacBook Pro (my daily driver) and the Mac mini. Now I can SSH in or screen share from anywhere — my couch, a coffee shop, wherever.",[775,3809,3810,3813],{},[778,3811,3812],{},"Named the agent GG."," This might sound silly, but it matters. When you give the agent an identity, you start treating it like a collaborator instead of a search bar. The dynamic shifts.",[758,3815,3816],{},"Within two hours, I had a 24/7 AI agent living on my network, accessible through WhatsApp, with the ability to browse the web, read and write files, run shell commands, and remember everything across sessions.",[758,3818,3819],{},[3036,3820],{"alt":3821,"src":3822},"The always-on AI hub: Mac mini at the center of a connected ecosystem","/blogs-img/2026-02-18-mac-mini-01.png",[753,3824,3826],{"id":3825},"day-1-first-boot-to-first-value","Day 1: First Boot to First Value",[758,3828,3829],{},"The first thing I did was set up scheduled jobs — a morning briefing, daily content research, and a nightly report. I wanted GG to proactively work for me, not just respond when I asked.",[758,3831,3832],{},[778,3833,3834],{},"The morning briefing failed.",[758,3836,3837],{},"The gog CLI (the tool GG uses to check my email and calendar) wasn't authenticated yet. So the scheduled job ran, hit an auth error, and produced nothing. I woke up to an empty briefing.",[758,3839,3840],{},"I'm sharing this because it's important. AI workflows break. They break just like any other software. The morning briefing failed because I hadn't completed a setup step. The fix took five minutes — authenticate the CLI, test it manually, confirm it works, re-schedule the job. Done.",[758,3842,3843],{},"The difference between \"AI as hype\" and \"AI as workflow\" is right here: you don't panic when it breaks. You fix it and keep going.",[758,3845,3846],{},[3036,3847],{"alt":3848,"src":3849},"The builder's iteration loop: break, fix, ship, repeat","/blogs-img/2026-02-18-mac-mini-02.png",[758,3851,3852],{},"After fixing the briefing, I gave GG its first real task: explore my content studio workflow. I have a system of chained AI skills — brainstorm, illustrate, generate video, publish — built with tools that talk to each other. I wanted GG to understand the whole pipeline.",[758,3854,3855],{},"That's when the first \"aha moment\" hit. GG didn't just read the files. It browsed my blog, analyzed the skill definitions, connected the dots between tools, and came back with a coherent summary of how my pipeline works — including suggestions for improvements I hadn't considered.",[758,3857,3858],{},"This wasn't a chatbot spitting back my own words. GG had context. It could see my files, read my code, and understand the relationships between systems. And it remembered all of it for later.",[753,3860,3862],{"id":3861},"day-2-the-compounding-effect","Day 2: The Compounding Effect",[758,3864,3865],{},"Day 2 is when everything changed.",[758,3867,3868],{},"I woke up to a working morning briefing (the fix from Day 1 held), and then GG and I went on a tear. Here's what we accomplished in a single day:",[758,3870,3871,3874],{},[778,3872,3873],{},"Deep blog analysis."," GG crawled aaronguo.com and came back with specific improvement recommendations — not generic \"add more keywords\" advice, but structural feedback about navigation, content gaps, and positioning clarity. It had read every post.",[758,3876,3877,3880],{},[778,3878,3879],{},"Learning from the best."," I asked GG to analyze five creators I admire: Justin Welsh, Pieter Levels, Lenny Rachitsky, Sahil Bloom, and Dickie Bush. GG didn't just summarize their bios. It analyzed their content strategies — posting frequency, topic patterns, monetization models, audience engagement tactics — and told me what I could learn from each. From Justin Welsh, it was the simplicity of single-idea threads. From Pieter Levels, radical transparency with revenue dashboards. From Sahil Bloom, the newsletter-first brand strategy.",[758,3882,3883,3886],{},[778,3884,3885],{},"Newsletter setup."," GG researched newsletter platforms, recommended Beehiiv, and walked me through setting up \"Ship with AI\" — the newsletter you might be reading right now. It then researched how to integrate Beehiiv with my Nuxt 3 blog so subscribers get the full experience.",[758,3888,3889,3892],{},[778,3890,3891],{},"X/Twitter strategy."," This was the big one. GG did deep research on the X algorithm — how replies are weighted 27x more than likes, why external links get penalized ~50%, which content formats get the most reach. Then it synthesized everything into a full content strategy: weekly posting cadence, content pillars, hook formulas, reply engagement tactics, and a publishing timeline for every blog post.",[758,3894,3895,3898],{},[778,3896,3897],{},"Homepage design review."," GG pulled up my site, analyzed the layout, and gave specific design recommendations with reasoning for each change.",[758,3900,3901],{},"All of this happened in one day.",[758,3903,3904],{},"But here's the thing that matters: none of these tasks existed in isolation. The blog analysis informed the content strategy. The creator research shaped the positioning. The newsletter setup was driven by the strategy. Each task built on the context from the previous one.",[758,3906,3907],{},"By the end of Day 2, GG knew my positioning (\"Ship with AI, not about AI\"), my content pillars (AI-Native Execution, Product Leadership, Building in Public), my target audience, my tone of voice, and the specific projects I'm working on. It wasn't starting from zero anymore.",[758,3909,3910],{},"What OpenClaw does exceptionally well is manage this context. GG maintains structured memory files — daily notes and long-term memory — and ties them to everything it does: scheduled jobs, file access, shell commands, web browsing. The context isn't just \"remembered\" — it's actively used across tasks. When GG studied top creators in the morning and then built my content strategy in the afternoon, it didn't need me to connect the dots. The memory system did that automatically.",[758,3912,3913],{},"That's the compounding effect. And it's the real reason this setup is powerful.",[758,3915,3916],{},[3036,3917],{"alt":3918,"src":3919},"The compounding effect: each task builds on the context of the previous one","/blogs-img/2026-02-18-mac-mini-03.png",[753,3921,3923],{"id":3922},"what-i-actually-learned","What I Actually Learned",[758,3925,3926],{},"The shift isn't about OpenClaw specifically, or the Mac mini, or any single tool. It's about moving from \"AI as search engine\" to \"AI as persistent partner.\"",[758,3928,3929],{},"When you use AI in a browser tab, you're doing one-off queries. You get a response, maybe it's useful, and you move on. There's no memory, no context, no compounding.",[758,3931,3932],{},"When you set up a persistent agent — one that lives on your network, has access to your files, remembers your conversations, runs scheduled jobs while you sleep, and builds context over time — something fundamentally different happens. The agent stops being a tool and starts being a teammate.",[758,3934,3935],{},"After 48 hours, GG had enough context about my work that it stopped asking obvious questions. It knew my goals without being reminded. It referenced decisions from earlier conversations. It started connecting dots I hadn't connected myself.",[758,3937,3938],{},"And right now, as I write this blog post? GG helped plan it. The content plan, the distribution strategy, the X thread outline — all of it was a collaboration. This isn't \"AI-generated content.\" It's AI-augmented building. I make the decisions. GG does the heavy lifting.",[758,3940,3941],{},"If you have a Mac mini — or any always-on machine — the barrier to trying this is one afternoon. Two hours of setup, and you have a partner that never sleeps, never loses context, and gets better every day.",[758,3943,3944],{},"The hardest part isn't the setup. It's changing your mental model from \"AI is a tool I use\" to \"AI is a partner I work with.\"",[758,3946,3947],{},[3036,3948],{"alt":3949,"src":3950},"The shift: from AI as a tool you query to AI as a partner you work with","/blogs-img/2026-02-18-mac-mini-04.png",[758,3952,3953,3954,3957],{},"I'm documenting this entire journey — the wins, the failures, the real numbers — in my newsletter, ",[778,3955,3956],{},"Ship with AI",". If you want to follow along as I build in public with an AI partner, subscribe below.",{"title":839,"searchDepth":708,"depth":708,"links":3959},[3960,3961,3962,3963],{"id":3774,"depth":708,"text":3775},{"id":3825,"depth":708,"text":3826},{"id":3861,"depth":708,"text":3862},{"id":3922,"depth":708,"text":3923},"I set up OpenClaw on a Mac mini, named the agent GG, and within 48 hours it analyzed my blog, studied top creators, set up my newsletter, and built my entire X strategy. Here's what happened.",{"date":3966,"image":3967,"alt":3968,"tags":3969,"category":3758,"published":859},"18th Feb 2026","/blogs-img/2026-02-18-mac-mini-cover.png","Mac mini with OpenClaw lobster mascot as the always-on AI content hub",[30,3970,3971,3972,3182],"openclaw","mac-mini","productivity","/blogs/en/mac-mini-ai-agent-content-team",{"title":3766,"description":3964},"blogs/en/16.mac-mini-ai-agent-content-team","GaEqe2L3jy6usgaIYpfKBTj26TdhLiqsmvu8zkLK5X4",{"id":3978,"title":3979,"body":3980,"description":4273,"extension":2126,"meta":4274,"navigation":859,"ogImage":4276,"path":4283,"seo":4284,"stem":4285,"__hash__":4286},"en/blogs/en/17.mckinsey-wealth-management-2035-ai-insider-take.md","McKinsey Says 2035. I Say They're Off by 5 Years.",{"type":750,"value":3981,"toc":4265},[3982,3985,3988,3991,3995,3998,4036,4039,4048,4054,4058,4064,4070,4076,4079,4083,4086,4091,4094,4097,4100,4103,4108,4111,4118,4121,4126,4129,4132,4138,4142,4145,4150,4161,4164,4167,4170,4175,4178,4189,4192,4197,4200,4203,4206,4212,4214,4221,4227,4233,4239,4243,4246,4249,4252,4255,4258,4260],[758,3983,3984],{},"McKinsey just dropped a 10-page report predicting how AI, demographics, and trust will reshape US wealth management by 2035. It's a solid report — well-researched, carefully hedged, consultant-perfect. The data is real, the trends are undeniable, and the six themes they lay out are directionally correct.",[758,3986,3987],{},"But I read it differently than most people would, because I'm not a consultant. I'm a Head of Products — technology product, not financial products — at a financial firm who still codes every day, and I'm building AI-native systems on the side. The future they're describing? Parts of it are already here. The rest isn't 10 years away — it's 5.",[758,3989,3990],{},"The gap between consultant timelines and builder reality is the real story. And firms that plan for 2035 will be disrupted by firms building for 2030.",[753,3992,3994],{"id":3993},"the-report-in-60-seconds","The Report in 60 Seconds",[758,3996,3997],{},"McKinsey identifies six themes that will define US wealth management by 2035:",[772,3999,4000,4006,4012,4018,4024,4030],{},[775,4001,4002,4005],{},[778,4003,4004],{},"Demographic shift"," — a $22 trillion wealth transfer to Gen X and millennials, women controlling 40%+ of wealth, and a projected 100,000 advisor shortfall",[775,4007,4008,4011],{},[778,4009,4010],{},"Agentic AI"," — a shift from task automation to AI systems that reason, plan, and act autonomously",[775,4013,4014,4017],{},[778,4015,4016],{},"Asset class expansion"," — private markets, real assets, and digital/tokenized assets moving mainstream",[775,4019,4020,4023],{},[778,4021,4022],{},"Trust rebuilt through transparency"," — institutional reputation giving way to demonstrated performance and radical openness",[775,4025,4026,4029],{},[778,4027,4028],{},"Every client becomes a family office"," — holistic \"life management\" beyond investments into tax, estate, and lifestyle",[775,4031,4032,4035],{},[778,4033,4034],{},"Advisors evolving from planners to life coaches"," — the human role shifting to relationships and guidance, not information delivery",[758,4037,4038],{},"They also identify four competitive archetypes that will dominate: mega-platforms that win on scale, boutiques that win on relationships, independent platforms that win on flexibility, and AI-native firms that win on efficiency.",[758,4040,4041,4042,4047],{},"The full report is worth reading — ",[2177,4043,4046],{"href":4044,"rel":4045},"https://www.mckinsey.com/industries/financial-services/our-insights/us-wealth-management-in-2035-a-transformative-decade-begins",[2181],"link here",". But here's my take from someone inside the industry.",[758,4049,4050],{},[3036,4051],{"alt":4052,"src":4053},"McKinsey's six themes for wealth management — the map is right, but the timeline is off","/blogs-img/2026-02-17-mckinsey-01.jpg",[753,4055,4057],{"id":4056},"where-mckinsey-is-right","Where McKinsey Is Right",[758,4059,4060,4063],{},[778,4061,4062],{},"The advisor shortage is real and urgent."," I don't need McKinsey data to confirm this — I see it in my own firm every day. Experienced advisors are retiring, junior hires take years to ramp, and the pipeline is not keeping pace with client demand. McKinsey's 100,000 advisor shortfall might actually be conservative. The math doesn't work without AI augmentation, and most firms know it.",[758,4065,4066,4069],{},[778,4067,4068],{},"Trust is already shifting with younger clients."," McKinsey cites 76% of Gen Z seeking financial advice online, versus only 14% who would turn to a financial professional first (compared to 39% of boomers). This tracks exactly with what I see in our product data. Younger clients don't want to be told what to do by someone with credentials — they want to be shown, in real time, with transparency. The \"show me, don't tell me\" dynamic shows up in every client interaction now.",[758,4071,4072,4075],{},[778,4073,4074],{},"The \"family office for everyone\" trend is happening."," McKinsey cites 52% of investors seeking holistic advice in 2023, up from 29% in 2018. Clients are already demanding that wealth managers help with more than just portfolios — they want tax strategy, estate planning, insurance, and even lifestyle decisions bundled together. The tech to deliver this at lower wealth tiers is arriving. We're actively building toward this in our product roadmap right now.",[758,4077,4078],{},"On these three points, McKinsey is not just right — they're understating the urgency.",[753,4080,4082],{"id":4081},"where-mckinsey-is-too-conservative","Where McKinsey Is Too Conservative",[758,4084,4085],{},"Here's where I push back.",[758,4087,4088],{},[778,4089,4090],{},"Agentic AI is not a 2035 story — it's 2026.",[758,4092,4093],{},"McKinsey frames agentic AI as an emerging future capability. Meanwhile, I set up an AI agent on a Mac mini last week. It runs 24/7, does deep research, manages my content pipeline, sends me briefings every morning, and runs scheduled tasks autonomously while I sleep. I talked to it through WhatsApp yesterday about a blog post I was planning.",[758,4095,4096],{},"This isn't a research lab prototype. It's a dusty Mac mini with $15/month in API costs running production workflows for my content business.",[758,4098,4099],{},"In my day job at the financial firm, we're already using AI to augment product workflows, research, and reporting. The 62% of advisors who \"intend to use AI\" in McKinsey's 2024 data — the movers in that group have already moved. They're not intending anymore. They're shipping.",[758,4101,4102],{},"What used to take a research team multiple days now takes my AI agent a few hours. That's not 2035. That's today.",[758,4104,4105],{},[778,4106,4107],{},"The four archetypes underestimate AI-native disruption.",[758,4109,4110],{},"McKinsey's four competitive categories (mega-platforms, boutiques, independent platforms, AI-native firms) treat \"AI-native\" as a separate niche. I think that's the wrong frame.",[758,4112,4113,4114,4117],{},"McKinsey actually hints at the right answer with the concept of the \"firm of one\" — a single advisor supported by hundreds of AI agents, serving a client load that would have required a team of ten. But they don't follow this thread to its conclusion: when one person can do what ten did, the economics of ",[2156,4115,4116],{},"every"," model break. Not just the AI-native category.",[758,4119,4120],{},"AI-native isn't a niche. It's a capability that will eat into every category from the inside out. The boutique that onboards AI-native tools becomes a different kind of boutique. The mega-platform that figures out agentic client service at scale looks nothing like its current form. Every archetype gets disrupted, not just the ones in the bottom-right quadrant.",[758,4122,4123],{},[778,4124,4125],{},"Portfolio transformation is already shipping.",[758,4127,4128],{},"Tokenization of real assets? Direct indexing? Digital assets going institutional? McKinsey frames these as 2030-2035 concepts. But spot crypto ETFs and ETPs already exceeded $150 billion in AUM. Direct indexing platforms exist and are competing for clients today. \"Unified managed household\" systems are in early production at forward-leaning firms.",[758,4130,4131],{},"These aren't pilots. They're products. The question isn't when they'll arrive — it's when the laggards will catch up.",[758,4133,4134],{},[3036,4135],{"alt":4136,"src":4137},"The timeline gap: McKinsey's 2035 map vs. builder reality in 2026","/blogs-img/2026-02-17-mckinsey-02.jpg",[753,4139,4141],{"id":4140},"what-mckinsey-is-missing","What McKinsey Is Missing",[758,4143,4144],{},"Beyond the timeline, there are a few things the report doesn't adequately address.",[758,4146,4147],{},[778,4148,4149],{},"The builder vs. buyer divide.",[758,4151,4152,4153,4156,4157,4160],{},"The biggest competitive gap in wealth management over the next five years won't be between firm types. It'll be between firms that ",[2156,4154,4155],{},"build"," AI and firms that ",[2156,4158,4159],{},"buy"," it.",[758,4162,4163],{},"We've seen this pattern before. During the digital transformation wave, firms that outsourced their tech strategy to vendors consistently lost to firms that built internal capability. The vendor solutions were always one generation behind. The firms that owned their tech stacks could iterate in weeks; the vendor-dependent firms waited for release cycles.",[758,4165,4166],{},"AI is the same pattern, amplified. Firms that build AI systems internalize the feedback loops — they learn what clients actually respond to, they tune the models for their specific use cases, they own the data advantage. Firms that buy AI solutions get a generalized product built for the average case, not their clients.",[758,4168,4169],{},"This distinction doesn't appear in McKinsey's four archetypes. But it's the most important differentiator over the next five years.",[758,4171,4172],{},[778,4173,4174],{},"The real talent crisis isn't just about advisors.",[758,4176,4177],{},"McKinsey identifies the advisor shortage as a core challenge. Correct. But there's a second talent crisis hiding underneath it: who builds the AI systems?",[758,4179,4180,4181,4184,4185,4188],{},"You need product leaders and engineers who understand ",[2156,4182,4183],{},"both"," financial services ",[2156,4186,4187],{},"and"," AI architecture. Not advisors learning to prompt ChatGPT. Not AI engineers who have never touched a financial product. The hybrid — someone who can design an agentic workflow for client onboarding and also understands fiduciary duty, compliance constraints, and the emotional dynamics of a client who just lost 20% in a market drawdown — that person is genuinely rare.",[758,4190,4191],{},"I'm in this intersection. I lead technology product and engineering teams at a financial firm by day — building the systems that run the business, not managing financial products. I build AI-native systems on the side. I can tell you from direct experience: this skillset is the actual bottleneck. Not advisors. Not capital. Not regulation. Talent that bridges finance and AI is the constraint.",[758,4193,4194],{},[778,4195,4196],{},"Client experience will leapfrog, not evolve.",[758,4198,4199],{},"McKinsey describes a gradual evolution from \"clicks to conversations\" — a smooth improvement in client-facing interfaces as AI gets better. I think this mental model misses how phase changes actually work.",[758,4201,4202],{},"When a client can have a real conversation with an AI that knows their complete financial picture — every account, every goal, every life event on record — and get instant, personalized, data-grounded answers at 11pm on a Sunday, the bar for the human advisor conversation changes overnight. Not gradually. Overnight.",[758,4204,4205],{},"The advisors who thrive won't be the ones who incrementally improve their service model. They'll be the ones who completely redefine what the human value-add is when AI handles everything it can. That's not an evolution. That's a phase change.",[758,4207,4208],{},[3036,4209],{"alt":4210,"src":4211},"The \"firm of one\" — a single advisor supported by hundreds of AI agents","/blogs-img/2026-02-17-mckinsey-03.jpg",[753,4213,3684],{"id":3683},[758,4215,4216,4217,4220],{},"If you're a ",[778,4218,4219],{},"financial advisor",": The 62% of advisors \"intending\" to use AI will split into two groups. The ones who started in 2024 or 2025 will be building a capability moat that compounds over time. The ones who start in 2027 will be playing catch-up against people who already have 2-3 years of AI-augmented practice. Start now. Don't wait for a firm-wide rollout.",[758,4222,4216,4223,4226],{},[778,4224,4225],{},"product leader at a financial firm",": Push for building, not buying. Every year you spend waiting for a vendor to deliver a solution is a year your competitor's internal team is learning faster than your vendor's product team. The AI wave moves at software speed, not vendor contract speed.",[758,4228,4216,4229,4232],{},[778,4230,4231],{},"builder or engineer",": Financial services AI is massively under-served. Most fintech focuses on consumer payments or lending. Wealth management — the actual allocation and management of trillions in assets — is still being run with advisor workflows that haven't fundamentally changed in 20 years. The opportunity is enormous, and the intersection of finance + AI + product is where the biggest openings are.",[758,4234,4216,4235,4238],{},[778,4236,4237],{},"client",": Ask your wealth manager how they're using AI. Not \"are you exploring it\" — that's a yes/no anyone can answer. Ask what they built in the last 12 months. Ask what their junior advisors can do that they couldn't do two years ago. \"We're exploring it\" is a red flag in 2026.",[753,4240,4242],{"id":4241},"the-map-is-right-the-timeline-isnt","The Map Is Right. The Timeline Isn't.",[758,4244,4245],{},"McKinsey's report is a solid map. The six themes are real. The demographic pressures are real. The technology trajectory is real. As a research artifact, it's excellent.",[758,4247,4248],{},"But maps drawn by consultants tend to underestimate builder speed. The people inside firms who are actually shipping AI systems aren't waiting for 2035. They're not waiting for 2030. They're iterating right now, this week, and the pace is accelerating.",[758,4250,4251],{},"I'm in both worlds simultaneously — Head of Technology Products at a financial firm by day, AI-native builder by night. The future McKinsey describes isn't 10 years out. For the firms moving fastest, it's already happening. For the firms still planning, it's about 5 years away — not 10.",[758,4253,4254],{},"The question isn't whether wealth management will be transformed by AI. That question is settled. The question is whether your firm will be a leader in that transformation or a case study in why the laggards lost.",[758,4256,4257],{},"Plan for 2035, get disrupted by 2030. Build for 2030, lead through 2035.",[2163,4259],{},[758,4261,4262],{},[2156,4263,4264],{},"This is part of my AI Native Journey series, where I share what AI transformation actually looks like from inside a financial firm. If you're building at the intersection of finance and AI, I'd love to connect.",{"title":839,"searchDepth":708,"depth":708,"links":4266},[4267,4268,4269,4270,4271,4272],{"id":3993,"depth":708,"text":3994},{"id":4056,"depth":708,"text":4057},{"id":4081,"depth":708,"text":4082},{"id":4140,"depth":708,"text":4141},{"id":3683,"depth":708,"text":3684},{"id":4241,"depth":708,"text":4242},"McKinsey's new wealth management report is directionally right but too conservative on timing. As a technology product leader inside a financial firm who also builds AI systems, I break down where they're right, where they're 5 years too late, and what they're missing entirely.",{"date":4275,"image":4276,"alt":4277,"tags":4278,"category":3417,"published":859},"17th Feb 2026","/blogs-img/2026-02-17-mckinsey-cover.jpg","McKinsey 2035 consultant report vs builder reality 2030 — split comparison illustration",[4279,4280,4281,2134,4282],"wealth-management","agentic-ai","financial-services","product-leadership","/blogs/en/mckinsey-wealth-management-2035-ai-insider-take",{"title":3979,"description":4273},"blogs/en/17.mckinsey-wealth-management-2035-ai-insider-take","bYvn6ObMXk2BeUu0Z3mdc--DeMoUPjvmQethTGX0U2U",{"id":4288,"title":4289,"body":4290,"description":4497,"extension":2126,"meta":4498,"navigation":859,"ogImage":4500,"path":4507,"seo":4508,"stem":4509,"__hash__":4510},"en/blogs/en/18.ai-wont-make-everyone-a-creator.md","The Camera Didn't Make Everyone a Spielberg. AI Won't Either.",{"type":750,"value":4291,"toc":4489},[4292,4295,4299,4302,4305,4312,4315,4318,4324,4328,4331,4334,4337,4344,4351,4354,4357,4363,4367,4370,4373,4376,4382,4385,4391,4395,4398,4401,4404,4411,4414,4417,4420,4426,4430,4436,4439,4442,4445,4448,4451,4457,4463,4467,4470,4473,4476,4479,4482,4484],[758,4293,4294],{},"Every few decades, a new technology promises to democratize creativity. The camera. The synthesizer. Photoshop. Auto-Tune. Each time, the same prediction: now everyone can be a creator. Each time, the same result: the tool got easier, but making something great didn't. AI is the latest chapter in this story — and I'm living it right now.",[753,4296,4298],{"id":4297},"the-pattern","The Pattern",[758,4300,4301],{},"Think about it. When the camera became affordable, the prediction was that professional photography would be over. Anyone could capture an image. And that was true — anyone could. But Ansel Adams wasn't Ansel Adams because he owned a camera. He was Ansel Adams because he had an eye.",[758,4303,4304],{},"When the synthesizer hit the market, the prediction was that anyone could make music. No orchestra needed. And that was true — the barrier to producing a track dropped to nearly zero. But having a synthesizer didn't give you taste. The charts didn't suddenly fill with masterpieces. They filled with noise, and the great musicians still rose to the top because they had something to say.",[758,4306,4307,4308,4311],{},"Photoshop. Desktop publishing. Auto-Tune. GarageBand. TikTok. The story repeats. The tool makes ",[2156,4309,4310],{},"production"," easy. But production was never the hard part. The hard part was always knowing what to produce, why it matters, and when it's actually good.",[758,4313,4314],{},"Now it's AI's turn. And the prediction is louder than ever: AI will let everyone create professional content, videos, music, art. Anyone can be a creator now.",[758,4316,4317],{},"I use AI every day. I love it. And I'm here to tell you — that prediction is telling half the story.",[758,4319,4320],{},[3036,4321],{"alt":4322,"src":4323},"The same pattern repeating: each new tool lowers the floor, but never raises the ceiling","/blogs-img/2026-02-20-ai-creator-01.png",[753,4325,4327],{"id":4326},"what-i-learned-creating-with-ai","What I Learned Creating with AI",[758,4329,4330],{},"I've been using AI to generate blog posts, videos, and images for weeks now. Not as an experiment — as my actual workflow. AI is how I produce content.",[758,4332,4333],{},"Here's what AI does brilliantly: first drafts. Variations. Visual generation. It takes ideas from my head and turns them into something tangible faster than I ever could alone. The speed is real. What used to take me a full day now takes a few hours. That part of the promise is true.",[758,4335,4336],{},"Here's what AI doesn't do: decide what's worth making.",[758,4338,4339,4340,4343],{},"It doesn't know when a generated image is ",[2156,4341,4342],{},"almost"," right versus actually right. It doesn't bring personal experience or conviction. It can't tell you whether your blog post has a point or is just words arranged in a plausible order. It doesn't have taste.",[758,4345,4346,4347,4350],{},"The effort didn't disappear when I started using AI. It shifted. I spend less time on execution — writing first drafts, generating visuals, formatting content. I spend ",[2156,4348,4349],{},"more"," time on curation, refinement, and creative direction. More time deciding what to keep and what to throw away. More time asking: does this actually say something?",[758,4352,4353],{},"That shift is the part nobody talks about. The people who show you \"I made this in 5 minutes with AI\" aren't showing you the 2 hours they spent figuring out what to make, or the 30 iterations they threw away, or the final pass where they rewrote the parts that sounded like a robot wrote them.",[758,4355,4356],{},"The gap between \"AI generated this\" and \"this is genuinely good\" is still filled by the human.",[758,4358,4359],{},[3036,4360],{"alt":4361,"src":4362},"The effort shifted: less execution, more judgment and curation","/blogs-img/2026-02-20-ai-creator-02.png",[753,4364,4366],{"id":4365},"the-ai-slop-problem","The AI Slop Problem",[758,4368,4369],{},"If AI made quality effortless, we'd be living in a creative golden age right now. Instead, we got \"AI slop.\"",[758,4371,4372],{},"That term — AI slop — was named Word of the Year 2025 by both Merriam-Webster and the Australian National Dictionary. It describes the flood of low-effort, AI-generated content that's drowning the internet. And the numbers are staggering: research by Kapwing estimates that 21 to 33 percent of YouTube's feed may now consist of AI-generated junk content, generating around $117 million in ad revenue annually.",[758,4374,4375],{},"In January 2026, Bandcamp banned AI-generated music entirely. A survey found that 52 percent of consumers reduce engagement when they suspect content was made by AI. Audiences have developed what one researcher called a \"sixth sense\" for lazy AI — they can spot the overly enthusiastic adjectives, the repetitive structures, the lack of specific references that come from someone who clicked \"generate\" and called it a day.",[758,4377,4378,4379,4381],{},"The problem isn't AI. The problem is people who think the tool ",[2156,4380,3394],{}," the product.",[758,4383,4384],{},"A camera doesn't make a photograph great. Photoshop doesn't make a design great. And AI doesn't make content great. The tool is just the tool. What makes it great is the person behind it — their judgment, their taste, their willingness to throw away the first ten outputs and keep pushing until it's right.",[758,4386,4387],{},[3036,4388],{"alt":4389,"src":4390},"AI slop: when the tool becomes the product, quality disappears","/blogs-img/2026-02-20-ai-creator-03.png",[753,4392,4394],{"id":4393},"the-hype-machine","The Hype Machine",[758,4396,4397],{},"So why does the narrative persist that AI can do everything?",[758,4399,4400],{},"Because it's a convenient story to tell. And many people repeat it — not because it matches their experience, but because it's the trend.",[758,4402,4403],{},"TechCrunch declared 2026 the year AI moves \"from hype to pragmatism.\" The market is sobering up. But the hype machine is still running, fueled by people selling AI tools, AI courses, and AI-powered dreams.",[758,4405,4406,4407,4410],{},"Here's a data point that tells the real story. Researchers tested AI against 100,000 humans on creativity. The headline: \"AI beats humans.\" The actual data? AI beat the ",[2156,4408,4409],{},"average"," human. But the top 10 percent of humans far outperformed every AI system tested. The best humans weren't close — they demolished the machines.",[758,4412,4413],{},"The headline says \"AI beats humans.\" The data says \"AI beats mediocre humans.\" That's a very different story.",[758,4415,4416],{},"There's a concept emerging called the \"Humanity Moat.\" In a world where everyone has access to the same AI tools, the competitive advantage isn't the tool — it's trust, authenticity, and taste. The things AI can't generate.",[758,4418,4419],{},"The people telling you AI makes everything effortless are usually selling something. The people actually building with AI every day know the truth: AI is incredible, and it still requires you to show up with something worth amplifying.",[758,4421,4422],{},[3036,4423],{"alt":4424,"src":4425},"The headline says \"AI beats humans\" — the data tells a different story","/blogs-img/2026-02-20-ai-creator-04.png",[753,4427,4429],{"id":4428},"ai-is-a-power-tool-not-a-magic-wand","AI Is a Power Tool, Not a Magic Wand",[758,4431,4432,4433],{},"Here's the mental model I've landed on after weeks of building with AI: ",[778,4434,4435],{},"AI is the best amplifier ever built.",[758,4437,4438],{},"If you have taste, vision, and something to say — AI makes you 10x faster. It handles the grunt work so you can focus on the creative decisions. It generates variations so you can choose the best one. It turns rough ideas into polished outputs in minutes instead of hours.",[758,4440,4441],{},"If you don't have taste, vision, or something to say — AI helps you produce generic content faster. That's it. Faster mediocrity.",[758,4443,4444],{},"A survey of 1,100 music producers found that what they want from AI is tools that \"save time without flattening creativity.\" They want amplification, not replacement. They want the instrument to be powerful while the musician stays in charge.",[758,4446,4447],{},"This is the right analogy. AI gives everyone a camera. It doesn't give everyone an eye. AI gives everyone a studio. It doesn't give everyone a song worth recording.",[758,4449,4450],{},"I want to be very clear: this is not an argument against AI. I use it every single day. It has genuinely transformed how I work. I produce more, I iterate faster, I explore ideas I wouldn't have attempted before.",[758,4452,4453,4454,4456],{},"But I'm not confused about what's happening. AI is doing the heavy lifting on ",[2156,4455,2135],{},". The taste, the judgment, the perspective — that's still me. And that's the part that makes the work actually good.",[758,4458,4459],{},[3036,4460],{"alt":4461,"src":4462},"AI is an amplifier: what you bring in determines what comes out","/blogs-img/2026-02-20-ai-creator-05.png",[753,4464,4466],{"id":4465},"the-constant","The Constant",[758,4468,4469],{},"Technology changes. The fundamental equation doesn't.",[758,4471,4472],{},"In the age of the printing press, the bottleneck wasn't printing — it was having something worth printing. In the age of YouTube, the bottleneck isn't uploading — it's having something worth watching. In the age of AI, the bottleneck isn't generating — it's having something worth generating.",[758,4474,4475],{},"The creators who will thrive in the AI era aren't the ones who lean on AI the hardest. They're the ones who know what they want to say — and use AI thoughtfully to say it better.",[758,4477,4478],{},"AI is the best creative tool of our generation. Use it. Embrace it. But don't mistake the tool for the talent.",[758,4480,4481],{},"Technology was never the bottleneck. Your taste, your judgment, your perspective — that was always the hard part. It still is.",[2163,4483],{},[758,4485,4486,4487,2480],{},"I write about building with AI every week — the real experience, not the hype. If you want honest takes on what actually works, subscribe to ",[778,4488,3956],{},{"title":839,"searchDepth":708,"depth":708,"links":4490},[4491,4492,4493,4494,4495,4496],{"id":4297,"depth":708,"text":4298},{"id":4326,"depth":708,"text":4327},{"id":4365,"depth":708,"text":4366},{"id":4393,"depth":708,"text":4394},{"id":4428,"depth":708,"text":4429},{"id":4465,"depth":708,"text":4466},"AI is the most powerful creative tool ever built, but technology was never the bottleneck for great work. Taste, vision, and judgment always were — and still are.",{"date":4499,"image":4500,"alt":4501,"tags":4502,"category":3758,"youtube":4506,"published":859},"20th Feb 2026","/blogs-img/2026-02-20-ai-creator-cover.png","The evolution of creative tools from camera to AI with the constant human figure",[30,4503,4504,4505],"creativity","content-creation","opinion","https://youtu.be/S6WF_0KDynM","/blogs/en/ai-wont-make-everyone-a-creator",{"title":4289,"description":4497},"blogs/en/18.ai-wont-make-everyone-a-creator","IxQIPxaT-Z6BHtPEnoA9egK0pm73XIi0NOaip4WYFm8",{"id":4512,"title":4513,"body":4514,"description":4785,"extension":2126,"meta":4786,"navigation":859,"ogImage":4788,"path":4797,"seo":4798,"stem":4799,"__hash__":4800},"en/blogs/en/19.ai-two-work-modes.md","AI Made Me 10x More Productive. Then I Almost Burned Out.",{"type":750,"value":4515,"toc":4777},[4516,4519,4522,4526,4529,4532,4535,4538,4544,4548,4551,4554,4557,4560,4563,4569,4573,4580,4583,4590,4593,4596,4599,4605,4609,4619,4629,4634,4651,4658,4663,4680,4683,4690,4696,4700,4703,4709,4715,4721,4727,4733,4737,4740,4751,4754,4757,4760,4762,4765,4771],[758,4517,4518],{},"I've been running 5 AI agents in parallel, shipping content, generating videos, and automating workflows. I was more productive than ever. Then I hit a wall — not because AI failed, but because I was using the wrong mode for the wrong task.",[758,4520,4521],{},"This is the part nobody warns you about.",[753,4523,4525],{"id":4524},"the-10x-trap","The 10x Trap",[758,4527,4528],{},"A few weeks ago, I built an AI content pipeline that goes from a rough idea to a published blog post, illustrations, a video, and distribution across X and newsletter — all AI-assisted. I was running multiple agents simultaneously: one generating images, another drafting copy, a third assembling video. It felt like a superpower.",[758,4530,4531],{},"And the numbers backed it up. I was producing more than ever. What used to take me a full week took a day and a half. I was a one-person content studio.",[758,4533,4534],{},"But something was off. I was shipping faster, yet I felt more tired — not less. My decision quality was dropping. I'd review my own output and catch mistakes I normally wouldn't make. I was switching between five parallel tasks and doing none of them with the attention they deserved.",[758,4536,4537],{},"More output. Less clarity. I'm not alone in this.",[758,4539,4540],{},[3036,4541],{"alt":4542,"src":4543},"The 10x trap: more output, less clarity, growing fatigue","/blogs-img/2026-02-22-ai-modes-01.png",[753,4545,4547],{"id":4546},"the-research-backs-it-up","The Research Backs It Up",[758,4549,4550],{},"In February 2026, a UC Berkeley research team published a study that went viral. The headline: AI doesn't reduce work — it intensifies it. Workers who adopted AI tools didn't work less. They took on a broader scope of tasks, worked at a faster pace, and extended work into more hours of the day. As one worker put it: \"You had thought that maybe you could work less. But you don't work less. You just work the same amount or even more.\"",[758,4552,4553],{},"Harvard Business Review identified three forms of AI work intensification: task expansion (you take on things you used to delegate), blurred boundaries (you prompt AI during lunch, commute, evenings), and increased multitasking (you manage parallel AI workflows on top of your regular work).",[758,4555,4556],{},"The most striking finding came from TechCrunch: the most enthusiastic AI adopters are burning out first. The people who embraced AI the hardest are the ones hitting the wall. Not the skeptics. The believers.",[758,4558,4559],{},"Meanwhile, thousands of CEOs surveyed in early 2026 admitted AI had no measurable impact on their companies' productivity. The Solow paradox — \"you can see the computer age everywhere except in the productivity statistics\" — is back, repackaged for AI.",[758,4561,4562],{},"The problem isn't the tool. The problem is that we don't have a framework for how to use it.",[758,4564,4565],{},[3036,4566],{"alt":4567,"src":4568},"Research data: AI intensifies work rather than reducing it","/blogs-img/2026-02-22-ai-modes-02.png",[753,4570,4572],{"id":4571},"kahneman-had-the-answer-15-years-ago","Kahneman Had the Answer 15 Years Ago",[758,4574,4575,4576,4579],{},"Daniel Kahneman's ",[2156,4577,4578],{},"Thinking, Fast and Slow"," introduced a framework that explains exactly what's happening. He described two systems of thinking: System 1 — fast, intuitive, automatic — and System 2 — slow, deliberate, effortful.",[758,4581,4582],{},"System 1 is how you recognize a face, react to a loud noise, or skim an email. It's quick and low-cost. System 2 is how you solve a complex math problem, make a difficult career decision, or debug a subtle software issue. It's slow and draining.",[758,4584,4585,4586,4589],{},"Here's the connection: AI is the greatest System 1 accelerator ever built. It makes fast work ",[2156,4587,4588],{},"faster",". Drafting, generating, researching, scaffolding, exploring — all the parallel, generative tasks that benefit from speed and breadth — AI supercharges them.",[758,4591,4592],{},"But System 2 work — deep reasoning, judgment, resolving contradictions, making calls where the answer isn't obvious — can't be parallelized. You can't run five deep thoughts simultaneously any more than you can have five serious conversations at the same time.",[758,4594,4595],{},"We've known this tension for decades. Some people thrive in divergent, exploratory thinking — brainstorming, ideation, switching between threads. Others do their best work in convergent, focused thinking — going deep on one problem until it breaks open. The left brain / right brain distinction is a myth, but the underlying intuition is real: different tasks require fundamentally different cognitive modes.",[758,4597,4598],{},"AI didn't create this tension. It amplified it. And most of us haven't adapted.",[758,4600,4601],{},[3036,4602],{"alt":4603,"src":4604},"Kahneman's System 1 and System 2, mapped to AI work modes","/blogs-img/2026-02-22-ai-modes-03.png",[753,4606,4608],{"id":4607},"two-modes-scatter-and-laser","Two Modes: Scatter and Laser",[758,4610,4611,4612,4615,4616,2480],{},"After hitting my own wall, I started paying attention to when AI made me better and when it made me worse. The pattern was clear. It wasn't about using AI more or less. It was about matching the right ",[2156,4613,4614],{},"mode"," to the right ",[2156,4617,4618],{},"task",[758,4620,4621,4622,4625,4626,2480],{},"I call them ",[778,4623,4624],{},"Scatter Mode"," and ",[778,4627,4628],{},"Laser Mode",[758,4630,4631,4633],{},[778,4632,4624],{}," is parallel, exploratory, and AI-amplified. It's System 1 on steroids. This is when you want breadth:",[804,4635,4636,4639,4642,4645,4648],{},[775,4637,4638],{},"Researching a topic across multiple sources simultaneously",[775,4640,4641],{},"Generating drafts, variations, and options",[775,4643,4644],{},"Running parallel AI agents on independent tasks",[775,4646,4647],{},"Exploring new ideas, prototyping, scaffolding code",[775,4649,4650],{},"Building a content pipeline: blog, video, images, distribution",[758,4652,4653,4654,4657],{},"Scatter Mode is where AI shines brightest. You can run five agents, explore ten directions, generate twenty variations — and converge later. The key word is ",[2156,4655,4656],{},"independent",". If the tasks don't depend on each other, scatter them.",[758,4659,4660,4662],{},[778,4661,4628],{}," is sequential, focused, and human-led. It's System 2, protected from distraction. This is when you need depth:",[804,4664,4665,4668,4671,4674,4677],{},[775,4666,4667],{},"Debugging a subtle, layered problem",[775,4669,4670],{},"Making a product architecture decision with competing tradeoffs",[775,4672,4673],{},"Polishing a piece of writing until it says exactly what you mean",[775,4675,4676],{},"Designing a complex workflow where every step depends on the one before it",[775,4678,4679],{},"Having a difficult conversation with a teammate",[758,4681,4682],{},"Laser Mode is where the human is irreplaceable. You can use AI here — a single, focused conversation — but you can't parallelize. The work requires holding the entire problem in your head, following chains of dependency, and exercising judgment that only comes from deep engagement.",[758,4684,4685,4686,4689],{},"The insight that changed my workflow: ",[778,4687,4688],{},"the mode matters more than the tool."," AI in Scatter Mode (five agents generating options) is a completely different thing from AI in Laser Mode (one focused dialogue solving a hard problem). Same tool. Different cognitive context. Wildly different results.",[758,4691,4692],{},[3036,4693],{"alt":4694,"src":4695},"Two Modes: Scatter (parallel, breadth, AI-amplified) vs Laser (sequential, depth, human-led)","/blogs-img/2026-02-22-ai-modes-04.png",[753,4697,4699],{"id":4698},"what-i-actually-do","What I Actually Do",[758,4701,4702],{},"Here's how this plays out in my daily work.",[758,4704,4705,4708],{},[778,4706,4707],{},"Mornings are for Scatter Mode."," I start the day with exploratory, generative work. Research across multiple topics. Kick off parallel AI tasks — image generation, draft writing, code scaffolding. Queue up things that can run in the background while I move between threads. This is when I feel most like a conductor — orchestrating multiple streams, not executing any one of them deeply.",[758,4710,4711,4714],{},[778,4712,4713],{},"Afternoons are for Laser Mode."," I close the parallel streams. One problem. One screen. Sometimes one AI conversation, but often just me and the code or the document. This is where I make the hard calls: product decisions, workflow design, debugging something that resists easy answers.",[758,4716,4717,4720],{},[778,4718,4719],{},"My rule of thumb:"," If the task has dependencies or requires judgment calls between competing tradeoffs — go Laser. If it's exploratory, generative, or the subtasks are independent — go Scatter.",[758,4722,4723,4726],{},[778,4724,4725],{},"The transition ritual matters."," Switching from Scatter to Laser isn't instant. I close tabs, stop background agents, and give myself a five-minute buffer to shift cognitive gears. Going from \"managing five parallel streams\" to \"solving one hard problem\" requires a mental reset. Trying to jump straight from Scatter to Laser is how mistakes happen.",[758,4728,4729],{},[3036,4730],{"alt":4731,"src":4732},"Daily practice: Scatter mornings, Laser afternoons, with a transition ritual","/blogs-img/2026-02-22-ai-modes-05.png",[753,4734,4736],{"id":4735},"the-constant-is-you","The Constant Is You",[758,4738,4739],{},"In my last post, I wrote that AI is the best amplifier ever built. I still believe that. But now I see a more nuanced picture.",[758,4741,4742,4743,4746,4747,4750],{},"In Scatter Mode, AI amplifies your ",[2156,4744,4745],{},"breadth"," — how many directions you can explore, how many options you can generate, how many tasks you can run in parallel. In Laser Mode, AI amplifies your ",[2156,4748,4749],{},"depth"," — the quality of your reasoning, the precision of your output, the speed at which you reach a well-considered decision.",[758,4752,4753],{},"But in both modes, the human chooses. You decide when to scatter and when to focus. You decide when breadth serves you and when it's just noise. You decide when to let five agents run and when to shut everything down and think.",[758,4755,4756],{},"The people who will thrive in the AI era aren't the fastest multitaskers or the deepest focusers. They're the ones who can switch between both — who know that the meta-skill isn't using AI, but knowing which mode to be in.",[758,4758,4759],{},"Kahneman was right: knowing when to think fast and when to think slow has always been the human edge. AI just raised the stakes.",[2163,4761],{},[758,4763,4764],{},"What's your experience? Do you find yourself stuck in one mode — always scattering or always lasering? I'm genuinely curious. Reply and tell me.",[758,4766,4767,4768,4770],{},"I write about building with AI every week — honest takes on what actually works. Subscribe to ",[778,4769,3956],{}," for the real story.",[758,4772,4773],{},[3036,4774],{"alt":4775,"src":4776},"The constant: you choose the mode","/blogs-img/2026-02-22-ai-modes-06.png",{"title":839,"searchDepth":708,"depth":708,"links":4778},[4779,4780,4781,4782,4783,4784],{"id":4524,"depth":708,"text":4525},{"id":4546,"depth":708,"text":4547},{"id":4571,"depth":708,"text":4572},{"id":4607,"depth":708,"text":4608},{"id":4698,"depth":708,"text":4699},{"id":4735,"depth":708,"text":4736},"AI doesn't just change what we can do — it changes how we should think about work itself. A framework for knowing when to scatter (parallel, exploratory, AI-amplified) and when to laser (deep, focused, human-led).",{"date":4787,"image":4788,"alt":4789,"tags":4790,"category":3016,"youtube":4796,"published":859},"22nd Feb 2026","/blogs-img/2026-02-22-ai-modes-cover.png","Two work modes — Scatter and Laser — and the human choosing between them",[4791,4792,4793,4794,4795],"ai-productivity","work-modes","deep-work","system-thinking","burnout","https://youtu.be/QeRfP_nFFZE","/blogs/en/ai-two-work-modes",{"title":4513,"description":4785},"blogs/en/19.ai-two-work-modes","Ak8bZaGfa53fEp4atGS6eVR1YPqlzV4RqGIVK4tnzTw",{"id":4802,"title":4803,"body":4804,"description":5037,"extension":2126,"meta":5038,"navigation":859,"ogImage":5040,"path":5049,"seo":5050,"stem":5051,"__hash__":5052},"en/blogs/en/20.chamath-process-over-goals.md","I Chased Goals for 15 Years. Chamath Showed Me What I Got Wrong.",{"type":750,"value":4805,"toc":5030},[4806,4809,4812,4816,4819,4825,4828,4831,4834,4837,4842,4849,4852,4858,4862,4865,4871,4877,4883,4889,4895,4905,4911,4915,4922,4925,4928,4934,4937,4940,4947,4950,4956,4960,4963,4966,4972,4978,4984,4990,4997,5000,5006,5010,5013,5016,5019,5021,5024],[758,4807,4808],{},"I've set goals my entire career. Become a Senior Manager by 30. Ship the product by Q3. Hit the revenue target by year-end. I checked every box. And then I watched Chamath Palihapitiya — the guy who built Facebook's growth engine and Social Capital — explain why goal-orientation is the wrong framework entirely.",[758,4810,4811],{},"His argument: the people who win aren't chasing endpoints. They're committed to a process of continuous learning that never stops. I've been testing this idea against my own career — and he's uncomfortably right.",[753,4813,4815],{"id":4814},"the-goal-trap","The Goal Trap",[758,4817,4818],{},"For fifteen years, I optimized my career around milestones. Senior Manager by 30. Partner by 35. Ship the product. Hit the number. Each goal was clear, measurable, and motivating — exactly what every management book tells you to set.",[758,4820,4821,4822],{},"And I hit them. Most of them, anyway. But here's what nobody prepares you for: the moment after you arrive. There's a brief rush — a week, maybe two — and then a quiet emptiness. And the only response your brain offers is: ",[2156,4823,4824],{},"now what?",[758,4826,4827],{},"So you set another goal. And another. Goal → achieve → emptiness → new goal → repeat.",[758,4829,4830],{},"Psychologist Tal Ben-Shahar calls this the \"arrival fallacy\" — the illusion that reaching a destination will bring lasting happiness. He discovered it through his own experience as an elite squash player: every victory brought a brief euphoria that faded almost immediately. The destination was never the point. But he didn't realize that until he'd spent years chasing it.",[758,4832,4833],{},"I didn't either.",[758,4835,4836],{},"Then I watched Chamath's video — \"30 Years of Business Advice\" — and something clicked.",[4838,4839],"youtube-embed",{"title":4840,"videoId":4841},"Chamath Palihapitiya — 30 Years of Business Advice in 39 Minutes","0-LAT4HjWPo",[758,4843,4844,4845,4848],{},"His first rule is \"Never Stop.\" But he's not talking about hustle culture. He's talking about committing to the ",[2156,4846,4847],{},"process"," of learning and risk-taking, not to endpoints. Warren Buffett at 95 isn't chasing a goal. He's running a process. Charlie Munger was the same way. The people who sustain excellence across decades aren't the ones with the best goals. They're the ones who never stopped learning.",[758,4850,4851],{},"Chamath compressed 30 years into 6 rules. When you look at them through the lens of \"process over goals,\" they stop being a list of tips and start functioning as a coherent operating system.",[758,4853,4854],{},[3036,4855],{"alt":4856,"src":4857},"The goal trap: achieve, feel empty, repeat","/blogs-img/2026-02-23-chamath-process-01.png",[753,4859,4861],{"id":4860},"the-6-rules-as-a-system","The 6 Rules as a System",[758,4863,4864],{},"Here's what most summaries of Chamath's video miss: these aren't 6 independent tips. They're interconnected principles that all protect the same thing — your ability to keep learning. Remove any one of them, and the system breaks.",[758,4866,4867,4870],{},[778,4868,4869],{},"Rule 1: Never Stop."," This is the foundation. Don't treat life as a series of objectives where you declare victory and stop. Commit to continuous learning and continuous risk-taking. Not \"work 80 hours a week\" — rather, never allow yourself to stop growing. The most stagnant years of my career were when I was \"executing against a plan\" instead of learning. I had a roadmap, a timeline, clear deliverables — and I was coasting intellectually. The most productive period? The last six months. No plan. Just learning AI by building with it every day.",[758,4872,4873,4876],{},[778,4874,4875],{},"Rule 2: No Debt."," Debt — financial and technical — forces you into goal-orientation. When you owe, you optimize for short-term payoff, not long-term learning. Debt constrains optionality. Chamath points out how social media's curated lifestyles push young people toward material consumption → debt → being trapped in jobs they hate just to service payments. I see the same pattern in product work. When your codebase is buried in technical debt, every sprint becomes about patching, not learning. Same principle at the personal level: financial freedom isn't about luxury. It's about preserving the freedom to learn.",[758,4878,4879,4882],{},[778,4880,4881],{},"Rule 3: Manage with Humility."," Be absolutely honest about where you are right now. Not where you wish you were. Not where your goals say you should be. Honest assessment of current state. The hardest part of my AI transition was admitting I was starting from zero. Not \"I have 15 years of experience, I'll figure it out fast.\" The honest assessment: I'm a beginner at AI-native building. I don't know what I don't know. That humility was uncomfortable — but it unlocked faster learning than any confidence could have.",[758,4884,4885,4888],{},[778,4886,4887],{},"Rule 4: Surround Yourself with Younger People."," Chamath says your cognition solidifies at some point. Young people are \"early warning systems for the future\" — they help you see your blind spots. This isn't mentorship, which is a goal-oriented frame (\"I'll teach them\"). It's staying connected to emerging patterns, which is a process. The junior developers on my team adopted AI tools faster than anyone. They didn't have 15 years of \"the way things are done\" to unlearn. They were my signal that the shift was real — and that I needed to move.",[758,4890,4891,4894],{},[778,4892,4893],{},"Rule 5: Preserve Optionality."," Pursue win-win in business, negotiation, and life. Don't lock into paths that close doors. This is fundamentally anti-goal: goals narrow your focus to one outcome, while optionality keeps it wide. I've been building side projects while keeping my day job. Not \"I'm quitting to go all-in\" — that's goal-oriented thinking. Instead: \"I'm building capability and keeping doors open.\" Process-oriented thinking protects relationships, preserves options, and reduces the self-destructive moves that come from being locked into a single path.",[758,4896,4897,4900,4901,4904],{},[778,4898,4899],{},"Rule 6: Status is Manufactured."," All external recognition — the lists, the clubs, the invitations — are hooks that give others judgment over you. Chasing status is the ultimate goal-orientation trap. It optimizes for ",[2156,4902,4903],{},"other people's"," scorecards, not your own learning. Chamath calls disengaging from the status game a superpower. I've experienced this firsthand: titles and promotions felt great for a week. The actual skill-building — learning to code AI agents, publishing daily, building a content studio from scratch — has compounded in ways no title ever did. Nobody gave me a certificate for it. And that's exactly the point.",[758,4906,4907],{},[3036,4908],{"alt":4909,"src":4910},"The 6 rules as an interconnected system, all protecting continuous learning","/blogs-img/2026-02-23-chamath-process-02.png",[753,4912,4914],{"id":4913},"why-this-matters-more-now-than-ever","Why This Matters More Now Than Ever",[758,4916,4917,4918,4921],{},"You could argue that \"process over goals\" has always been good advice. And you'd be right — Kahneman, Seligman, and a century of psychology research back it up. But there's a reason this framework is ",[2156,4919,4920],{},"especially"," critical right now.",[758,4923,4924],{},"AI is accelerating the rate at which the landscape shifts. When the ground moves every six months, goals based on today's assumptions become obsolete before you reach them. The person who set a goal to \"master prompt engineering\" in January 2025 watched AI agents make that skill dramatically less relevant by December 2025. The person who committed to \"learning how AI systems work\" adapted seamlessly — because their north star was the process, not the destination.",[758,4926,4927],{},"Process-oriented people adapt automatically. They don't need to tear up their plan and start over every time the technology shifts. They were never attached to a specific plan in the first place.",[758,4929,4930,4931,2480],{},"Chamath references a 1957 experiment by Curt Richter that makes this visceral. Richter placed wild rats in water jars to measure how long they could swim before drowning. Most gave up within 15 minutes. But when he rescued them briefly — pulled them out, let them rest, then put them back — the rats that had been saved once swam for up to 60 hours. Not 60 minutes. 60 ",[2156,4932,4933],{},"hours",[758,4935,4936],{},"The difference wasn't physical strength. It was belief. The rats that had been rescued believed the situation wasn't hopeless. They had evidence that persistence could lead to being saved.",[758,4938,4939],{},"For builders navigating AI's uncertainty, the equivalent isn't believing a specific outcome will happen. It's believing in your own process — that if you keep learning, keep building, keep adapting, you'll find your way. That belief is what sustains you when everything else is shifting.",[758,4941,4942,4943,4946],{},"Chamath's practical advice for young people reinforces this: \"Go to the main stage.\" If you're in tech, go to Silicon Valley. If you're in finance, go to New York. This sounds like a goal — move to a specific city — but it's actually process-thinking. You're optimizing for ",[2156,4944,4945],{},"learning environment and opportunity density",", not for a specific reward.",[758,4948,4949],{},"His most controversial point: \"Work-life balance is a false issue.\" Before you dismiss this as billionaire hustle culture, listen to the reframe. When you find flow in a process you love, the balance question dissolves. Not because you work 24/7, but because the work doesn't feel like a sacrifice. I experience this with AI building. I work on it at 11pm sometimes — not because I have to, but because I'm in flow. That's not imbalance. That's alignment.",[758,4951,4952],{},[3036,4953],{"alt":4954,"src":4955},"Why process-orientation matters in the AI era: the landscape shifts too fast for fixed goals","/blogs-img/2026-02-23-chamath-process-03.png",[753,4957,4959],{"id":4958},"the-anti-goal-operating-system","The Anti-Goal Operating System",[758,4961,4962],{},"After absorbing Chamath's framework, I looked at how I'd been operating during my AI transition — the period where I've been most productive and most fulfilled — and realized I'd already been following a version of this system. I just hadn't articulated it.",[758,4964,4965],{},"Here's what it looks like in practice:",[758,4967,4968,4971],{},[778,4969,4970],{},"No 5-year plan."," Instead: daily learning rituals. Build something with AI every day. Some days it's a new workflow. Some days it's a blog post. Some days it's a failed experiment that teaches me something I couldn't have learned any other way.",[758,4973,4974,4977],{},[778,4975,4976],{},"No revenue targets for side projects."," Instead: ship and iterate. Let the market teach me. The feedback from publishing and building is more valuable than any forecast I could write in a spreadsheet.",[758,4979,4980,4983],{},[778,4981,4982],{},"No follower goals for the newsletter."," Instead: publish consistently, and make each piece better than the last. The growth has been a byproduct of the process, not the target.",[758,4985,4986,4989],{},[778,4987,4988],{},"No career destination."," Instead: accumulate capabilities that compound. Every skill I build feeds into the next one. Writing improves my thinking. Building AI workflows teaches me what's possible. Publishing teaches me what resonates.",[758,4991,4992,4993,4996],{},"The paradox is real: process-orientation often achieves \"goals\" faster than goal-orientation does. When you stop wasting energy on the anxiety gap — the distance between where you are and where your goal says you ",[2156,4994,4995],{},"should"," be — you redirect that energy toward actual work. James Clear articulated this perfectly: \"You do not rise to the level of your goals. You fall to the level of your systems.\" Chamath is saying the same thing from a different angle.",[758,4998,4999],{},"The compounding effect is the key. When your north star is \"learn continuously,\" every day's output — even the failures — feeds into tomorrow's capability. When your north star is \"hit X by date Y,\" the days that don't advance the metric feel wasted. One framework compounds. The other creates anxiety.",[758,5001,5002],{},[3036,5003],{"alt":5004,"src":5005},"The anti-goal operating system: daily learning, no targets, compound capability","/blogs-img/2026-02-23-chamath-process-04.png",[753,5007,5009],{"id":5008},"the-process-is-the-point","The Process Is the Point",[758,5011,5012],{},"Chamath's 6 rules aren't 6 separate lessons. They're one lesson with six facets: commit to the process of learning, remove the things that stop you — debt, status-seeking, fixed goals, ego — and stay relentlessly honest about where you actually are.",[758,5014,5015],{},"The best career advice I've received in recent memory isn't \"set bigger goals\" or \"hustle harder.\" It's this: the people who win are the ones who never stop learning. Not because they have a target. Because the process itself is the point.",[758,5017,5018],{},"My AI transition is living proof. No business plan. No timeline. No revenue target. Just a commitment to building with AI every single day. And somehow, it's producing more results than any goal-oriented plan I've ever written.",[2163,5020],{},[758,5022,5023],{},"I'm documenting this process — building with AI, learning in public, shipping weekly. Follow along if that resonates.",[758,5025,5026],{},[3036,5027],{"alt":5028,"src":5029},"The process is the point: commit to learning, let results compound","/blogs-img/2026-02-23-chamath-process-05.png",{"title":839,"searchDepth":708,"depth":708,"links":5031},[5032,5033,5034,5035,5036],{"id":4814,"depth":708,"text":4815},{"id":4860,"depth":708,"text":4861},{"id":4913,"depth":708,"text":4914},{"id":4958,"depth":708,"text":4959},{"id":5008,"depth":708,"text":5009},"Chamath Palihapitiya's 6 rules aren't independent tips — they form one operating system that protects continuous learning. I map each rule against my 15-year career and current AI transition to show why process-orientation matters more than ever.",{"date":5039,"image":5040,"alt":5041,"tags":5042,"category":3016,"youtube":5048,"published":859},"23rd Feb 2026","/blogs-img/2026-02-23-chamath-process-cover.png","Process over Goals — the operating system behind Chamath's 30 years of business advice",[5043,5044,5045,5046,4282,5047],"chamath","process-over-goals","career-strategy","continuous-learning","ai-career","https://youtu.be/tyK4-inXdSs","/blogs/en/chamath-process-over-goals",{"title":4803,"description":5037},"blogs/en/20.chamath-process-over-goals","9kYM4MWVZMsORLqR6sc74WPFMiC2UlIqnbJnkwle6m0",{"id":5054,"title":5055,"body":5056,"description":5523,"extension":2126,"meta":5524,"navigation":859,"ogImage":5526,"path":5533,"seo":5534,"stem":5535,"__hash__":5536},"en/blogs/en/21.anthropic-finance-plugins-insider-take.md","Anthropic Just Handed Finance Teams a Complete AI Stack. I Ran It on Real Work — Here's What I Found.",{"type":750,"value":5057,"toc":5512},[5058,5061,5064,5068,5071,5131,5148,5151,5154,5160,5164,5167,5170,5173,5179,5182,5186,5189,5192,5197,5200,5214,5220,5225,5228,5234,5243,5246,5250,5253,5256,5259,5265,5268,5274,5280,5286,5290,5293,5296,5299,5302,5305,5309,5312,5315,5318,5321,5324,5327,5333,5342,5345,5348,5354,5358,5361,5371,5377,5386,5392,5398,5402,5405,5408,5411,5414,5417,5423,5427,5430,5495,5497,5502],[758,5059,5060],{},"Anthropic released its financial-services-plugins this week — five Claude plugins covering DCF models, LBO analysis, equity research reports, IC memos, and wealth management workflows, wired to 11 institutional data sources including FactSet, S&P Global, Morningstar, and Moody's.",[758,5062,5063],{},"I build software at a financial firm. I spent time this week going through the repo — then ran it on real work. I produced a DCF model and a client report. Here's what I actually found.",[753,5065,5067],{"id":5066},"whats-in-the-box","What's in the Box",[758,5069,5070],{},"Five plugins. A core foundation layer that installs first — Financial Analysis — then four specialist modules you add based on what your team does:",[5072,5073,5074,5087],"table",{},[5075,5076,5077],"thead",{},[5078,5079,5080,5084],"tr",{},[5081,5082,5083],"th",{},"Plugin",[5081,5085,5086],{},"What it does",[5088,5089,5090,5101,5111,5121],"tbody",{},[5078,5091,5092,5098],{},[5093,5094,5095],"td",{},[778,5096,5097],{},"Investment Banking",[5093,5099,5100],{},"Drafts CIMs, builds buyer lists, runs merger models, tracks deals",[5078,5102,5103,5108],{},[5093,5104,5105],{},[778,5106,5107],{},"Equity Research",[5093,5109,5110],{},"Writes earnings reports, builds investment theses, runs sector coverage",[5078,5112,5113,5118],{},[5093,5114,5115],{},[778,5116,5117],{},"Private Equity",[5093,5119,5120],{},"Sources deals, runs due diligence, drafts IC memos, monitors portfolio KPIs",[5078,5122,5123,5128],{},[5093,5124,5125],{},[778,5126,5127],{},"Wealth Management",[5093,5129,5130],{},"Preps client meetings, rebalances portfolios, builds financial plans",[758,5132,5133,5134,5137,5138,5137,5141,5137,5144,5147],{},"Each plugin adds slash commands you run directly — ",[841,5135,5136],{},"/comps [company]",", ",[841,5139,5140],{},"/dcf [company]",[841,5142,5143],{},"/earnings [company] [quarter]",[841,5145,5146],{},"/ic-memo [project]",". Output lands in Excel with working formulas, PowerPoint with your firm's branded templates, or Word documents. No code. No setup beyond installing the plugin.",[758,5149,5150],{},"What most coverage glosses over: the skill files. These aren't generic AI prompts dressed in financial vocabulary. They encode the kind of reasoning a senior analyst actually applies to a comp table or an IC memo. Read through a few and you'll see what thoughtful design looks like.",[758,5152,5153],{},"Neither the commands nor the skill files are the most interesting part of this launch.",[758,5155,5156],{},[3036,5157],{"alt":5158,"src":5159},"Repo architecture: financial-analysis core plugin, four specialist add-ons, 11 MCP data connectors wired to institutional data sources","/blogs-img/2026-02-26-anthropic-finance-01.jpg",[753,5161,5163],{"id":5162},"the-data-incumbents-just-declared-a-standard","The Data Incumbents Just Declared a Standard",[758,5165,5166],{},"FactSet. S&P Global. Morningstar. Moody's. LSEG. PitchBook — eleven institutional data vendors, all building native MCP connectors in the same launch window. These aren't startups. They're the data backbone of institutional finance, with multi-year enterprise contracts embedded at every major firm.",[758,5168,5169],{},"When incumbents this entrenched all move in the same direction at once, that's a standard being declared. The historical parallel: when Bloomberg and FactSet opened REST APIs in the 2010s, a generation of fintech was built on top. MCP is the 2026 version of that moment.",[758,5171,5172],{},"Data vendors are the pipes. Claude is the plumber.",[758,5174,5175],{},[3036,5176],{"alt":5177,"src":5178},"MCP convergence: 11 institutional data vendors all connecting to MCP Protocol, feeding into Claude and its outputs","/blogs-img/2026-02-26-anthropic-finance-02.jpg",[758,5180,5181],{},"In my firm, the question isn't \"should we adopt AI?\" — that's settled. It's: what does it cost to turn on the MCP connector for the data vendor we already have under contract? That's a different, and much more tractable, conversation.",[753,5183,5185],{"id":5184},"i-actually-ran-it-dcf-and-client-report","I Actually Ran It — DCF and Client Report",[758,5187,5188],{},"Reading the repo is one thing. I wanted to know what it actually produces on real work.",[758,5190,5191],{},"I ran two tests: a DCF model and a client-facing report. The output files below are real — unmodified outputs from Claude, shared here exactly as generated so you can see what the plugin actually produces.",[758,5193,5194],{},[778,5195,5196],{},"The DCF:",[758,5198,5199],{},"The output came back as a proper Excel workbook — not a static table, not a screenshot, but a live model with formula references, linked cells, and a sensitivity table built in. The structure followed investment banking conventions: revenue build, margin assumptions, terminal value, WACC calculation, implied valuation range. It wasn't perfect — there were assumptions I'd adjust and a few places where I'd want to override the defaults — but the scaffolding was right, and editing it from that starting point is materially faster than building from scratch.",[758,5201,5202,5203,5209,5210,5213],{},"📎 ",[778,5204,5205],{},[2177,5206,5208],{"href":5207},"/blog-assets/AMZN_DCF_Model_2026-02-25.xlsx","Download: AMZN DCF Model (Excel)"," — actual output from ",[841,5211,5212],{},"/dcf AMZN",", unedited after generation",[758,5215,5216],{},[3036,5217],{"alt":5218,"src":5219},"AMZN DCF model in Excel — scenario assumptions, revenue build, WACC, and sensitivity table as generated","/blogs-img/2026-02-26-anthropic-finance-03.png",[758,5221,5222],{},[778,5223,5224],{},"The Client Report:",[758,5226,5227],{},"The report output (generated using sample data, not real client information) was similarly structured — executive summary, key metrics, portfolio commentary, formatted for a client-facing context. The language was professional. The structure was standard. Again, not something I'd send as-is without review, but a solid first draft that cuts the \"blank page\" problem entirely.",[758,5229,5230],{},[3036,5231],{"alt":5232,"src":5233},"Client report output — formatted for client-facing use","/blogs-img/2026-02-26-anthropic-finance-04.png",[758,5235,5202,5236,5242],{},[778,5237,5238],{},[2177,5239,5241],{"href":5240},"/blog-assets/Sample_Customer_Performance_Report_FY2025.xlsx","Download: Client Performance Report (Excel)"," — actual output from the client report command, real test run",[758,5244,5245],{},"What I took away from both: the value is not in the final output being perfect. It's in the starting point being right. A blank Excel and a blank document are very different from a correctly-structured workbook and a properly-formatted draft. The former is a task. The latter is an editing job. That gap — from blank page to structured starting point — is where most of the time goes in practice.",[753,5247,5249],{"id":5248},"skills-are-becoming-the-product","Skills Are Becoming the Product",[758,5251,5252],{},"After using this, I came away thinking about something broader than just these specific plugins.",[758,5254,5255],{},"The architecture here — domain knowledge encoded in Markdown files, organized into skills and commands, wired to data sources via JSON — is a pattern that's going to generalize far beyond financial services. The software isn't the Python or the API calls. The software is the Markdown.",[758,5257,5258],{},"This is a meaningful shift in what building looks like. Skills-oriented project design — where the core IP of the system lives in structured, human-readable domain knowledge files rather than in code — is going to become more common. The codebase of the future, for a lot of knowledge-work automation, might primarily be a collection of well-structured Markdown files.",[758,5260,5261],{},[3036,5262],{"alt":5263,"src":5264},"Skills Oriented Programming — expertise encoded in structured files, not in code","/blogs-img/2026-02-26-anthropic-finance-08.jpg",[758,5266,5267],{},"A few things follow from this:",[758,5269,5270,5273],{},[778,5271,5272],{},"The skill files themselves are the moat."," Anyone can install this repo. The differentiation comes from the firm-specific skills layered on top — the proprietary deal frameworks, the house views encoded as constraints, the client communication style formalized into a skill file. That institutional knowledge, codified and version-controlled, is harder to replicate than any SaaS subscription.",[758,5275,5276,5279],{},[778,5277,5278],{},"There's a real consulting opportunity around skills configuration."," Writing and structuring skill files for specific industries and workflows is a craft. A firm that hires someone to configure these properly — translating their actual processes into well-structured domain knowledge files — gets something qualitatively different from a firm that installs the defaults and leaves them. As this pattern spreads, the people who know how to do that translation well will be in demand.",[758,5281,5282,5285],{},[778,5283,5284],{},"\"Just Markdown\" is the wrong way to read the skepticism."," One critique circulating this week: these skills are \"just prompts,\" and anyone could replicate them with a cheaper model. Technically true. Practically irrelevant. Firms don't want to write and maintain custom system prompts for 41 financial workflows. The files in this repo are pre-built, version-controlled, professionally designed domain knowledge. The fact that it's Markdown is a feature — any senior person who understands the workflow can read, verify, and extend it without touching a line of code.",[753,5287,5289],{"id":5288},"what-im-actually-considering-internally","What I'm Actually Considering Internally",[758,5291,5292],{},"Based on what I saw, I'm seriously thinking about rolling out Claude Code plus skills to both our sales team and our R&D team.",[758,5294,5295],{},"The skills format is what makes this realistic for non-technical teams. I don't need to build a custom tool, train people on a new interface, or maintain an integration. I write the skill files that encode our specific workflows, configure the data connectors we already have contracts for, and the team has an AI co-pilot that understands our domain from day one.",[758,5297,5298],{},"For sales: a skills set that knows our products, our typical client profiles, our common objections and responses, our proposal formats.",[758,5300,5301],{},"For R&D: a skills set that knows our research frameworks, our data sources, our documentation standards.",[758,5303,5304],{},"The institutional knowledge that currently lives in people's heads — and walks out the door when they leave — gets encoded and made queryable. That's a different kind of value from a generic AI assistant.",[753,5306,5308],{"id":5307},"a-real-example-from-agentic-workflow-to-skill-file","A Real Example: From Agentic Workflow to Skill File",[758,5310,5311],{},"This shift became concrete for me through a conversation I'd been having with our client-facing team a few weeks before this launch.",[758,5313,5314],{},"The problem we were trying to solve: automating personalized investment recommendations. A client comes in with a specific portfolio, specific goals, specific risk profile. Producing a thoughtful, firm-standard recommendation used to take significant manual work — pulling the data, reasoning through the allocation logic, drafting the document in our format.",[758,5316,5317],{},"The original plan was to build an agentic workflow from scratch. Design the data pipeline, write the orchestration logic, build the output template, wire it to our systems, test it, maintain it. Real engineering work — weeks of development, ongoing maintenance, a dedicated person to own it.",[758,5319,5320],{},"After reading through these plugins, I realized we could get most of that done with a skill file.",[758,5322,5323],{},"One skill file that encodes our investment recommendation logic — our allocation framework, our client communication standards, our required disclosures, our output format. One command that pulls the client's portfolio data and runs it through that reasoning. The output: a structured recommendation document, formatted to our standard, ready for advisor review.",[758,5325,5326],{},"I leveraged the wealth management skill to run it against a sample client profile — and the output speaks for itself.",[758,5328,5329],{},[3036,5330],{"alt":5331,"src":5332},"Output: personalized investment recommendation generated from the skill, formatted to firm standard","/blogs-img/2026-02-26-anthropic-finance-07.png",[758,5334,5202,5335,5341],{},[778,5336,5337],{},[2177,5338,5340],{"href":5339},"/blog-assets/Guo_Aaron_Investment_Proposal.md","Download: Investment Proposal (Markdown)"," — actual skill output for a sample client profile, unedited",[758,5343,5344],{},"The thing that surprised me wasn't the quality of the output — it was the speed of building it. What I estimated as a several-week engineering project became a few hours of writing a well-structured Markdown file. The skill file itself is readable by any senior person on the team. It can be reviewed, questioned, and refined without touching a line of code.",[758,5346,5347],{},"That gap — between \"we need to build an agentic system\" and \"we need to write a skill file\" — is the paradigm shift. And it's not just for investment recommendations. Any workflow that can be encoded as structured reasoning belongs in a skill file now, not in a custom-built agentic pipeline.",[758,5349,5350],{},[3036,5351],{"alt":5352,"src":5353},"Paradigm shift: building an agentic workflow (weeks of engineering) vs writing a skill file (hours)","/blogs-img/2026-02-26-anthropic-finance-05.jpg",[753,5355,5357],{"id":5356},"what-would-actually-get-adopted","What Would Actually Get Adopted",[758,5359,5360],{},"A few things from the technical and hands-on evaluation:",[758,5362,5363,5366,5367,5370],{},[778,5364,5365],{},"What moves fast:"," Deliverables in known formats where the data subscription is already in place. Earnings summaries, comp tables, first-draft research notes, client-facing reports. The incremental lift: install the plugin, configure ",[841,5368,5369],{},".mcp.json"," to point at the provider you already use, done. Months, not years.",[758,5372,5373,5376],{},[778,5374,5375],{},"What moves slower:"," Full models going directly into client-facing materials without review. Before an AI-generated DCF goes in a pitch deck, compliance and risk need to sign off on the model architecture and the review process. That's not a Claude problem — it's a controls problem that exists with Excel templates already. It just gets more formal when AI is in the loop.",[758,5378,5379,5382,5383,5385],{},[778,5380,5381],{},"The real blocker is data, not tooling."," If your firm doesn't have FactSet or Capital IQ under contract, you're not unlocking the full depth of these connectors. The plugins function without the MCP data feeds — but it's the difference between a Bloomberg terminal with a subscription and one without. The open-source nature changes the calculus for mid-market though: fork the repo, configure ",[841,5384,5369],{}," to point at the one provider you do have, add your firm's own models and terminology. What used to require a software project now requires someone who can edit Markdown.",[758,5387,5388,5391],{},[778,5389,5390],{},"Start the compliance conversation before you're ready to deploy."," The technical lift is lighter than most teams expect. The organizational lift — compliance sign-off, IT security review, data governance documentation — is where the timeline actually lives. If you wait until you're technically ready to start that conversation, you've already lost two or three months.",[758,5393,5394],{},[3036,5395],{"alt":5396,"src":5397},"What moves fast vs. what takes longer — adoption framework for mid-market financial firms","/blogs-img/2026-02-26-anthropic-finance-06.jpg",[753,5399,5401],{"id":5400},"the-blueprint","The Blueprint",[758,5403,5404],{},"Anthropic didn't just release plugins this week. They released a design specification for what the AI-native financial firm's tooling stack looks like.",[758,5406,5407],{},"The architecture: a core plugin with shared modeling infrastructure, specialist add-ons per function, MCP connectors for institutional data, workflow outputs directly to Excel and PowerPoint. Modular, data-connected, workflow-specific, customizable in Markdown.",[758,5409,5410],{},"The more I think about it, the more I think the skill-file pattern is the lasting contribution here — not any specific command or output. Domain knowledge encoded in structured, version-controlled, human-readable files that any sufficiently senior person in the relevant field can read and extend. That's a genuinely new software primitive.",[758,5412,5413],{},"The firms figuring out how to encode their institutional knowledge into skill files right now — their deal frameworks, their research methodologies, their client communication standards — are building something that compounds. Not just a productivity tool. An organizational knowledge system that gets more useful over time.",[758,5415,5416],{},"The firms that wait for a vendor to package this into a licensed product will pay three times as much for something already a generation behind. We've seen this pattern before in financial technology. It's playing out again.",[758,5418,5419],{},[3036,5420],{"alt":5421,"src":5422},"The Blueprint — AI-native financial firm architecture: institutional knowledge in skill files, core infrastructure with specialist modules, data connectors feeding from below, outputs to Excel/PowerPoint/Word","/blogs-img/2026-02-26-anthropic-finance-09.jpg",[753,5424,5426],{"id":5425},"downloads-real-outputs-from-this-post","Downloads — Real Outputs From This Post",[758,5428,5429],{},"All files below are unmodified outputs from Claude, generated during the tests described above. Shared as-is to show what the plugin actually produces.",[5072,5431,5432,5445],{},[5075,5433,5434],{},[5078,5435,5436,5439,5442],{},[5081,5437,5438],{},"File",[5081,5440,5441],{},"Format",[5081,5443,5444],{},"Section",[5088,5446,5447,5460,5471,5484],{},[5078,5448,5449,5454,5457],{},[5093,5450,5451],{},[2177,5452,5453],{"href":5207},"AMZN DCF Model",[5093,5455,5456],{},"Excel",[5093,5458,5459],{},"I Actually Ran It",[5078,5461,5462,5467,5469],{},[5093,5463,5464],{},[2177,5465,5466],{"href":5240},"Client Performance Report",[5093,5468,5456],{},[5093,5470,5459],{},[5078,5472,5473,5479,5481],{},[5093,5474,5475],{},[2177,5476,5478],{"href":5477},"/blog-assets/Guo_Aaron_Investment_Proposal.xlsx","Investment Proposal",[5093,5480,5456],{},[5093,5482,5483],{},"From Agentic Workflow to Skill File",[5078,5485,5486,5490,5493],{},[5093,5487,5488],{},[2177,5489,5478],{"href":5339},[5093,5491,5492],{},"Markdown",[5093,5494,5483],{},[2163,5496],{},[758,5498,5499],{},[2156,5500,5501],{},"I write about building AI-native at the intersection of financial services and tech product — what it actually looks like from inside a firm. Newsletter below:",[758,5503,5504],{},[2156,5505,5506,5507],{},"→ ",[2177,5508,5511],{"href":5509,"rel":5510},"https://shipwithai.beehiiv.com",[2181],"Ship with AI on Beehiiv",{"title":839,"searchDepth":708,"depth":708,"links":5513},[5514,5515,5516,5517,5518,5519,5520,5521,5522],{"id":5066,"depth":708,"text":5067},{"id":5162,"depth":708,"text":5163},{"id":5184,"depth":708,"text":5185},{"id":5248,"depth":708,"text":5249},{"id":5288,"depth":708,"text":5289},{"id":5307,"depth":708,"text":5308},{"id":5356,"depth":708,"text":5357},{"id":5400,"depth":708,"text":5401},{"id":5425,"depth":708,"text":5426},"Anthropic released financial-services-plugins this week — five Claude plugins wired to 11 institutional data sources. I'm a Head of Technology Products at a financial firm. I ran it on a real DCF and client report. Here's what I actually found.",{"date":5525,"image":5526,"alt":5527,"tags":5528,"category":3417,"published":859},"26th Feb 2026","/blogs-img/2026-02-26-anthropic-finance-cover.jpg","Blueprint architecture diagram of Anthropic's financial AI stack — Claude and Skills at top, four specialist modules in the middle, 11 institutional data connectors at the bottom",[30,5529,5530,5531,5532],"finance","anthropic","mcp","financial-analysis","/blogs/en/anthropic-finance-plugins-insider-take",{"title":5055,"description":5523},"blogs/en/21.anthropic-finance-plugins-insider-take","I8LrCy-qy0cKG5rNKr29Oyh9lSsHmOJ-oIGATWvOUFk",{"id":5538,"title":5539,"body":5540,"description":6858,"extension":2126,"meta":6859,"navigation":859,"ogImage":6861,"path":6869,"seo":6870,"stem":6871,"__hash__":6872},"en/blogs/en/22.chatgpt-explained-kitchen-metaphor.md","No Magic: How ChatGPT Actually Works, Explained in One Kitchen",{"type":750,"value":5541,"toc":6828},[5542,5553,5555,5558,5561,5564,5571,5578,5581,5587,5589,5592,5595,5601,5604,5615,5618,5621,5624,5626,5630,5633,5636,5643,5650,5656,5662,5669,5675,5682,5684,5688,5691,5697,5704,5707,5710,5716,5731,5737,5740,5747,5749,5753,5759,5762,5765,5771,5777,5780,5782,5786,5789,5792,5803,5814,5817,5820,5826,5833,5835,5839,5842,5845,5848,5854,5861,5863,5867,5870,5875,5878,5885,5887,5891,5894,5900,5906,5909,5915,5917,5921,5924,5935,5942,5945,5951,5957,5964,5966,5970,5973,5976,5982,5985,5989,5992,5999,6005,6009,6015,6029,6032,6036,6039,6119,6122,6128,6132,6135,6152,6155,6163,6170,6177,6184,6186,6190,6193,6213,6216,6219,6223,6226,6232,6238,6244,6248,6308,6314,6317,6323,6330,6332,6336,6339,6356,6359,6363,6368,6388,6394,6397,6401,6404,6407,6417,6424,6426,6430,6433,6520,6526,6529,6533,6536,6542,6548,6551,6557,6564,6566,6570,6573,6576,6579,6582,6589,6592,6595,6598,6600,6609,6611,6615],[758,5543,5544],{},[2156,5545,5546,5547,5552],{},"Inspired by Andrej Karpathy's ",[2177,5548,5551],{"href":5549,"rel":5550},"https://karpathy.github.io/2026/02/12/microgpt/",[2181],"MicroGPT"," — 200 lines of code that prove it's all just a kitchen.",[2163,5554],{},[758,5556,5557],{},"You walk into an omakase restaurant for the first time.",[758,5559,5560],{},"The host guides you to a counter seat. There's no menu. The chef nods and begins.",[758,5562,5563],{},"Dishes arrive one after another — sashimi first, then something warm, then something acidic, then rich, then light. Each course feels inevitable. By the time dessert arrives, you realize you never once thought \"this doesn't belong here.\" The whole meal felt like it was designed specifically for you, by someone who understood something you couldn't quite articulate.",[758,5565,5566,5567,5570],{},"Here's the question that lodged in my head after reading a recent post by Andrej Karpathy (OpenAI co-founder, former Tesla AI director): how does the chef ",[2156,5568,5569],{},"know","?",[758,5572,5573,5574,5577],{},"Not the senior chef with 20 years of experience. I mean — how does ",[2156,5575,5576],{},"any"," chef learn, from scratch, what order of courses will feel right to a stranger they've never met? No one handed them a rulebook. They didn't memorize every possible omakase sequence.",[758,5579,5580],{},"The answer is Ming.",[758,5582,5583],{},[3036,5584],{"alt":5585,"src":5586},"Ming stands at the omakase counter facing a wall of 4,192 judgment dials","/blogs-img/2026-03-01-chatgpt-kitchen-01.png",[2163,5588],{},[758,5590,5591],{},"Meet Ming: a kitchen apprentice with a wall of 4,192 dials and 32,000 omakase records. (4,192 is not a rounded number chosen for effect — it's the exact parameter count in Karpathy's 200-line code.)",[758,5593,5594],{},"The 32,000 records are the master's complete portfolio — every omakase set ever designed, every course written down in order. Ming has studied all of them.",[758,5596,5597,5598],{},"The 4,192 dials are something different. They're in Ming's head. Each dial encodes one tiny piece of his learned judgment: ",[2156,5599,5600],{},"\"How much should a rich third course influence what comes fourth? How strongly does a seafood-heavy opening push me toward something different mid-meal? How much does texture contrast matter across a full set?\"",[758,5602,5603],{},"4,192 tiny opinions. Together, they are his entire culinary intuition.",[758,5605,5606,5607,5610,5611,5614],{},"In this kitchen, Ming is the sole decision-maker — every judgment about what the next course should be is his. You'll meet two other characters along the way. Brief introductions: the ",[778,5608,5609],{},"sous chefs"," aren't separate people — they're Ming's thinking assistants, offering input but never making the final call. ",[778,5612,5613],{},"Adam"," is a training sensei who handles all the dial adjustments during practice, then steps away once real service begins. When diners sit down, it's just Ming.",[758,5616,5617],{},"He starts knowing nothing. Every dial is set to a random position. By the end of training, those dials will encode something close to a master's taste.",[758,5619,5620],{},"This is how ChatGPT works. Not a metaphor. Literally.",[758,5622,5623],{},"Andrej Karpathy proved it in 200 lines of code. Today I'll walk you through the kitchen.",[2163,5625],{},[753,5627,5629],{"id":5628},"chapter-1-the-ingredients-32000-omakase-records","Chapter 1: The Ingredients — 32,000 Omakase Records",[758,5631,5632],{},"Ming has the master's portfolio: 32,000 complete omakase sets, every course written down in sequence.",[758,5634,5635],{},"\"emma\" is a 4-course set: e → m → m → a. \"sophia\" is a 6-course set: s → o → p → h → i → a. Ming can see what was served, and in what order. No recipes. No explanations. Just the sequence.",[758,5637,5638,5639,5642],{},"His goal: after studying enough sets, design ",[2156,5640,5641],{},"new"," omakase that make diners say \"this feels right.\"",[758,5644,5645,5646,5649],{},"One thing worth noting: in this simplified kitchen, the menu has only 27 items, each labeled with a letter code. \"e\" isn't \"eel nigiri\" — it's just Course #5 in a 27-item repertoire. The codes don't carry meaning on their own. What matters is purely the ",[2156,5647,5648],{},"sequence",": which code follows which, and what patterns emerge across thousands of sets.",[758,5651,5652,5655],{},[778,5653,5654],{},"Why does sequence matter so much?"," Because the same four courses — served as sashimi → grilled → soup → dessert — create a completely different experience than the same four courses in a different order. The omakase \"emma\" and the sequence \"amme\" are entirely different experiences, even though they contain the same ingredients. In language, \"dog bites man\" and \"man bites dog\" are the same four words, entirely different meanings.",[758,5657,5658,5661],{},[778,5659,5660],{},"The extension:"," ChatGPT's portfolio isn't 32,000 names. It's the entire internet — every book, article, conversation, and post ever written. Trillions of \"omakase sets,\" all consumed in sequence.",[758,5663,5664,5665,5668],{},"The key feeling: Ming doesn't memorize every sequence he's seen. He develops ",[2156,5666,5667],{},"intuition"," — a feel for what tends to follow what, which combinations feel \"right,\" which feel off. There's no rulebook. Just pattern, absorbed until it becomes instinct.",[758,5670,5671],{},[3036,5672],{"alt":5673,"src":5674},"32,000 omakase records as scrolls — one unrolled showing the sequence e→m→m→a. A callout shows ChatGPT's trillions of sequences.","/blogs-img/2026-03-01-chatgpt-kitchen-02.png",[3590,5676,5677],{},[758,5678,5679],{},[2156,5680,5681],{},"Ming isn't memorizing. He's developing taste.",[2163,5683],{},[753,5685,5687],{"id":5686},"chapter-2-breaking-it-down-one-course-at-a-time","Chapter 2: Breaking It Down — One Course at a Time",[758,5689,5690],{},"A full omakase is too complex to learn all at once. The first step is to break it into individual courses.",[758,5692,5693,5694],{},"\"emma\" becomes: ",[844,5695,5696],{},"🛎️, e, m, m, a, 🛎️",[758,5698,5699,5700,5703],{},"That service bell 🛎️ marks the start and end of every set. Each individual course is a ",[778,5701,5702],{},"token"," — the smallest unit of meaning the kitchen works with.",[758,5705,5706],{},"But here's the thing: Ming's kitchen doesn't understand names. It only understands numbers. So every course gets a number:",[758,5708,5709],{},"a = 0, b = 1, c = 2 ... z = 25, 🛎️ = 26",[758,5711,5712,5713],{},"The omakase \"emma\" becomes: ",[844,5714,5715],{},"26, 4, 12, 12, 0, 26",[758,5717,5718,5719,5722,5723,5726,5727,5730],{},"Numbers alone still aren't enough. A number is just a label — it tells you ",[2156,5720,5721],{},"which"," course, but nothing about its ",[2156,5724,5725],{},"character",". So each number also comes with a ",[778,5728,5729],{},"flavor profile card",": 16 numbers describing properties like richness, temperature, texture, and acidity. These profiles start completely random. Through training, they become meaningful.",[758,5732,5733],{},[3036,5734],{"alt":5735,"src":5736},"Tokenization pipeline: \"emma\" → tokens with service bells → numbers 26,4,12,12,0,26 → flavor profile card with 16 slots","/blogs-img/2026-03-01-chatgpt-kitchen-03.png",[758,5738,5739],{},"For ChatGPT, the numbering is smarter — common word combinations get a single number instead of being spelled out letter by letter. \"hello\" is one dish — a single course code — not five letters served separately. More efficient, more expressive. About 100,000 possible \"courses\" instead of 27.",[3590,5741,5742],{},[758,5743,5744],{},[2156,5745,5746],{},"Every word you type becomes a number. Every number carries a flavor profile. That's how a machine starts to understand language.",[2163,5748],{},[753,5750,5752],{"id":5751},"chapter-3-the-kitchen-assembly-line-what-course-comes-next","Chapter 3: The Kitchen Assembly Line — What Course Comes Next?",[758,5754,5755,5756],{},"Here's where everything comes together. Ming's task at every moment is to answer one question: ",[778,5757,5758],{},"what should the next course be?",[758,5760,5761],{},"Say the set \"emma\" is underway. The first two courses (e, m) have been served. Course #3 — another \"m\" — has just arrived at the kitchen. Ming's job: decide what #4 should be.",[758,5763,5764],{},"This course passes through three key stations. The answer comes out the other side.",[834,5766,5769],{"className":5767,"code":5768,"language":3579},[3577],"Course #3 \"m\" arrives\n        ↓\n[1] Look backward — Sous Chef Roundtable  ← the only place you look back\n        ↓\n[2] Think it through — Back Kitchen\n        ↓\n[3] Place your bet — Final Vote\n        ↓\nProbability ranking of all 27 possible next courses\n",[841,5770,5768],{"__ignoreMap":839},[758,5772,5773],{},[3036,5774],{"alt":5775,"src":5776},"The three-station kitchen assembly line: Sous Chef Roundtable → Back Kitchen → Final Vote, with 27 possible next courses as output","/blogs-img/2026-03-01-chatgpt-kitchen-04.png",[758,5778,5779],{},"Three steps. Between each one, two backstage actions happen quietly. Let's understand the three steps first, then the two supporting roles.",[2163,5781],{},[826,5783,5785],{"id":5784},"step-1-look-backward-the-sous-chef-roundtable","Step 1: Look Backward — The Sous Chef Roundtable",[758,5787,5788],{},"This is the only station in the kitchen that looks at previous courses.",[758,5790,5791],{},"Course #3 \"m\" is placed at the center of the table. Four sous chefs sit around it, each tracking a different quality of the meal so far:",[804,5793,5794,5797,5800],{},[775,5795,5796],{},"Sous Chef A tracks richness, consulting her notebook: \"Course 1 was light, course 2 was rich...\"",[775,5798,5799],{},"Sous Chef B tracks texture: \"Two soft courses in a row — time for something with crunch?\"",[775,5801,5802],{},"Sous Chef C tracks temperature. Sous Chef D tracks acidity.",[758,5804,5805,5806,5809,5810,5813],{},"Each sous chef asks: ",[2156,5807,5808],{},"\"Which previous course is most relevant to what I'm tracking right now?\""," The previous two courses respond in turn: ",[2156,5811,5812],{},"\"Here's my profile.\""," Good matches pass their full details forward. Poor matches get skipped.",[758,5815,5816],{},"All four combine their findings and pass them to the next station.",[758,5818,5819],{},"One detail worth noting: the sous chefs don't re-taste old courses. From the very first course, they've been writing everything down in notebooks. Each new course adds a page — this is the kitchen's memory system, and nothing ever needs to be re-read from scratch.",[758,5821,5822],{},[3036,5823],{"alt":5824,"src":5825},"Top-down view of the sous chef roundtable: four sous chefs (A=Richness, B=Texture, C=Temperature, D=Acidity) around the central dish \"Course #3: m\"","/blogs-img/2026-03-01-chatgpt-kitchen-05.png",[3590,5827,5828],{},[758,5829,5830],{},[2156,5831,5832],{},"The roundtable does one thing: tell Ming which parts of the past are most relevant to the decision at hand.",[2163,5834],{},[826,5836,5838],{"id":5837},"step-2-think-it-through-the-back-kitchen","Step 2: Think It Through — The Back Kitchen",[758,5840,5841],{},"The sous chefs have done their part. Now Ming thinks for himself.",[758,5843,5844],{},"The back kitchen has a working method: it first opens the problem completely — entertaining every possible direction at once in a much larger mental space. Then it filters: anything that clearly doesn't work gets zeroed out. Finally, it compresses back down into a clear conclusion.",[758,5846,5847],{},"Open wide, filter, compress. The result is a refined judgment about what course #4 should be.",[758,5849,5850],{},[3036,5851],{"alt":5852,"src":5853},"Back kitchen method: Open Wide (fan of arrows) → Filter (crossed-out paths) → Compress (converging single arrow)","/blogs-img/2026-03-01-chatgpt-kitchen-06.png",[3590,5855,5856],{},[758,5857,5858],{},[2156,5859,5860],{},"The roundtable looks outward: what does the past tell me? The back kitchen looks inward: what do I actually think?",[2163,5862],{},[826,5864,5866],{"id":5865},"step-3-place-your-bet-the-final-vote","Step 3: Place Your Bet — The Final Vote",[758,5868,5869],{},"Back kitchen done. Ming scores all 27 possible next courses:",[758,5871,5872],{},[2156,5873,5874],{},"\"70% chance course #4 should be 'a'. 15% it's 'e'. 8% it's 'i'...\"",[758,5876,5877],{},"Raw scores get converted into clean percentages that sum to 100%. Not a certain answer — a confident bet.",[3590,5879,5880],{},[758,5881,5882],{},[2156,5883,5884],{},"Ming doesn't know the answer. He makes a bet. The whole magic is in how he gets better at betting.",[2163,5886],{},[826,5888,5890],{"id":5889},"two-supporting-roles","Two Supporting Roles",[758,5892,5893],{},"Two other actions happen throughout the process — not the main story, but the kitchen breaks without them.",[758,5895,5896,5899],{},[778,5897,5898],{},"Palate Cleanser (before each main station):"," Before entering the roundtable or the back kitchen, the palate gets reset. This ensures a consistent baseline for each judgment. Taste something intensely spicy, then immediately evaluate something delicate — your perception is off. The cleanser means every judgment starts from the same neutral point.",[758,5901,5902,5905],{},[778,5903,5904],{},"Spoonful of the Original (after each main station):"," After the roundtable, and after the back kitchen, one action happens: the course's original profile gets mixed back in. Like making a reduction — always keep a pot of the original stock nearby, add some back after each step. This prevents the course from losing itself through layer after layer of processing.",[758,5907,5908],{},"In Ming's kitchen (MicroGPT), these three steps run once per course — a single pass. In ChatGPT's kitchen, the same three steps are stacked dozens of times: the output of one full pass becomes the input of the next. Same sous chefs, same back kitchen — but each round reaches a deeper level of understanding.",[758,5910,5911],{},[3036,5912],{"alt":5913,"src":5914},"Two supporting roles: Palate Cleanser (reset to neutral before each station) and Spoonful of Original (mix back in after each station)","/blogs-img/2026-03-01-chatgpt-kitchen-07.png",[2163,5916],{},[753,5918,5920],{"id":5919},"chapter-4-the-diners-scorecard-how-bad-did-you-do","Chapter 4: The Diner's Scorecard — How Bad Did You Do?",[758,5922,5923],{},"Ming designs an omakase. The diner scores each course transition.",[804,5925,5926,5929,5932],{},[775,5927,5928],{},"If Ming was 100% confident the next course should be \"a\" — and it was — perfect score, zero penalty.",[775,5930,5931],{},"If he only gave \"a\" a 10% chance — big penalty.",[775,5933,5934],{},"If he gave it a 0.1% chance — massive penalty.",[758,5936,5937,5938,5941],{},"This penalty is the ",[778,5939,5940],{},"Loss",". Lower loss = better cooking.",[758,5943,5944],{},"Starting loss: 3.3. That's what pure random guessing looks like when you have 27 options. A complete kitchen novice, eyes closed, putting dishes down at random.",[758,5946,5947],{},[3036,5948],{"alt":5949,"src":5950},"The Diner's Scorecard: three confidence levels with penalties, and the loss curve dropping from 3.3 to 2.37 over 1,000 dishes","/blogs-img/2026-03-01-chatgpt-kitchen-08.png",[758,5952,5953,5954],{},"The entire training process has one mission: ",[778,5955,5956],{},"get that number down.",[3590,5958,5959],{},[758,5960,5961],{},[2156,5962,5963],{},"Every decision Ming makes can be judged with a single number. That number is the north star.",[2163,5965],{},[753,5967,5969],{"id":5968},"chapter-5-tracing-the-problem-who-made-it-too-salty","Chapter 5: Tracing the Problem — Who Made It Too Salty?",[758,5971,5972],{},"This is the core of everything — the part that makes the whole system actually intelligent.",[758,5974,5975],{},"The score comes back: \"Terrible.\" Ming is staring at a wall of 4,192 dials with no idea where to start. Test them one by one? That would take years.",[758,5977,5978,5979],{},"But there's something smarter: ",[778,5980,5981],{},"trace backwards from the plate.",[758,5983,5984],{},"The diner takes a bite: \"Too salty.\"",[826,5986,5988],{"id":5987},"from-plate-to-kitchen","From Plate to Kitchen",[758,5990,5991],{},"\"This bite is too salty → last step was plating (no salt added there) → step before was the back kitchen (added soy sauce) → step before was the roundtable (referenced course 2's profile) → traced back to dial #347 (soy sauce intensity).\"",[758,5993,5994,5995,5998],{},"This is ",[778,5996,5997],{},"backpropagation"," — following the chain of cause and effect backward from the result to the source.",[758,6000,6001],{},[3036,6002],{"alt":6003,"src":6004},"Backpropagation chain: \"Too Salty!\" traced backwards through Plating → Back Kitchen → Roundtable → Dial #347 \"soy sauce intensity\"","/blogs-img/2026-03-01-chatgpt-kitchen-09.png",[826,6006,6008],{"id":6007},"the-relay-chain-rule","The Relay (Chain Rule)",[758,6010,6011,6012],{},"Each station knows one simple thing: ",[2156,6013,6014],{},"\"If my input changes by a tiny amount, how much does my output change?\"",[804,6016,6017,6020,6023],{},[775,6018,6019],{},"Turn dial #347 up by 1 → broth saltiness increases by 3",[775,6021,6022],{},"Broth saltiness up by 1 → final taste changes by 2",[775,6024,6025,6026],{},"Dial #347's total impact on final taste = 3 × 2 = ",[778,6027,6028],{},"6",[758,6030,6031],{},"Walk along the assembly line and multiply at each step. That's the chain rule — no calculus required, just multiplication along a path.",[826,6033,6035],{"id":6034},"six-basic-techniques","Six Basic Techniques",[758,6037,6038],{},"The entire kitchen uses only six cooking operations. Every checkpoint, every station, every calculation reduces to one of these six. And crucially — each one knows exactly how to trace backwards through itself. That's what makes it possible to walk from the plate all the way back to dial #347.",[5072,6040,6041,6051],{},[5075,6042,6043],{},[5078,6044,6045,6048],{},[5081,6046,6047],{},"Technique",[5081,6049,6050],{},"What It Does",[5088,6052,6053,6064,6075,6086,6097,6108],{},[5078,6054,6055,6061],{},[5093,6056,6057,6060],{},[778,6058,6059],{},"Combine"," (addition)",[5093,6062,6063],{},"Pour two things together",[5078,6065,6066,6072],{},[5093,6067,6068,6071],{},[778,6069,6070],{},"Blend"," (multiplication)",[5093,6073,6074],{},"Two ingredients amplify each other",[5078,6076,6077,6083],{},[5093,6078,6079,6082],{},[778,6080,6081],{},"Reduce"," (power)",[5093,6084,6085],{},"Concentrate the flavor",[5078,6087,6088,6094],{},[5093,6089,6090,6093],{},[778,6091,6092],{},"Extract"," (log)",[5093,6095,6096],{},"A little goes a long way",[5078,6098,6099,6105],{},[5093,6100,6101,6104],{},[778,6102,6103],{},"Ferment"," (exp)",[5093,6106,6107],{},"Exponential growth",[5078,6109,6110,6116],{},[5093,6111,6112,6115],{},[778,6113,6114],{},"Quality Check"," (ReLU)",[5093,6117,6118],{},"Good flavors pass, bad ones get dumped",[758,6120,6121],{},"ChatGPT's entire kitchen uses only these six. There is no seventh.",[758,6123,6124],{},[3036,6125],{"alt":6126,"src":6127},"Six kitchen techniques in a 2×3 grid: Combine, Blend, Reduce, Extract, Ferment, Quality Check — each as a hand-drawn cooking icon","/blogs-img/2026-03-01-chatgpt-kitchen-10.png",[826,6129,6131],{"id":6130},"a-concrete-example","A Concrete Example",[758,6133,6134],{},"Ming makes a simple two-step dish:",[804,6136,6137,6140,6146],{},[775,6138,6139],{},"Ingredients: 2 parts salt, 3 parts sugar",[775,6141,6142,6143,6145],{},"Step 1: ",[778,6144,6070],{}," → salt-sugar base = 6",[775,6147,6148,6149,6151],{},"Step 2: ",[778,6150,6059],{}," → final taste = base + extra pinch of salt = 8",[758,6153,6154],{},"Diner says the taste is off. Trace back:",[804,6156,6157,6160],{},[775,6158,6159],{},"Final taste = base + salt → Combine → base's impact = 1, salt's impact = 1",[775,6161,6162],{},"Base = salt × sugar → Blend → salt's impact here = sugar's value = 3",[758,6164,6165,6166,6169],{},"Salt appeared at two steps. Total impact = 3 + 1 = ",[778,6167,6168],{},"4",". Sugar appeared once: impact = 2.",[758,6171,6172,6173,6176],{},"Now Ming knows: salt influences the outcome twice as much as sugar. That's what ",[778,6174,6175],{},"gradient = 4"," means. Adjust salt carefully. Be bolder with sugar.",[3590,6178,6179],{},[758,6180,6181],{},[2156,6182,6183],{},"Ming doesn't guess which dial to turn. He calculates exactly how much each one matters.",[2163,6185],{},[753,6187,6189],{"id":6188},"chapter-6-the-masters-training-method-1000-dishes","Chapter 6: The Master's Training Method — 1,000 Dishes",[758,6191,6192],{},"Ming drills. Every dish, the same three steps:",[772,6194,6195,6201,6207],{},[775,6196,6197,6200],{},[778,6198,6199],{},"Cook (forward pass)"," — run the current dish through the assembly line with current dial positions",[775,6202,6203,6206],{},[778,6204,6205],{},"Trace (backward pass)"," — after the diner scores it, trace from plate back through every checkpoint to find each dial's impact number",[775,6208,6209,6212],{},[778,6210,6211],{},"Adjust (update)"," — based on each dial's impact number, nudge it in the direction that reduces the penalty",[758,6214,6215],{},"The first two steps have clear rules to follow. The third is where it gets genuinely hard: 4,192 dials — how do you adjust them? Same amount for each? Which ones need a light touch, which ones can take more? Bold moves early, or cautious moves throughout?",[758,6217,6218],{},"Ming can't figure this out alone. So during training, a sensei stands beside him — a specialist focused entirely on fine-tuning the dials. His name is Adam.",[826,6220,6222],{"id":6221},"adam-the-sensei","Adam the Sensei",[758,6224,6225],{},"Not every dial gets the same treatment. Adam is smarter than \"turn everything by the same amount.\"",[758,6227,6228,6231],{},[778,6229,6230],{},"Memory (Momentum):"," Adam remembers the last several rounds. If the last five rounds all said \"turn down dial #347,\" round six goes bolder. Like a ball rolling downhill — it builds momentum instead of bouncing back and forth.",[758,6233,6234,6237],{},[778,6235,6236],{},"Personalized touch (Adaptive Learning Rate):"," Some dials are hypersensitive — a tiny turn causes a huge change in the final taste. Adam adjusts those gently. Some dials are stubborn — big turns barely matter. Adam cranks those harder.",[758,6239,6240,6243],{},[778,6241,6242],{},"Lighter over time (Learning Rate decay):"," Bold adjustments early in training — you're so far from good that aggressive moves are fine. Increasingly delicate adjustments as you approach excellence. The closer to perfect, the more carefully you tune.",[826,6245,6247],{"id":6246},"the-growth-curve","The Growth Curve",[5072,6249,6250,6262],{},[5075,6251,6252],{},[5078,6253,6254,6257,6259],{},[5081,6255,6256],{},"Dish",[5081,6258,5940],{},[5081,6260,6261],{},"What's Happening in the Kitchen",[5088,6263,6264,6275,6286,6297],{},[5078,6265,6266,6269,6272],{},[5093,6267,6268],{},"#1",[5093,6270,6271],{},"3.3",[5093,6273,6274],{},"Blindfolded — dessert after sashimi, deep-fried everything, pure random chaos",[5078,6276,6277,6280,6283],{},[5093,6278,6279],{},"#100",[5093,6281,6282],{},"~2.8",[5093,6284,6285],{},"Learned \"omakase usually starts light, ends rich\"",[5078,6287,6288,6291,6294],{},[5093,6289,6290],{},"#500",[5093,6292,6293],{},"~2.5",[5093,6295,6296],{},"Learned \"miso and rice always appear together,\" \"never three fried courses in a row\"",[5078,6298,6299,6302,6305],{},[5093,6300,6301],{},"#1000",[5093,6303,6304],{},"2.37",[5093,6306,6307],{},"Designing omakase that diners actually believe in",[758,6309,6310,6313],{},[778,6311,6312],{},"Ming never memorized a single rule."," Nobody told him \"don't serve three fried courses back-to-back.\" Nobody explained that miso pairs with rice. He just ran the cook → trace → adjust loop until the dials naturally settled into positions that reflect those patterns.",[758,6315,6316],{},"That's what \"learning intuition\" means.",[758,6318,6319],{},[3036,6320],{"alt":6321,"src":6322},"Training loop: Cook → Score → Trace → Adjust (1,000 rounds), with Adam the Sensei's three methods: Memory, Personalized Touch, Lighter Over Time","/blogs-img/2026-03-01-chatgpt-kitchen-11.png",[3590,6324,6325],{},[758,6326,6327],{},[2156,6328,6329],{},"Intelligence didn't come from rules. It came from repetition.",[2163,6331],{},[753,6333,6335],{"id":6334},"chapter-7-graduation-ming-cooks-solo","Chapter 7: Graduation — Ming Cooks Solo",[758,6337,6338],{},"Training's over. Dials locked in their final positions. This isn't practice anymore — real diners are here.",[772,6340,6341,6344,6347,6350,6353],{},[775,6342,6343],{},"Service bell rings 🛎️ — \"begin a new omakase\"",[775,6345,6346],{},"Ming designs the first course based on current dial positions",[775,6348,6349],{},"That course travels through the assembly line, and from it, Ming designs the second",[775,6351,6352],{},"The second informs the third. The third informs the fourth.",[775,6354,6355],{},"Continues until Ming naturally produces the closing bell 🛎️",[758,6357,6358],{},"Every course is improvised — in the sense that no predetermined menu exists — but informed by everything that came before. Just like a great omakase chef reading the table: the first course shapes the second, the mood of the meal shapes the end.",[826,6360,6362],{"id":6361},"the-risk-dial-temperature","The Risk Dial (Temperature)",[758,6364,6365,6366,2480],{},"There's one dial by the kitchen door that doesn't change skill level — it changes ",[2156,6367,2104],{},[804,6369,6370,6376,6382],{},[775,6371,6372,6375],{},[778,6373,6374],{},"Temperature = 0.1"," (ultra-safe) → always picks the statistically safest choice → predictable, solid, a little boring",[775,6377,6378,6381],{},[778,6379,6380],{},"Temperature = 0.5"," (slightly adventurous) → mostly coherent with surprises",[775,6383,6384,6387],{},[778,6385,6386],{},"Temperature = 1.5"," (bold risk-taker) → might be brilliant, might be incomprehensible",[758,6389,6390],{},[3036,6391],{"alt":6392,"src":6393},"Temperature dial spanning from 0.1 (Ultra-Safe) through 0.7 (ChatGPT marked with a star) to 1.5 (Bold Risk-Taker), with example outputs below each","/blogs-img/2026-03-01-chatgpt-kitchen-12.png",[758,6395,6396],{},"ChatGPT runs at around 0.7. A little creative, but won't go off the rails.",[826,6398,6400],{"id":6399},"on-inventing-dishes-hallucination","On \"Inventing Dishes\" (Hallucination)",[758,6402,6403],{},"After training, Ming generates courses called \"Karia,\" \"Yeran,\" \"Liole.\" These are actual outputs from Karpathy's trained model — names that sound plausible, feel right, could belong to a real person. Most probably don't exist anywhere in the training data.",[758,6405,6406],{},"He's not lying. He's following patterns he internalized — combinations that feel statistically plausible. He generates them with complete confidence. But they were never real.",[758,6408,6409,6410,6413,6414,2480],{},"When ChatGPT confidently cites a paper that doesn't exist, or invents a specific date or name, it's doing exactly what Ming does with \"karia.\" It has no fact-checking station. It only knows what ",[2156,6411,6412],{},"tastes right",", not what ",[2156,6415,6416],{},"is real",[3590,6418,6419],{},[758,6420,6421],{},[2156,6422,6423],{},"Ming can design a flawless omakase using an ingredient that was never harvested. That's the trade-off built into how the kitchen works.",[2163,6425],{},[753,6427,6429],{"id":6428},"chapter-8-from-neighborhood-cart-to-michelin-3-star","Chapter 8: From Neighborhood Cart to Michelin 3-Star",[758,6431,6432],{},"Ming learned to cook with 4,192 judgment dials and 32,000 training records. What about ChatGPT?",[5072,6434,6435,6447],{},[5075,6436,6437],{},[5078,6438,6439,6441,6444],{},[5081,6440],{},[5081,6442,6443],{},"Ming (Neighborhood Omakase)",[5081,6445,6446],{},"ChatGPT (Michelin 3-Star)",[5088,6448,6449,6460,6471,6482,6493,6504],{},[5078,6450,6451,6454,6457],{},[5093,6452,6453],{},"Judgment dials",[5093,6455,6456],{},"4,192",[5093,6458,6459],{},"Hundreds of billions",[5078,6461,6462,6465,6468],{},[5093,6463,6464],{},"Training records",[5093,6466,6467],{},"32,000 names",[5093,6469,6470],{},"Trillions of sequences (the entire internet)",[5078,6472,6473,6476,6479],{},[5093,6474,6475],{},"Course vocabulary",[5093,6477,6478],{},"27 letters + a bell",[5093,6480,6481],{},"~100,000 word chunks",[5078,6483,6484,6487,6490],{},[5093,6485,6486],{},"Kitchen",[5093,6488,6489],{},"One stove (a MacBook)",[5093,6491,6492],{},"Thousands of burners in parallel (GPU cluster)",[5078,6494,6495,6498,6501],{},[5093,6496,6497],{},"Training time",[5093,6499,6500],{},"1 minute",[5093,6502,6503],{},"Months",[5078,6505,6506,6511,6516],{},[5093,6507,6508],{},[778,6509,6510],{},"Cooking principles",[5093,6512,6513],{},[778,6514,6515],{},"Identical",[5093,6517,6518],{},[778,6519,6515],{},[758,6521,6522],{},[3036,6523],{"alt":6524,"src":6525},"Scale comparison: Ming's humble street cart (4,192 dials, MacBook, 32,000 records) vs. ChatGPT's Michelin kitchen (billions of dials, thousands of GPUs, trillions of records). Footer: \"Same cooking principles. Wildly different scale.\"","/blogs-img/2026-03-01-chatgpt-kitchen-13.png",[758,6527,6528],{},"Same kitchen. Wildly different scale.",[826,6530,6532],{"id":6531},"the-3-star-kitchens-extra-steps","The 3-Star Kitchen's Extra Steps",[758,6534,6535],{},"ChatGPT didn't stop at basic training. Two additional stages took it from technically correct to genuinely useful.",[758,6537,6538,6541],{},[778,6539,6540],{},"Stage 1 — Switch the Menu (SFT):"," Ming first trained on simple name sequences to develop his foundational sequencing feel. Then he switched to multi-turn conversations — complex back-and-forths — and kept training. Same cook → trace → adjust algorithm. Different training material. A chef who masters eggs before moving to haute cuisine. The fundamentals don't change; the repertoire expands.",[758,6543,6544,6547],{},[778,6545,6546],{},"Stage 2 — Bring in the Critics (RLHF):"," Ming designs two dishes. A critic chooses the better one. The dials adjust based on critic preference, repeated millions of times.",[758,6549,6550],{},"This is why ChatGPT is \"polite\" and \"helpful\" — not because a rule was programmed in. Because human critics, across millions of comparisons, consistently selected responses that felt helpful and respectful. Those preferences got baked into the dials.",[758,6552,6553,6554],{},"Through all of it, the core never changes: ",[778,6555,6556],{},"cook → score → trace → adjust.",[3590,6558,6559],{},[758,6560,6561],{},[2156,6562,6563],{},"From 200 lines to hundreds of billions of parameters: same six techniques, same three steps. Just more dials, more dishes, more diners.",[2163,6565],{},[753,6567,6569],{"id":6568},"the-kitchen-simplified","The Kitchen, Simplified",[758,6571,6572],{},"Next time someone tells you AI is mysterious, terrifying, or about to replace humanity — remember the diner sitting down at that omakase counter.",[758,6574,6575],{},"No menu. Dishes arriving with quiet confidence. Everything feeling inevitable.",[758,6577,6578],{},"Behind that experience: one apprentice, a wall of judgment dials, and 32,000 records. No rules given. Just cook → score → trace → adjust. Do it enough times, and the dials settle.",[758,6580,6581],{},"This isn't intelligence. It's statistical intuition so refined it becomes indistinguishable from taste.",[758,6583,6584,6585,6588],{},"But the fact that such a simple mechanism — six operations, three steps, one loop — produces something that can hold a conversation, explain a concept, write a story, and feel ",[2156,6586,6587],{},"present"," in the exchange?",[758,6590,6591],{},"That's what's truly awe-inspiring.",[758,6593,6594],{},"200 lines of code. Six techniques. One kitchen.",[758,6596,6597],{},"That's all of ChatGPT.",[2163,6599],{},[758,6601,6602],{},[2156,6603,6604,6605,6608],{},"If you want to see the actual 200 lines, Andrej Karpathy's ",[2177,6606,5551],{"href":5549,"rel":6607},[2181]," is worth the read. He built the kitchen. I just gave it a name and walked you through it.",[2163,6610],{},[753,6612,6614],{"id":6613},"metaphor-reference","Metaphor Reference",[5072,6616,6617,6627],{},[5075,6618,6619],{},[5078,6620,6621,6624],{},[5081,6622,6623],{},"GPT Concept",[5081,6625,6626],{},"Kitchen Metaphor",[5088,6628,6629,6637,6645,6653,6661,6669,6677,6685,6693,6701,6709,6717,6725,6733,6741,6749,6757,6764,6772,6780,6788,6796,6804,6812,6820],{},[5078,6630,6631,6634],{},[5093,6632,6633],{},"Model",[5093,6635,6636],{},"Ming, the apprentice chef",[5078,6638,6639,6642],{},[5093,6640,6641],{},"Parameters",[5093,6643,6644],{},"4,192 judgment dials — encoding Ming's learned intuition about sequence",[5078,6646,6647,6650],{},[5093,6648,6649],{},"Dataset",[5093,6651,6652],{},"32,000 omakase records from the master (not recipes — just the sequences)",[5078,6654,6655,6658],{},[5093,6656,6657],{},"Token",[5093,6659,6660],{},"A single course in the omakase",[5078,6662,6663,6666],{},[5093,6664,6665],{},"Tokenizer",[5093,6667,6668],{},"The system that breaks a full set into individual courses",[5078,6670,6671,6674],{},[5093,6672,6673],{},"BOS / EOS Token",[5093,6675,6676],{},"The service bell 🛎️ — \"new set begins\" / \"set complete\"",[5078,6678,6679,6682],{},[5093,6680,6681],{},"Embedding",[5093,6683,6684],{},"Flavor profile card — 16 numbers encoding a course's character",[5078,6686,6687,6690],{},[5093,6688,6689],{},"Position Embedding",[5093,6691,6692],{},"Sequence tag — \"this is course #3\"",[5078,6694,6695,6698],{},[5093,6696,6697],{},"Attention",[5093,6699,6700],{},"Sous chefs consulting their notebooks on previous courses",[5078,6702,6703,6706],{},[5093,6704,6705],{},"Query / Key / Value",[5093,6707,6708],{},"Q = \"what do I need?\" K = \"here's what I offered\" V = \"here's my recipe if you pick me\"",[5078,6710,6711,6714],{},[5093,6712,6713],{},"Multi-head Attention",[5093,6715,6716],{},"4 sous chefs, each tracking a different dimension simultaneously",[5078,6718,6719,6722],{},[5093,6720,6721],{},"KV Cache",[5093,6723,6724],{},"The notebooks — no need to re-taste what's already been recorded",[5078,6726,6727,6730],{},[5093,6728,6729],{},"MLP",[5093,6731,6732],{},"The back kitchen — think for yourself after consulting the room",[5078,6734,6735,6738],{},[5093,6736,6737],{},"ReLU",[5093,6739,6740],{},"Quality control — bad flavors get zeroed out, good ones pass through",[5078,6742,6743,6746],{},[5093,6744,6745],{},"Residual Connection",[5093,6747,6748],{},"The spoonful of original stock mixed back in after every step",[5078,6750,6751,6754],{},[5093,6752,6753],{},"RMSNorm",[5093,6755,6756],{},"Palate cleanser before each station",[5078,6758,6759,6761],{},[5093,6760,5940],{},[5093,6762,6763],{},"The diner's penalty score — how far from perfect?",[5078,6765,6766,6769],{},[5093,6767,6768],{},"Softmax",[5093,6770,6771],{},"Converting raw scores into ranked percentages",[5078,6773,6774,6777],{},[5093,6775,6776],{},"Backpropagation",[5093,6778,6779],{},"Tracing \"too salty\" back through every station to dial #347",[5078,6781,6782,6785],{},[5093,6783,6784],{},"Gradient",[5093,6786,6787],{},"Each dial's influence number — how much does it shift the final result?",[5078,6789,6790,6793],{},[5093,6791,6792],{},"Chain Rule",[5093,6794,6795],{},"Multiply the impact at each step along the assembly line",[5078,6797,6798,6801],{},[5093,6799,6800],{},"Adam Optimizer",[5093,6802,6803],{},"The master chef who adjusts with memory, personalization, and a lighter touch over time",[5078,6805,6806,6809],{},[5093,6807,6808],{},"Temperature",[5093,6810,6811],{},"The risk dial — low = safe, high = creative",[5078,6813,6814,6817],{},[5093,6815,6816],{},"Hallucination",[5093,6818,6819],{},"Inventing \"karia\" — statistically plausible, never existed",[5078,6821,6822,6825],{},[5093,6823,6824],{},"MicroGPT → ChatGPT",[5093,6826,6827],{},"Neighborhood cart → Michelin 3-star — same principles, wildly different scale",{"title":839,"searchDepth":708,"depth":708,"links":6829},[6830,6831,6832,6838,6839,6845,6849,6853,6856,6857],{"id":5628,"depth":708,"text":5629},{"id":5686,"depth":708,"text":5687},{"id":5751,"depth":708,"text":5752,"children":6833},[6834,6835,6836,6837],{"id":5784,"depth":110,"text":5785},{"id":5837,"depth":110,"text":5838},{"id":5865,"depth":110,"text":5866},{"id":5889,"depth":110,"text":5890},{"id":5919,"depth":708,"text":5920},{"id":5968,"depth":708,"text":5969,"children":6840},[6841,6842,6843,6844],{"id":5987,"depth":110,"text":5988},{"id":6007,"depth":110,"text":6008},{"id":6034,"depth":110,"text":6035},{"id":6130,"depth":110,"text":6131},{"id":6188,"depth":708,"text":6189,"children":6846},[6847,6848],{"id":6221,"depth":110,"text":6222},{"id":6246,"depth":110,"text":6247},{"id":6334,"depth":708,"text":6335,"children":6850},[6851,6852],{"id":6361,"depth":110,"text":6362},{"id":6399,"depth":110,"text":6400},{"id":6428,"depth":708,"text":6429,"children":6854},[6855],{"id":6531,"depth":110,"text":6532},{"id":6568,"depth":708,"text":6569},{"id":6613,"depth":708,"text":6614},"You've heard ChatGPT is magic. It's not. It's a kitchen apprentice named Ming, a wall of 4,192 judgment dials, and one loop — repeated a thousand times. No formulas, no code. Just a kitchen.",{"date":6860,"image":6861,"alt":6862,"tags":6863,"category":2132,"youtube":6868,"published":859},"1st Mar 2026","/blogs-img/2026-03-01-chatgpt-kitchen-cover.png","Ming the apprentice chef faces a wall of 4,192 judgment dials in a Japanese omakase kitchen",[6864,6865,6866,6867,5997],"ChatGPT","AI explained","machine learning","transformers","https://www.youtube.com/watch?v=mfvkgxE7U4M","/blogs/en/chatgpt-explained-kitchen-metaphor",{"title":5539,"description":6858},"blogs/en/22.chatgpt-explained-kitchen-metaphor","b1aHtGHXg71axsAOJmLV6sc8j14lZQBCVafMfqOEtBI",{"id":6874,"title":6875,"body":6876,"description":7251,"extension":2126,"meta":7252,"navigation":859,"ogImage":7254,"path":7261,"seo":7262,"stem":7263,"__hash__":7264},"en/blogs/en/23.i-engineered-the-law-of-attraction-with-ai.md","I Engineered the Law of Attraction with AI",{"type":750,"value":6877,"toc":7243},[6878,6881,6884,6887,6890,6892,6896,6899,6902,6905,6908,6911,6916,6919,6921,6925,6928,6953,6956,6959,6965,6968,6974,7020,7023,7029,7032,7034,7038,7041,7044,7047,7053,7059,7065,7071,7077,7080,7106,7116,7123,7126,7132,7134,7138,7141,7144,7160,7163,7169,7175,7178,7180,7184,7187,7193,7196,7199,7201,7205,7208,7211,7222,7225,7236,7238],[758,6879,6880],{},"Every morning, before I open WhatsApp or check my phone, something arrives in my inbox.",[758,6882,6883],{},"It's a digest. Around 15-20 items — papers, posts, threads, projects — each with an AI-generated summary and a relevance score. The high-scoring items are my \"must reads.\" Everything else I can skim or skip — and the ones worth keeping often make their way into my writing and blog posts.",[758,6885,6886],{},"This system — I call it Signal — runs 24/7. It monitors Hacker News, Reddit, ArXiv, X, Product Hunt, GitHub Trending, and a few other sources. Every item gets scored by AI against a profile of my interests, goals, and blind spots. Every hour it rescans. Every morning it delivers.",[758,6888,6889],{},"I built Signal because I believe in the Law of Attraction. And I wanted a better version of it.",[2163,6891],{},[753,6893,6895],{"id":6894},"the-pattern-i-couldnt-ignore","The Pattern I Couldn't Ignore",[758,6897,6898],{},"I've never been someone who dismissed the Law of Attraction as wishful thinking. I've lived it too many times to doubt it.",[758,6900,6901],{},"In university, I wanted to start a business. I had no capital, no business plan, no clear idea. What I had was obsession — I read about entrepreneurship, talked about it constantly, surrounded myself with people who were building things. A few months later, I was running a music instrument store with friends on campus. It wasn't the business I imagined. But it was a business.",[758,6903,6904],{},"Years later, I wanted to leave China and build a life abroad. I didn't have a clear path. I just kept moving toward it — learning, meeting people, taking the small steps that were available. Slowly, a series of doors opened. I immigrated to Canada.",[758,6906,6907],{},"Later, I wanted to work in finance. I took every opportunity to get closer to that world — conversations, roles, projects. And eventually, I became a partner at a finance firm.",[758,6909,6910],{},"None of these happened in a straight line. None of them happened because I simply wished hard enough. But all of them followed the same pattern:",[758,6912,6913],{},[778,6914,6915],{},"Intense focus → Immersion → A series of \"coincidences\" that lined up → Action on those coincidences.",[758,6917,6918],{},"I didn't just want things. I saturated myself in the world of what I wanted. And then things found me.",[2163,6920],{},[753,6922,6924],{"id":6923},"what-the-law-of-attraction-actually-is","What the Law of Attraction Actually Is",[758,6926,6927],{},"Strip away the mysticism and the mechanism is pretty clear:",[772,6929,6930,6936,6941,6947],{},[775,6931,6932,6935],{},[778,6933,6934],{},"Intention"," — You define what matters to you",[775,6937,6938,6940],{},[778,6939,6697],{}," — You expose yourself to relevant information and people",[775,6942,6943,6946],{},[778,6944,6945],{},"Pattern Recognition"," — Your mind starts filtering the world for relevant signals",[775,6948,6949,6952],{},[778,6950,6951],{},"Action"," — You act on the \"coincidences\" that aren't coincidences",[758,6954,6955],{},"There's actual neuroscience behind step 3. Your brain's Reticular Activating System (RAS) processes roughly 11 million bits of information per second — but your conscious mind handles only about 40. The RAS is the filter. It decides what reaches awareness and what gets ignored.",[758,6957,6958],{},"When you set a clear intention, you reprogram the filter. Suddenly you notice the article that was always there, the person you should have met months ago, the opportunity that was sitting in plain sight.",[758,6960,6961],{},[3036,6962],{"alt":6963,"src":6964},"The RAS: how intention reprograms your brain's filter, from 11M bits/sec down to ~40","/blogs-img/2026-03-08-signal-loa-01.webp",[758,6966,6967],{},"This is not magic. This is signal acquisition.",[758,6969,6970,6971],{},"And here's the part that clicked for me: ",[778,6972,6973],{},"this is exactly how a recommendation algorithm works.",[5072,6975,6976,6986],{},[5075,6977,6978],{},[5078,6979,6980,6983],{},[5081,6981,6982],{},"The Law of Attraction",[5081,6984,6985],{},"A Recommendation System",[5088,6987,6988,6996,7004,7012],{},[5078,6989,6990,6993],{},[5093,6991,6992],{},"Set an intention",[5093,6994,6995],{},"Define a user profile",[5078,6997,6998,7001],{},[5093,6999,7000],{},"Expose yourself to information",[5093,7002,7003],{},"Ingest a content feed",[5078,7005,7006,7009],{},[5093,7007,7008],{},"RAS filters for relevance",[5093,7010,7011],{},"Ranking model scores relevance",[5078,7013,7014,7017],{},[5093,7015,7016],{},"\"Coincidences\" surface",[5093,7018,7019],{},"Personalized feed appears",[758,7021,7022],{},"The Law of Attraction is a biological recommendation engine. And like any recommendation engine, it has a hard ceiling — bandwidth, processing speed, the finite hours you can read and connect and explore.",[758,7024,7025],{},[3036,7026],{"alt":7027,"src":7028},"The Law of Attraction vs a Recommendation Algorithm — the same 4-step structure","/blogs-img/2026-03-08-signal-loa-02.webp",[758,7030,7031],{},"AI doesn't have those limitations.",[2163,7033],{},[753,7035,7037],{"id":7036},"how-signal-works","How Signal Works",[758,7039,7040],{},"In early 2025 I started learning AI seriously — agents, LLMs, the whole landscape. One theme kept surfacing: the best AI researchers were pouring money into systems that build richer internal representations of the world, so machines could filter and reason more like humans do. Yann LeCun left Meta to build AMI Labs around this idea. Fei-Fei Li raised at a $5B valuation for World Labs. The direction was clear: the future of AI isn't just generating text, it's modeling what matters.",[758,7042,7043],{},"I wasn't trying to build AGI. I was trying to build a better filter for myself. That's when Signal came together.",[758,7045,7046],{},"The architecture is simple:",[758,7048,7049,7052],{},[778,7050,7051],{},"Sources"," → Hacker News, Reddit, ArXiv, X/Twitter, Product Hunt, GitHub Trending, Lobsters",[758,7054,7055,7058],{},[778,7056,7057],{},"Scoring"," → Every item is evaluated by an LLM against a personal interest profile: my current focus areas, the questions I'm trying to answer, the blind spots I want to address",[758,7060,7061,7064],{},[778,7062,7063],{},"Output"," → High-scoring items are \"must reads.\" They hit my morning digest. Everything else is available but deprioritized.",[758,7066,7067,7070],{},[778,7068,7069],{},"Cadence"," → Rescan every hour. Daily digest every morning.",[758,7072,7073],{},[3036,7074],{"alt":7075,"src":7076},"Signal system architecture — sources, AI scoring, and morning digest","/blogs-img/2026-03-08-signal-loa-03.webp",[758,7078,7079],{},"What this actually does:",[804,7081,7082,7088,7094,7100],{},[775,7083,7084,7087],{},[778,7085,7086],{},"Expands my attention bandwidth roughly 100x."," I \"read\" 500+ items a day through AI summaries. I decide which ones are worth full attention.",[775,7089,7090,7093],{},[778,7091,7092],{},"Eliminates the noise loop."," No doomscrolling. No algorithmic distraction designed to maximize my time on a platform, not my growth.",[775,7095,7096,7099],{},[778,7097,7098],{},"Manufactures lucky coincidences."," Papers I wouldn't have found. People I wouldn't have known to look for. Ideas I wouldn't have connected.",[775,7101,7102,7105],{},[778,7103,7104],{},"Compounds over time."," The sharper I define my interests, the sharper the signal gets.",[758,7107,7108,7109,7112,7113,2480],{},"This is the shift the tech world is calling the move from the ",[778,7110,7111],{},"Attention Economy"," to the ",[778,7114,7115],{},"Intention Economy",[758,7117,7118,7119,7122],{},"The attention economy — social media as we know it — optimizes for platform engagement. It captures your focus and sells it. The intention economy is different: systems that optimize for ",[2156,7120,7121],{},"your"," goals, not the platform's metrics.",[758,7124,7125],{},"Researchers at Cambridge are already warning that AI will soon intercept your developing intentions in real time and sell them to advertisers. The arms race is coming either way. The question is whether you build your own system, or let someone else build one for you.",[758,7127,7128],{},[3036,7129],{"alt":7130,"src":7131},"Attention Economy vs Intention Economy — who does the algorithm serve?","/blogs-img/2026-03-08-signal-loa-04.webp",[2163,7133],{},[753,7135,7137],{"id":7136},"the-flywheel","The Flywheel",[758,7139,7140],{},"The traditional Law of Attraction is passive: set your intention and wait for the universe to deliver. That framing always bothered me. It leaves too much to chance.",[758,7142,7143],{},"The AI-augmented version is an active flywheel:",[758,7145,7146,7149,7150,7153,7154,7156,7157,7159],{},[778,7147,7148],{},"Signal"," → captures what matters in your domain\n",[778,7151,7152],{},"Knowledge"," → deepens as you process curated information every day\n",[778,7155,6951],{}," → becomes more informed — better decisions, sharper writing, the right conversations\n",[778,7158,7148],{}," → improves as your actions generate new connections, new sources, new feedback",[758,7161,7162],{},"A concrete example: Signal started surfacing world model research in late 2025 — LeCun's papers, Jim Fan's robotics work, early coverage of AMI Labs. I went deep on it. I started writing and talking about world model applications in ways that weren't common yet. That positioned me in conversations I wouldn't have been part of otherwise. Those conversations fed new sources back into Signal.",[758,7164,7165],{},[3036,7166],{"alt":7167,"src":7168},"The Signal flywheel — Signal → Knowledge → Action → Signal","/blogs-img/2026-03-08-signal-loa-05.webp",[758,7170,5994,7171,7174],{},[778,7172,7173],{},"engineered serendipity"," — designing the conditions where valuable accidents happen more often.",[758,7176,7177],{},"It's not that the universe delivers. It's that you build a system that makes delivery more likely, more frequent, and more precisely targeted.",[2163,7179],{},[753,7181,7183],{"id":7182},"why-this-matters-now","Why This Matters Now",[758,7185,7186],{},"We're in the age of information abundance and attention scarcity. The average person encounters thousands of pieces of content every day and retains almost none of it — because the algorithms deciding what they see are optimized for platform engagement, not personal growth.",[758,7188,7189,7190,7192],{},"The people who win in this environment won't be the ones with access to the best AI tools. Everyone has access to GPT now. What matters is whether you've built a system that uses AI in service of ",[2156,7191,7121],{}," specific intentions — consistently, automatically, while you sleep.",[758,7194,7195],{},"Signal didn't change how smart I am. It changed what I'm exposed to. And over time, what you're exposed to shapes what you know, who you meet, and what you're able to do.",[758,7197,7198],{},"That's not manifestation. That's architecture.",[2163,7200],{},[753,7202,7204],{"id":7203},"build-your-own-signal","Build Your Own Signal",[758,7206,7207],{},"I didn't stop believing in the Law of Attraction. I just decided to engineer it.",[758,7209,7210],{},"The questions I'd ask anyone thinking about this:",[804,7212,7213,7216,7219],{},[775,7214,7215],{},"What are you trying to attract into your life right now?",[775,7217,7218],{},"What information would you need to see — consistently, every day — to get there?",[775,7220,7221],{},"What if you stopped leaving that to an algorithm that doesn't know you, and built one that does?",[758,7223,7224],{},"You don't need to believe in the Law of Attraction to find value here. You just need to believe in better signal processing.",[758,7226,7227,7228,7233,7234,2480],{},"I'll be sharing more about how Signal is built — the architecture, the scoring prompts, the profile design — in future posts. If that's something you want to follow, I write about building AI-native systems at ",[2177,7229,7232],{"href":7230,"rel":7231},"https://aaronguo.com",[2181],"aaronguo.com"," and in my newsletter, ",[778,7235,3956],{},[2163,7237],{},[758,7239,7240],{},[2156,7241,7242],{},"Follow for more on building AI systems that serve your intentions — not the algorithm's.",{"title":839,"searchDepth":708,"depth":708,"links":7244},[7245,7246,7247,7248,7249,7250],{"id":6894,"depth":708,"text":6895},{"id":6923,"depth":708,"text":6924},{"id":7036,"depth":708,"text":7037},{"id":7136,"depth":708,"text":7137},{"id":7182,"depth":708,"text":7183},{"id":7203,"depth":708,"text":7204},"The Law of Attraction works — so I built an AI system that does it programmatically. Signal monitors Hacker News, Reddit, ArXiv, and more, scoring every item against my personal interest profile. This is engineered serendipity.",{"date":7253,"image":7254,"alt":7255,"tags":7256,"category":3016,"youtube":7260,"published":859},"8th Mar 2026","/blogs-img/2026-03-08-signal-loa-cover.webp","A person surrounded by glowing signal streams, representing intentional information filtering",[30,7257,7258,7259],"signal","personal-systems","intention-economy","https://youtu.be/JApDiwbocck","/blogs/en/i-engineered-the-law-of-attraction-with-ai",{"title":6875,"description":7251},"blogs/en/23.i-engineered-the-law-of-attraction-with-ai","rv1-2iZ2sYi0dDH0-h-bNgfTy61SYhxrJiwDWTQHaTc",{"id":7266,"title":7267,"body":7268,"description":7360,"extension":2126,"meta":7361,"navigation":859,"ogImage":7363,"path":7370,"seo":7371,"stem":7372,"__hash__":7373},"en/blogs/en/24.in-the-age-of-ai-look-up-at-the-stars.md","In the Age of AI, Look Up at the Stars",{"type":750,"value":7269,"toc":7353},[7270,7273,7276,7280,7283,7286,7292,7296,7299,7302,7308,7312,7315,7318,7321,7327,7331,7334,7337,7340,7344,7347,7350],[758,7271,7272],{},"I keep seeing the same pattern at work: same model, same task, very different outcomes. One person keeps rewriting prompts and still gets generic output; another takes one or two rounds and gets to something usable.",[758,7274,7275],{},"This looks like a prompting gap on the surface, but I think it is really a cognitive asset gap. AI lowers the cost of execution, which makes judgment more valuable. The stronger the tool becomes, the more it reveals the business context, customer sense, product judgment, and strategic altitude behind the person using it.",[753,7277,7279],{"id":7278},"a-prompt-is-an-interface-to-business-judgment","A prompt is an interface to business judgment",[758,7281,7282],{},"A good prompt is not a magic spell. It is a person's ability to break down a problem: the goal, the constraints, the context that matters, the details that will create noise, and the standard for what a usable answer looks like.",[758,7284,7285],{},"That does not come from a template. It comes from industry experience, customer understanding, product sense, taste, and decision-making reps. The best AI users are not simply writing longer prompts; they have clearer models in their heads. The prompt is where that business judgment becomes an interface.",[758,7287,7288],{},[3036,7289],{"alt":7290,"src":7291},"Prompt as an interface to business judgment","/blogs-img/2026-06-08-ai-stars-01.webp",[753,7293,7295],{"id":7294},"sales-works-the-same-way","Sales works the same way",[758,7297,7298],{},"You can see the same pattern in sales. Some people meet a customer once and the opportunity moves; others prepare more material, run more meetings, and still cannot move the deal. The difference is not effort. It is the ability to read the decision chain, incentives, risks, and real blocker.",[758,7300,7301],{},"On the surface, the difference is one meeting. In reality, it is business judgment. AI has moved that same gap into the prompt box: for one person, the question is just a question; for another, it already carries direction, tradeoffs, and standards.",[758,7303,7304],{},[3036,7305],{"alt":7306,"src":7307},"Sales judgment and customer understanding","/blogs-img/2026-06-08-ai-stars-02.webp",[753,7309,7311],{"id":7310},"ai-will-reprice-human-capabilities","AI will reprice human capabilities",[758,7313,7314],{},"For a long time, execution skills carried a lot of value by themselves. Writing copy, drafting plans, coding, analysis, research, documentation: these all took time and specialized training. With AI, execution gets cheaper, and whatever gets cheaper stops being the core moat.",[758,7316,7317],{},"The valuable layer moves up: problem definition, standards, business context, tradeoff judgment, taste, and understanding people. AI is not only an equalizer. It is leverage, and leverage does not make every operator equally strong. It magnifies the operator.",[758,7319,7320],{},"Someone with judgment will test faster, iterate faster, and learn faster. Someone without judgment may also produce more, but often more things that only look finished. That is one of the central shifts of the AI era: output becomes abundant, and judgment becomes scarce.",[758,7322,7323],{},[3036,7324],{"alt":7325,"src":7326},"Strategic altitude in the age of AI","/blogs-img/2026-06-08-ai-stars-03.webp",[753,7328,7330],{"id":7329},"looking-up-at-the-stars-is-strategy","Looking up at the stars is strategy",[758,7332,7333],{},"So when I say, in the age of AI, look up at the stars, I do not mean it romantically. I mean horizon and altitude. The faster execution becomes, the more valuable direction becomes. The cheaper content generation becomes, the more valuable point of view becomes. The easier information becomes to access, the more valuable judgment becomes.",[758,7335,7336],{},"For an entrepreneur, the real question is not only: how do I use AI better? That matters, but the deeper questions are: do I have the altitude to judge what is worth doing? Do I understand the market deeply enough to see where the opportunity is? Do I understand customers well enough to separate real demand from noise?",[758,7338,7339],{},"AI can amplify execution, but it cannot replace direction. Direction still comes from people, from accumulated cognitive assets, and from the ongoing judgment of markets, customers, and opportunities.",[753,7341,7343],{"id":7342},"my-conclusion","My conclusion",[758,7345,7346],{},"My investment in AI now has two layers. The first layer is tools and workflows; I want more speed. The second layer is cognitive assets: market understanding, business insight, customer sense, product judgment, taste, and historical context.",[758,7348,7349],{},"The first layer determines how fast I can run. The second layer determines where I am running, and how much AI can actually amplify me. The biggest risk in the AI era is not being bad at tools; it is mistaking the fluency created by tools for your own capability.",[758,7351,7352],{},"Same model. Same blank box. The difference is not who asks more cleverly. The difference is who has a larger world behind the question.",{"title":839,"searchDepth":708,"depth":708,"links":7354},[7355,7356,7357,7358,7359],{"id":7278,"depth":708,"text":7279},{"id":7294,"depth":708,"text":7295},{"id":7310,"depth":708,"text":7311},{"id":7329,"depth":708,"text":7330},{"id":7342,"depth":708,"text":7343},"In the age of AI, looking up at the stars means strategic altitude: the judgment, customer context, and direction that make AI useful.",{"date":7362,"image":7363,"alt":7364,"tags":7365,"category":3417,"youtube":7369,"published":859},"8th Jun 2026","/blogs-img/2026-06-08-ai-stars-cover.webp","A retro editorial illustration of a compass, market map, and AI prompt box pointing toward a strategic horizon",[30,7366,7367,7368],"entrepreneurship","judgment","strategy","https://youtu.be/6-X3NwICb7g","/blogs/en/in-the-age-of-ai-look-up-at-the-stars",{"title":7267,"description":7360},"blogs/en/24.in-the-age-of-ai-look-up-at-the-stars","x72AWiB7J3r7pAkBNqUHjg_lqFm2xuc20MaoJoYY880",{"id":7375,"title":7376,"body":7377,"description":7712,"extension":2126,"meta":7713,"navigation":859,"ogImage":7715,"path":7722,"seo":7723,"stem":7724,"__hash__":7725},"en/blogs/en/25.fable-5-managing-ai-autonomy.md","Fable 5 Changed the Unit of AI Work",{"type":750,"value":7378,"toc":7703},[7379,7382,7385,7388,7391,7394,7408,7411,7414,7418,7421,7424,7427,7430,7433,7439,7443,7452,7455,7458,7461,7464,7470,7474,7482,7485,7488,7491,7494,7497,7500,7506,7510,7513,7516,7519,7522,7525,7528,7531,7534,7538,7541,7556,7559,7562,7565,7585,7588,7591,7602,7605,7611,7615,7618,7621,7624,7630,7636,7642,7648,7654,7660,7666,7672,7675,7678,7681,7684,7688,7691,7694,7697,7700],[758,7380,7381],{},"I gave Claude Fable 5 a multi-hour task, and for the first time the interaction did not feel like prompting a model.",[758,7383,7384],{},"It felt like launching a work run.",[758,7386,7387],{},"The difference is subtle until you experience it. In the normal chat loop, I am inside the work. I ask a question, inspect the answer, correct it, add more context, ask again, and keep steering every few minutes. With Fable, the shape changed. I gave it an objective, it planned, delegated, researched, synthesized, and came back with polished pages. The visible result was the page. The more important result was the run itself.",[758,7389,7390],{},"That is the part I think matters.",[758,7392,7393],{},"Fable 5 did not only produce better output. It changed the unit of AI work from a response to a run.",[758,7395,7396,7401,7402,7407],{},[2177,7397,7400],{"href":7398,"rel":7399},"https://www.anthropic.com/news/claude-fable-5-mythos-5",[2181],"Anthropic launched Fable 5 and Mythos 5 on June 9, 2026",", describing models that can work autonomously for longer than previous Claude models. Three days later, Anthropic ",[2177,7403,7406],{"href":7404,"rel":7405},"https://www.anthropic.com/news/fable-mythos-access",[2181],"said it had received a US government export-control directive"," to suspend Fable 5 and Mythos 5 access for foreign nationals, with the practical effect of disabling access for all customers while it works to restore availability.",[758,7409,7410],{},"The launch and suspension made the same point from different directions. The launch showed what happens when agents can run longer, delegate more reliably, and use tools across larger tasks. The suspension showed that once agents become capable enough, control, governance, and boundaries stop being secondary concerns.",[758,7412,7413],{},"The lesson is not simply that Fable was powerful. The lesson is that when an agent can run for hours, the scarce skill is no longer writing the cleverest prompt. It is designing the operating contract that makes the run bounded, inspectable, and reversible.",[753,7415,7417],{"id":7416},"the-unit-changed-from-response-to-run","The unit changed from response to run",[758,7419,7420],{},"A response is small. It has text, maybe a tool call, maybe a file edit. If it is wrong, I correct it. If it is incomplete, I ask again. The cost of a mistake is usually one turn.",[758,7422,7423],{},"A run is different. A run has duration. It has state. It consumes tokens, calls tools, reads files, writes artifacts, dispatches subagents, makes intermediate assumptions, and may travel a long way before I inspect the result. The cost of a bad assumption is no longer one bad answer. It can become an entire wrong branch of work.",[758,7425,7426],{},"That is why longer autonomy changes the economics of attention. In prompt mode, my attention is the control loop. In run mode, my attention moves before and after the run. Before the run, I define the objective, context, constraints, permissions, checkpoints, and stop conditions. After the run, I inspect evidence and decide whether the work is usable.",[758,7428,7429],{},"The model may be smarter, but the work system becomes less forgiving. If the objective is vague, autonomy amplifies vagueness. If the constraints are unclear, autonomy explores the wrong space. If the agent can take actions without a boundary, autonomy turns capability into operational risk.",[758,7431,7432],{},"This is why \"managing agents\" is too soft a phrase. The work is not just managing. It is run design.",[758,7434,7435],{},[3036,7436],{"alt":7437,"src":7438},"Response versus run as the new unit of AI work","/blogs-img/2026-06-15-fable-01.webp",[753,7440,7442],{"id":7441},"fable-made-the-control-surface-visible","Fable made the control surface visible",[758,7444,7445,7446,7451],{},"Anthropic's own ",[2177,7447,7450],{"href":7448,"rel":7449},"https://platform.claude.com/docs/en/build-with-claude/prompt-engineering/prompting-claude-fable-5",[2181],"Fable prompting documentation"," is revealing because it is not only about better prompts. It talks about long runs, effort levels, progress claims, explicit boundaries, parallel subagents, memory systems, and scaffolding changes. That is a different category of guidance.",[758,7453,7454],{},"If a model can run for hours, the interface cannot just be a text box. The interface needs timeout behavior, progress indicators, asynchronous check-ins, evidence-backed status reports, and a way to define when the model should pause. If it can dispatch subagents, the harness needs to decide when delegation is useful, how subagents communicate, and how their work is reviewed. If it can maintain memory, the system needs a place to record lessons without polluting the work with bad assumptions.",[758,7456,7457],{},"In other words, the model improvement forces a product and workflow redesign around it.",[758,7459,7460],{},"This is the pattern I felt in the Fable run. I was not just looking at \"what did the model answer?\" I was implicitly asking a different set of questions: What did it decide to research? What did it delegate? Which assumptions did it carry forward? Where did it spend effort? What evidence supports the result? Where would I intervene if the run went off track?",[758,7462,7463],{},"Those questions are not prompt questions. They are operating-system questions.",[758,7465,7466],{},[3036,7467],{"alt":7468,"src":7469},"Run design as an operating contract around the agent","/blogs-img/2026-06-15-fable-03.webp",[753,7471,7473],{"id":7472},"superpowers-is-a-practical-rehearsal","Superpowers is a practical rehearsal",[758,7475,7476,7477,2480],{},"This is why my experience with Fable connected immediately to ",[2177,7478,7481],{"href":7479,"rel":7480},"https://github.com/obra/superpowers",[2181],"Superpowers",[758,7483,7484],{},"Superpowers is not another model. It is an open-source skills framework and software development methodology for coding agents. Its value is not that it makes the model magically smarter. Its value is that it wraps the model in a more disciplined way of working.",[758,7486,7487],{},"The pattern is simple: clarify the goal, tease out a spec, turn the spec into an implementation plan, execute through subagents or structured phases, review the work, and verify before claiming completion. The skills become a reusable operating contract for the agent.",[758,7489,7490],{},"That changes my daily work with Opus 4.8 and Codex 5.5. I can give the agent a meaningful objective and let it work for one or two hours, because the work is not just free-form generation. It has a process. It reads the codebase, writes a plan, executes against that plan, checks its own work, and comes back with artifacts I can review.",[758,7492,7493],{},"The model is the engine. The skills are the operating discipline around the engine.",[758,7495,7496],{},"This is a useful distinction because it prevents the wrong conclusion. The conclusion is not \"Fable is the future, so wait for better models.\" The conclusion is that better models make the surrounding work system more important. Superpowers is valuable because it trains the habit before every model can run for half a day. It makes autonomy legible.",[758,7498,7499],{},"That is also why the word \"skill\" is more important than it looks. A good skill is not a prompt template. It is a reusable piece of operating judgment: when to clarify, how to plan, what verification means, when to use subagents, when to ask for review, and what evidence is required before saying the work is done.",[758,7501,7502],{},[3036,7503],{"alt":7504,"src":7505},"Superpowers wraps the model in an operating discipline","/blogs-img/2026-06-15-fable-04.webp",[753,7507,7509],{"id":7508},"this-is-not-just-project-management","This is not just project management",[758,7511,7512],{},"The obvious objection is that all of this sounds like project management with AI branding.",[758,7514,7515],{},"Partly, yes. That is the point. As AI agents become more capable, the work starts to look less like chatting and more like delegation. Delegation has always required goals, context, review, and accountability.",[758,7517,7518],{},"But agent delegation has three differences that make the old habits insufficient.",[758,7520,7521],{},"First, agents move at machine speed. A human teammate who misunderstands a task usually creates friction quickly: a question, a delay, a visible mismatch. An agent can silently spend a large amount of compute on the wrong interpretation and return a polished artifact that hides the wrong path.",[758,7523,7524],{},"Second, agents operate through tools. They do not only think; they read, write, call APIs, modify files, send messages, and eventually may spend money or publish work. That moves risk from \"bad answer\" to \"bad action.\"",[758,7526,7527],{},"Third, agents are probabilistic workers without native accountability. A human can own a decision socially and organizationally. An agent can produce work, but the accountability remains with the person or team that designed the run.",[758,7529,7530],{},"That is why this is not just normal management. It is management with a new kind of worker: fast, tireless, useful, inconsistent, tool-using, and not actually responsible.",[758,7532,7533],{},"The operating contract is what makes that worker productive.",[753,7535,7537],{"id":7536},"governance-is-the-same-problem-at-larger-scale","Governance is the same problem at larger scale",[758,7539,7540],{},"The Fable suspension made this visible at the geopolitical level, but the pattern is the same at every scale.",[758,7542,7543,7544,7549,7550,7555],{},"Anthropic's statement talks about safeguards, red-teaming, jailbreak resistance, monitoring, customer data retention, and defense in depth. Databricks framed Fable 5 access around ",[2177,7545,7548],{"href":7546,"rel":7547},"https://www.databricks.com/blog/claude-fable-5-now-available-databricks-fully-governed-through-unity-ai-gateway",[2181],"governance, audit logs, tool-call policies, and cost controls",". Gartner ",[2177,7551,7554],{"href":7552,"rel":7553},"https://www.gartner.com/en/newsroom/press-releases/2026-05-26-gartner-says-applying-uniform-governance-across-ai-agents-will-lead-to-enterprise-ai-agent-failure",[2181],"warned in May 2026"," that applying uniform governance to all AI agents can fail because organizations must distinguish an agent's autonomy level from the scope of access it has been granted.",[758,7557,7558],{},"That distinction is exactly right. The dangerous combination is not intelligence by itself. It is autonomy plus access without a proportional control system.",[758,7560,7561],{},"An agent that only observes has one risk profile. An agent that drafts recommendations has another. An agent that can act with approval has another. An agent that can act autonomously across systems is a different category entirely.",[758,7563,7564],{},"This maps directly to individual builders. Before I let an agent run, I need to know what kind of autonomy I am granting:",[804,7566,7567,7570,7573,7576,7579,7582],{},[775,7568,7569],{},"Can it only read, summarize, and propose?",[775,7571,7572],{},"Can it edit files?",[775,7574,7575],{},"Can it install dependencies?",[775,7577,7578],{},"Can it call external services?",[775,7580,7581],{},"Can it create branches, push code, or publish content?",[775,7583,7584],{},"Can it continue when uncertain, or should it stop at specific boundaries?",[758,7586,7587],{},"These questions are not bureaucratic. They are the control surface of the work.",[758,7589,7590],{},"For every meaningful agent run, I now think there are three practical questions:",[772,7592,7593,7596,7599],{},[775,7594,7595],{},"What can it do?",[775,7597,7598],{},"How will I know what it did?",[775,7600,7601],{},"How do I stop, reverse, or repair it if it goes wrong?",[758,7603,7604],{},"If I cannot answer those, I have not designed the run. I have only launched it.",[758,7606,7607],{},[3036,7608],{"alt":7609,"src":7610},"Agent governance as autonomy plus access","/blogs-img/2026-06-15-fable-07.webp",[753,7612,7614],{"id":7613},"the-real-skill-is-designing-operating-contracts","The real skill is designing operating contracts",[758,7616,7617],{},"This is where I think AI-native work is going.",[758,7619,7620],{},"Prompting still matters. A prompt is an interface to judgment. But a prompt is not enough when the unit of work becomes a run. A run needs an operating contract.",[758,7622,7623],{},"The contract does not need to be complicated. For my own work, I think about seven parts:",[758,7625,7626,7629],{},[778,7627,7628],{},"Objective:"," What outcome matters? What would count as success?",[758,7631,7632,7635],{},[778,7633,7634],{},"Context:"," Which files, facts, docs, rules, or examples are authoritative?",[758,7637,7638,7641],{},[778,7639,7640],{},"Authority:"," What can the agent read, write, call, spend, or publish?",[758,7643,7644,7647],{},[778,7645,7646],{},"Checkpoints:"," Where should it pause, report, or ask for approval?",[758,7649,7650,7653],{},[778,7651,7652],{},"Evidence:"," What proof must support progress claims and final claims?",[758,7655,7656,7659],{},[778,7657,7658],{},"Budget:"," How much time, cost, and exploration is allowed?",[758,7661,7662,7665],{},[778,7663,7664],{},"Rollback:"," What happens if the run produces the wrong result?",[758,7667,7668],{},[3036,7669],{"alt":7670,"src":7671},"The seven layers of an operating contract for agent runs","/blogs-img/2026-06-15-fable-08.webp",[758,7673,7674],{},"This is a more useful frame than \"prompt engineering.\" It is closer to designing a small factory line. Some work can be handed over and run in the background. Some work requires my direct judgment. The leverage comes from knowing the difference.",[758,7676,7677],{},"That has been the larger shift in my own work. AI makes me feel more powerful because more projects can move at the same time. I can have one agent drafting, another researching, another coding, another turning a plan into distribution assets. But my focus did not become infinite. If anything, the value of focus increased.",[758,7679,7680],{},"The bottleneck moved from doing every task to routing the work correctly.",[758,7682,7683],{},"The best operator in this environment is not the person who watches every step. It is the person who designs the work so that the right parts can move without them, and the important parts still come back for judgment.",[753,7685,7687],{"id":7686},"the-durable-lesson","The durable lesson",[758,7689,7690],{},"Fable 5 may come back. It may change. Another model may take its place. That product cycle is not the durable lesson.",[758,7692,7693],{},"The durable lesson is that AI work is moving from prompt-response loops to longer, delegated, inspectable runs. Once that happens, the edge moves from access to orchestration.",[758,7695,7696],{},"The strongest model will still matter. But the strongest model inside a weak operating system will produce expensive chaos. A slightly weaker model inside a disciplined workflow may produce shippable work.",[758,7698,7699],{},"That is the part I want to keep practicing: not just how to ask for output, but how to design the conditions under which an agent can do real work.",[758,7701,7702],{},"The future does not belong to the best prompter. It belongs to the builder who can let execution run ahead without letting judgment fall behind.",{"title":839,"searchDepth":708,"depth":708,"links":7704},[7705,7706,7707,7708,7709,7710,7711],{"id":7416,"depth":708,"text":7417},{"id":7441,"depth":708,"text":7442},{"id":7472,"depth":708,"text":7473},{"id":7508,"depth":708,"text":7509},{"id":7536,"depth":708,"text":7537},{"id":7613,"depth":708,"text":7614},{"id":7686,"depth":708,"text":7687},"Fable 5 showed that AI work is shifting from prompt-response loops to long-running agent runs, making operating contracts the next scarce AI-native skill.",{"date":7714,"image":7715,"alt":7716,"tags":7717,"category":2132,"youtube":7721,"published":859},"15th Jun 2026","/blogs-img/2026-06-15-fable-cover.webp","A human operator supervising multiple AI work streams from a control room",[30,7718,7719,7720],"agents","workflow","ai-governance","https://youtu.be/jPHR-73HJa8","/blogs/en/fable-5-managing-ai-autonomy",{"title":7376,"description":7712},"blogs/en/25.fable-5-managing-ai-autonomy","YCXsS_ktdAsIdbDPNp2mfpyZUeaFiZVG-lXb5cgNmR4",{"id":7727,"title":7728,"body":7729,"description":7959,"extension":2126,"meta":7960,"navigation":859,"ogImage":7962,"path":7966,"seo":7967,"stem":7968,"__hash__":7969},"en/blogs/en/7.build-blog-in-one-day.md","Rapidly Building My Personal Blog",{"type":750,"value":7730,"toc":7951},[7731,7735,7741,7744,7746,7750,7753,7775,7778,7792,7798,7800,7804,7821,7837,7844,7846,7850,7853,7875,7878,7883,7886,7888,7892,7898,7905,7911,7913,7917,7920,7926,7932,7934,7940],[753,7732,7734],{"id":7733},"thoughts","Thoughts",[758,7736,7737,7738,2480],{},"Recently, I finally completed something I had been wanting to do for a long time—but kept putting off: ",[778,7739,7740],{},"building my own blog",[758,7742,7743],{},"I’ve always wanted a space of my own where I could share life experiences, reflect on my work, and document what I’ve learned. Social media is too fast, too fragmented. A blog feels more like a digital notebook—calm, self-owned, and free. Writing here isn’t about chasing traffic; it’s about organizing my thoughts, preserving value, and leaving a trail that my future self can look back on.",[2163,7745],{},[753,7747,7749],{"id":7748},"my-experience-with-ai-tools","My Experience with AI Tools",[758,7751,7752],{},"Like many others, I was initially excited about the idea of using AI tools to build websites. I tried several popular platforms:",[804,7754,7755,7765],{},[775,7756,7757,7764],{},[2177,7758,7761],{"href":7759,"rel":7760},"https://lovable.dev",[2181],[778,7762,7763],{},"Lovable.dev",": Friendly UI, low learning curve, and you can spin up a decent-looking page in minutes.",[775,7766,7767,7774],{},[2177,7768,7771],{"href":7769,"rel":7770},"https://v0.dev",[2181],[778,7772,7773],{},"V0.dev",": Great for quickly generating structured component code, especially when paired with shadcn/ui—it feels very intuitive.",[758,7776,7777],{},"However, I ran into two main issues:",[772,7779,7780,7786],{},[775,7781,7782,7785],{},[778,7783,7784],{},"Lack of control over details",": The first draft comes out fast, but if I wanted to tweak layout, adjust typography, or add multi-language support, the process became frustrating. It was often easier to just write the code myself.",[775,7787,7788,7791],{},[778,7789,7790],{},"Free plan limitations",": I was on the free tier, which meant the code wasn’t persistently saved and I couldn’t iterate freely. This seriously impacted continuity.",[758,7793,7794,7795],{},"So in the end, I decided: ",[778,7796,7797],{},"AI can assist my workflow, but the core architecture should be in my hands.",[2163,7799],{},[753,7801,7803],{"id":7802},"my-final-setup-nuxt-content-module","My Final Setup: Nuxt + Content Module",[758,7805,7806,7807,7814,7815,7820],{},"After comparing several options, I chose to use ",[2177,7808,7811],{"href":7809,"rel":7810},"https://nuxt.com",[2181],[778,7812,7813],{},"Nuxt 3"," with the official ",[2177,7816,7819],{"href":7817,"rel":7818},"https://content.nuxtjs.org/",[2181],"@nuxt/content"," module to build my blog. Why?",[804,7822,7823,7826,7831,7834],{},[775,7824,7825],{},"Nuxt is a framework I’m already familiar with. It combines Vue’s flexibility with powerful SSR capabilities.",[775,7827,7828,7830],{},[841,7829,7819],{}," allows me to write posts in Markdown, use frontmatter metadata, auto-generate TOCs, and get great SEO support out of the box.",[775,7832,7833],{},"It supports static deployment, making it easy to publish on Vercel or Netlify.",[775,7835,7836],{},"And in the future, if I want to add search, a recommendation system, multi-language support, or even embed AI-assisted reading—this stack is easy to extend.",[758,7838,7839,7840,7843],{},"Most importantly, this solution gives me ",[778,7841,7842],{},"full control over code and styles",", without being boxed in by a platform or a locked template.",[2163,7845],{},[753,7847,7849],{"id":7848},"my-ai-workflow-vibe-coding-augment-code","My AI Workflow: Vibe Coding + Augment Code",[758,7851,7852],{},"Although I didn’t stick with AI site builders, I did integrate AI into my development flow.",[758,7854,7855,7856,7863,7864,7867,7868,7871,7872,2480],{},"I chose ",[2177,7857,7860],{"href":7858,"rel":7859},"https://www.augmentcode.com/",[2181],[778,7861,7862],{},"Augment Code",". It doesn’t generate full pages for me, but instead acts as a ",[778,7865,7866],{},"semantic coding assistant"," while I work.",[7869,7870],"br",{},"\nI call this process: ",[778,7873,7874],{},"vibe coding",[758,7876,7877],{},"It feels more like this:",[3590,7879,7880],{},[758,7881,7882],{},"“I know I want to adjust this card component’s layout, but I don’t want to handwrite every CSS rule. Can you try a few versions that feel right?”",[758,7884,7885],{},"In this context, AI isn’t some robotic code generator—it’s more like a design-savvy partner who helps explore options. This significantly improves how I control the visual output and makes the building process way more enjoyable.",[2163,7887],{},[753,7889,7891],{"id":7890},"tech-is-just-a-mediumexpression-is-the-core","Tech is Just a Medium—Expression is the Core",[758,7893,7894,7895],{},"Building a blog isn’t hard. ",[778,7896,7897],{},"What’s hard is sticking with it.",[758,7899,7900,7901,7904],{},"My biggest takeaway during setup was this: ",[778,7902,7903],{},"Don’t aim for a perfect launch—just aim to write",". Early on, there’s no need for fancy design or heavy structure. Just get words on the page.",[758,7906,7907,7908,7910],{},"To me, this blog isn’t just a content platform—it’s a long-term system for self-construction.",[7869,7909],{},"\nIt connects my reflections on life, questions I explore while learning, lessons from work—and maybe even seeds of future product ideas.",[2163,7912],{},[753,7914,7916],{"id":7915},"closing-thoughts","Closing Thoughts",[758,7918,7919],{},"From today on, this blog isn’t just a record of my past, but a commitment to my future.",[758,7921,7922,7923,7925],{},"Here, I’ll continue documenting my thoughts and growth—",[7869,7924],{},"\non work, on family, on tech, and on life.",[758,7927,7928,7929,7931],{},"Not fancy, just real.",[7869,7930],{},"\nMay it become a consistent part of how I express and evolve.",[2163,7933],{},[758,7935,7936,7937],{},"🚀 ",[778,7938,7939],{},"Coming Soon:",[804,7941,7942,7945,7948],{},[775,7943,7944],{},"Next post: how I use Nuxt + Tailwind for progressive enhancement on my blog homepage",[775,7946,7947],{},"Monthly summaries of writing insights + blog improvements",[775,7949,7950],{},"Currently exploring GPT-powered blog integration (reader Q&A + multilingual content recommendations)",{"title":839,"searchDepth":708,"depth":708,"links":7952},[7953,7954,7955,7956,7957,7958],{"id":7733,"depth":708,"text":7734},{"id":7748,"depth":708,"text":7749},{"id":7802,"depth":708,"text":7803},{"id":7848,"depth":708,"text":7849},{"id":7890,"depth":708,"text":7891},{"id":7915,"depth":708,"text":7916},"How I used open-source frameworks to quickly build my blog, and why I plan to turn it into a long-term space for growth and reflection.",{"date":7961,"image":7962,"alt":7963,"tags":7964,"category":2463,"topics":7965,"published":859},"18th Apr 2025","/blogs-img/blog6.jpg","Rapidly Building My Blog",[2130,2131],[2135,3182],"/blogs/en/build-blog-in-one-day",{"title":7728,"description":7959},"blogs/en/7.build-blog-in-one-day","HRS07uV65WJDmdWHv6onsmkw2agJh3aFeqeJESn4sss",{"id":7971,"title":7972,"body":7973,"description":8224,"extension":2126,"meta":8225,"navigation":859,"ogImage":8227,"path":8230,"seo":8231,"stem":8232,"__hash__":8233},"en/blogs/en/8.learn-graph-db-neo4j.md","Learn Graph DB - Neo4j",{"type":750,"value":7974,"toc":8206},[7975,7977,7980,7984,7987,7991,8017,8021,8024,8029,8046,8050,8067,8071,8088,8092,8115,8122,8126,8132,8136,8140,8143,8146,8150,8153,8156,8160,8163,8166,8170,8173,8176,8180,8183,8186,8190,8193,8196,8200,8203],[753,7976,756],{"id":755},[758,7978,7979],{},"I recently saw they are some discussions about Graph Rag, which seems a powerful way to build knowledge systems. In the meanwhile, during KubeConf 2025, I met the folks at the Neo4j's booth. It's interesting to see how to use Graph DB to solve my problems.",[753,7981,7983],{"id":7982},"a-brief-overview-of-database-types","A Brief Overview of Database Types",[758,7985,7986],{},"Before diving deeper into graph databases like Neo4j, it's helpful to understand the broader landscape of database technologies. Different types of databases are optimized for different kinds of data and access patterns. Here's a quick summary:",[826,7988,7990],{"id":7989},"_1-relational-databases-sql","1. Relational Databases (SQL)",[804,7992,7993,7999,8005,8011],{},[775,7994,7995,7998],{},[778,7996,7997],{},"Concept:"," Stores data in predefined tables with rows and columns. Relationships between tables are explicitly defined using foreign keys. Data integrity is often enforced through schemas and constraints.",[775,8000,8001,8004],{},[778,8002,8003],{},"Use Cases:"," Structured data, transactions requiring ACID (Atomicity, Consistency, Isolation, Durability) compliance, complex queries involving multiple tables (joins).",[775,8006,8007,8010],{},[778,8008,8009],{},"Query Language:"," SQL (Structured Query Language).",[775,8012,8013,8016],{},[778,8014,8015],{},"Representatives:"," PostgreSQL, MySQL, Oracle Database, SQL Server, SQLite.",[826,8018,8020],{"id":8019},"_2-nosql-databases","2. NoSQL Databases",[758,8022,8023],{},"This is a broad category encompassing several models that differ from the traditional relational approach. They often prioritize scalability, performance, and flexibility over the strict consistency of SQL databases.",[8025,8026,8028],"h4",{"id":8027},"a-key-value-stores","a) Key-Value Stores",[804,8030,8031,8036,8041],{},[775,8032,8033,8035],{},[778,8034,7997],{}," Simplest NoSQL type. Stores data as a collection of key-value pairs. Think of it like a giant dictionary or hash map.",[775,8037,8038,8040],{},[778,8039,8003],{}," Caching, session management, user profiles, real-time data lookups where access is primarily by key.",[775,8042,8043,8045],{},[778,8044,8015],{}," Redis, Memcached, Amazon DynamoDB (also has document features).",[8025,8047,8049],{"id":8048},"b-document-databases","b) Document Databases",[804,8051,8052,8057,8062],{},[775,8053,8054,8056],{},[778,8055,7997],{}," Stores data in documents, typically in formats like JSON or BSON. Documents can be nested and don't require a strict, predefined schema for all documents in a collection.",[775,8058,8059,8061],{},[778,8060,8003],{}," Content management, catalogs, user profiles, applications where data structure evolves frequently.",[775,8063,8064,8066],{},[778,8065,8015],{}," MongoDB, CosmosDB, Couchbase, Firestore.",[8025,8068,8070],{"id":8069},"c-column-family-stores","c) Column-Family Stores",[804,8072,8073,8078,8083],{},[775,8074,8075,8077],{},[778,8076,7997],{}," Stores data in tables with rows and dynamic columns. Data for a row is stored together by column families, which allows for efficient reads/writes of specific columns across many rows.",[775,8079,8080,8082],{},[778,8081,8003],{}," Big data applications, logging, data warehousing, systems requiring high write throughput.",[775,8084,8085,8087],{},[778,8086,8015],{}," Apache Cassandra, HBase.",[826,8089,8091],{"id":8090},"_3-graph-databases","3. Graph Databases",[804,8093,8094,8099,8104,8110],{},[775,8095,8096,8098],{},[778,8097,7997],{}," Stores data as nodes (entities) and edges (relationships) connecting those nodes. Both nodes and edges can have properties. Optimized for traversing and querying complex relationships.",[775,8100,8101,8103],{},[778,8102,8003],{}," Social networks, recommendation engines, fraud detection, knowledge graphs, network and IT operations.",[775,8105,8106,8109],{},[778,8107,8108],{},"Query Languages:"," Often specific graph query languages like Cypher (Neo4j) or Gremlin (Apache TinkerPop).",[775,8111,8112,8114],{},[778,8113,8015],{}," Neo4j, ArangoDB (multi-model), Amazon Neptune, JanusGraph.",[758,8116,8117,8118,8121],{},"Understanding these different types helps you choose the right tool for the job. Now, let's focus on the unique strengths of ",[778,8119,8120],{},"Graph Databases"," like Neo4j...",[753,8123,8125],{"id":8124},"neo4j-cypher-queries-examples","Neo4j Cypher queries examples:",[834,8127,8130],{"className":8128,"code":8129,"language":3579},[3577],"MATCH (n:PLAYER) RETURN n\nMATCH (n:PLAYER) WHERE n.name = \"James Harden\" RETURN n\nMATCH (n:PLAYER {name: \"James Harden\"}) RETURN n\nMATCH (n:PLAYER) WHERE n.name \u003C> \"James Harden\" RETURN n\n\n# Aggregation\nMATCH (player:PLAYER) - [gamePlayed:PLAYED_AGAINST] -> (team:TEAM) RETURN player.name, COUNT(gamePlayed)\n\nMATCH (x:PLAYER {name: 'xx'}) DETACH DELETE ja\n\nMATCH (n) DETACH DELETE n\n\nCREATE (newPlayer:PLAYER:COACH:GENERAL_MANAGER { name: \"Aaron Guo\", height: 1.70 }) RETURN newPlayer\n\n# Create relation\nMATCH (player:PLAYER {name: \"Aaron Guo\"}), (lakers:TEAM {name: \"LA Lakers\"} )\nCREATE (player) - [:PLAY_FOR {salary: 2000000}] -> (lakers)\n",[841,8131,8129],{"__ignoreMap":839},[753,8133,8135],{"id":8134},"here-are-some-common-and-powerful-use-cases-for-neo4j","Here are some common and powerful use cases for Neo4j:",[826,8137,8139],{"id":8138},"recommendation-engines","Recommendation Engines:",[758,8141,8142],{},"How it works: Neo4j can easily model relationships like \"User A BOUGHT Product X\", \"User A is FRIENDS_WITH User B\", \"User B LIKED Product Y\".",[758,8144,8145],{},"Use Case: By traversing these relationships, you can generate recommendations like \"Users who bought X also bought Y\" or \"Your friends liked Z\". It's much faster and more intuitive to query these connections in a graph than joining many tables in SQL.  \nFraud Detection:",[826,8147,8149],{"id":8148},"fraud-detection","Fraud Detection",[758,8151,8152],{},"How it works: Fraud often involves complex rings of seemingly unrelated individuals, accounts, devices, or transactions. Neo4j can map these connections (e.g., \"Account 1 USED_DEVICE A\", \"Account 2 USED_DEVICE A\", \"Account 1 SENT_MONEY_TO Account 3\").  ",[758,8154,8155],{},"Use Case: Identifying shared identifiers (like devices, IP addresses, physical addresses) or unusual transaction patterns that link potentially fraudulent accounts. Graph queries can quickly uncover these hidden rings that might be hard to spot in tabular data.  \nKnowledge Graphs:",[826,8157,8159],{"id":8158},"knowledge-graphs","Knowledge Graphs",[758,8161,8162],{},"How it works: Representing complex domains by connecting entities (people, places, organizations, concepts, products) with various relationships.",[758,8164,8165],{},"Use Case: Building internal enterprise knowledge bases, enhancing search engines (like Google's Knowledge Graph), managing metadata, or creating semantic applications where understanding the context and links between information is key.  \nNetwork and IT Operations:",[826,8167,8169],{"id":8168},"network-and-it-operations","Network and IT Operations",[758,8171,8172],{},"How it works: Modeling dependencies between servers, applications, services, routers, data centers, and virtual machines.  ",[758,8174,8175],{},"Use Case: Visualizing network topology, performing impact analysis (\"If this server goes down, what services are affected?\"), identifying single points of failure, and speeding up root cause analysis during outages.  \nIdentity and Access Management (IAM):",[826,8177,8179],{"id":8178},"identity-and-access-management-iam","Identity and Access Management (IAM)",[758,8181,8182],{},"How it works: Modeling users, groups, roles, permissions, and resources, along with the relationships between them (e.g., \"User X BELONGS_TO Group Y\", \"Group Y HAS_PERMISSION Z\", \"Permission Z APPLIES_TO Resource W\").",[758,8184,8185],{},"Use Case: Quickly querying complex access rights (\"Who has access to this sensitive resource?\", \"What access does this user have?\", \"Why does this user have access?\"), managing hierarchies, and enforcing security policies.  \nSupply Chain Management & Logistics:",[826,8187,8189],{"id":8188},"supply-chain-management-logistics","Supply Chain Management & Logistics",[758,8191,8192],{},"How it works: Mapping suppliers, manufacturers, distributors, warehouses, products, and shipments as nodes, and their interactions as relationships.  ",[758,8194,8195],{},"Use Case: Tracking goods through the supply chain, identifying bottlenecks, optimizing routes, understanding dependencies, and assessing risks (e.g., \"If Supplier A has an issue, which products are affected?\").  \nSocial Networking:",[826,8197,8199],{"id":8198},"social-networking","Social Networking",[758,8201,8202],{},"How it works: The classic example – modeling users and their connections (friends, follows, blocks), group memberships, posts, likes, shares, etc.  ",[758,8204,8205],{},"Use Case: Finding friends-of-friends, suggesting connections, analyzing influence within the network, and powering social features efficiently.  ",{"title":839,"searchDepth":708,"depth":708,"links":8207},[8208,8209,8214,8215],{"id":755,"depth":708,"text":756},{"id":7982,"depth":708,"text":7983,"children":8210},[8211,8212,8213],{"id":7989,"depth":110,"text":7990},{"id":8019,"depth":110,"text":8020},{"id":8090,"depth":110,"text":8091},{"id":8124,"depth":708,"text":8125},{"id":8134,"depth":708,"text":8135,"children":8216},[8217,8218,8219,8220,8221,8222,8223],{"id":8138,"depth":110,"text":8139},{"id":8148,"depth":110,"text":8149},{"id":8158,"depth":110,"text":8159},{"id":8168,"depth":110,"text":8169},{"id":8178,"depth":110,"text":8179},{"id":8188,"depth":110,"text":8189},{"id":8198,"depth":110,"text":8199},"Recently read somethings about Graph Rag, started to interested into Graph DB, learn and see how to use it in my work.",{"date":8226,"image":8227,"alt":7972,"tags":8228,"category":2463,"topics":8229,"published":859,"featured":859},"5th Apr 2025","/blogs-img/8.learn-graph-db-neo4j.jpg",[2130,2131],[2134,2135],"/blogs/en/learn-graph-db-neo4j",{"title":7972,"description":8224},"blogs/en/8.learn-graph-db-neo4j","uTU4lrmH5mt4blsID3JMfaTPpc8eyeIpAX5HyrtS57o",{"id":8235,"title":8236,"body":8237,"description":8397,"extension":2126,"meta":8398,"navigation":859,"ogImage":8400,"path":8404,"seo":8405,"stem":8406,"__hash__":8407},"en/blogs/en/9.a-new-chapter.md","A New Chapter",{"type":750,"value":8238,"toc":8392},[8239,8242,8248,8251,8254,8257,8264,8270,8273,8280,8291,8295,8302,8305,8343,8350,8354,8357,8374,8377,8383,8386,8389],[753,8240,8236],{"id":8241},"a-new-chapter",[758,8243,8244,8245,5570],{},"Lately, what I’ve been thinking about the most is this: ",[778,8246,8247],{},"how can I bring about a real, qualitative change in my life",[758,8249,8250],{},"Over the past few years, I’ve thrown 200% of my energy into work. Every day felt like a battle, constantly pushing forward. But I’ve come to realize that this kind of single-threaded life simply isn’t sustainable. I’ve neglected my responsibilities at home, missed out on valuable time with loved ones, ignored my own health, and slowly started to operate on autopilot—becoming dull, rigid, and emotionally drained. This isn’t the life I want.",[758,8252,8253],{},"Not long ago, my family went through a difficult period, which forced me to face a tough question: as a son, husband, and father, how well have I really fulfilled my role?",[758,8255,8256],{},"Managing the relationship between my parents and my wife is far from a logical problem with a clear solution. It’s a complex emotional art that requires understanding, patience, empathy, and ongoing communication. During this time, I’ve had to dedicate more mental and emotional energy to navigating this web of relationships. And gradually, I realized: work may actually be the easiest part of my life to control—while family and close relationships are where I still have the most growing to do.",[758,8258,8259,8260,8263],{},"The truth is, as long as you’re reasonably smart and driven, career success is often just a matter of effort. But life isn’t just about your job. What’s truly worth pursuing is not individual achievement, but ",[778,8261,8262],{},"systematic balance","—where family, health, career, and personal growth all move forward in harmony.",[758,8265,8266,8267],{},"This reminded me of raising kids. We all know that parenting isn’t just about pouring in time, nor is there one-size-fits-all advice. It requires presence, observation, listening, and guidance—just like crafting a great product. You need breadth of knowledge, depth of vision, and consistency of execution. But most of all, you need to approach it wisely. ",[778,8268,8269],{},"With the right mindset, it’s possible to achieve compounding growth even with limited time.",[758,8271,8272],{},"And I’ve finally realized—I need to apply that same mindset to redesigning my own life.",[758,8274,8275,8276,8279],{},"From today onward, I want to approach every dimension of my life with a new perspective. In this age of AI, it’s no longer just a futuristic idea. It’s already woven into our lives. Instead of grinding along inefficient paths, I want to embrace AI as a powerful lever to ",[778,8277,8278],{},"reshape the way I live",":",[804,8281,8282,8285,8288],{},[775,8283,8284],{},"Using AI to help me manage my health more scientifically",[775,8286,8287],{},"Using AI to tackle complex work tasks more efficiently",[775,8289,8290],{},"Even using AI to improve communication, relationships, and emotional regulation",[753,8292,8294],{"id":8293},"why-ai","Why AI?",[758,8296,8297,8298,8301],{},"As AI continues to evolve, I’ve realized I can ",[778,8299,8300],{},"live smarter, not harder",". We’re living in an extraordinary era—AI is no longer just for programmers; it’s becoming an everyday companion in our lives.",[758,8303,8304],{},"And the data backs it up:",[804,8306,8307,8314,8321,8332],{},[775,8308,8309,8310,8313],{},"A McKinsey study shows that generative AI could unlock ",[778,8311,8312],{},"$2.6 to $4.4 trillion USD"," in productivity gains annually—especially for knowledge workers, by automating repetitive and information-heavy tasks.",[775,8315,8316,8317,8320],{},"A Harvard study found that employees using GPT-4 in complex writing and data analysis tasks saw a ",[778,8318,8319],{},"40% increase in accuracy",", halved completion time, and reported significantly higher satisfaction.",[775,8322,8323,8324,8327,8328,8331],{},"According to a survey by Jobscan, ",[778,8325,8326],{},"44% of users"," used AI to optimize resumes and prepare for interviews, resulting in an ",[778,8329,8330],{},"average 14% salary increase"," and better job matches.",[775,8333,8334,8335,8338,8339,8342],{},"In Japan, an AI emotional coach called ",[2156,8336,8337],{},"MindEase"," reduced ",[778,8340,8341],{},"anxiety by 31%"," among high-stress professionals by guiding daily self-awareness and stress management practices.",[758,8344,8345,8346,8349],{},"I believe that ",[778,8347,8348],{},"if I intentionally redesign my life with AI and optimize every detail",", I might just be able to: preserve my health, maintain strong family relationships, and still create professional impact—perhaps even more effectively than before.",[753,8351,8353],{"id":8352},"my-ai-powered-life-experiment-begins-today","My “AI-powered Life Experiment” Begins Today",[758,8355,8356],{},"I’ve decided to give myself a real reset. With the help of AI, I’m going to redesign every key aspect of my life: work, family, health, emotions, cognition.",[804,8358,8359,8362,8365,8368,8371],{},[775,8360,8361],{},"Using AI to plan daily meals, rest, and workouts—so I’m not relying on sheer willpower anymore",[775,8363,8364],{},"Using AI to streamline meetings, writing, and knowledge gathering—so I can get complex tasks done with less time",[775,8366,8367],{},"Using AI to generate parenting ideas and activities—so I can connect with my child more meaningfully and easily",[775,8369,8370],{},"Using AI to track emotions, journal, and reflect—so I can stay sharp and self-aware",[775,8372,8373],{},"Using AI to help me design better systems—and occasionally, more humorous ways to work",[758,8375,8376],{},"My 38th birthday is just around the corner. I don’t want to just blow out candles or drink a little wine. This year, I want to give myself a truly meaningful gift: a full reboot of life.",[758,8378,8379,8380,2480],{},"May the road ahead be one of ",[778,8381,8382],{},"clarity, balance, and creativity",[758,8384,8385],{},"This blog post is the first step of my “AI-powered Life Redesign” project. I’ll use this space to document the changes I experience, the challenges I encounter, and the breakthroughs—both technical and emotional—that come along the way.",[758,8387,8388],{},"Very soon, I’ll turn 38. Instead of celebrating, I hope this day marks the true beginning of a new life phase. Not just a new way of working, but a new way of living.",[758,8390,8391],{},"May this new chapter be one where I live with more clarity, more balance, and more creative energy.",{"title":839,"searchDepth":708,"depth":708,"links":8393},[8394,8395,8396],{"id":8241,"depth":708,"text":8236},{"id":8293,"depth":708,"text":8294},{"id":8352,"depth":708,"text":8353},"How to redesign life with AI—finding balance among family, health, career, and personal growth",{"date":8399,"image":8400,"alt":8401,"tags":8402,"category":3016,"topics":8403,"published":859,"featured":859},"3rd May 2025","/blogs-img/blog5.jpg","A New Chapter - Redesigning life with AI",[2130,3015],[3182],"/blogs/en/a-new-chapter",{"title":8236,"description":8397},"blogs/en/9.a-new-chapter","SqkwRcCTmaMX_XsjxLUPR4YXKtzYle5L5Bb8S-XD3PQ",[8409,9619,9907,10438,10591,10816,11125,11335,11630,11848,12124,12361,12807,14089,14469,14569,14896,15113,15135],{"id":8410,"title":8411,"body":8412,"description":9610,"extension":2126,"meta":9611,"navigation":859,"ogImage":7962,"path":9615,"seo":9616,"stem":9617,"__hash__":9618},"zh/blogs/zh/10.openai-serp-api-web-search-ai-summary.md","构建股票分析的网络搜索+AI摘要工具",{"type":750,"value":8413,"toc":9592},[8414,8417,8420,8423,8427,8430,8456,8459,8473,8477,8480,8484,8487,8520,8524,8527,8638,8642,8645,8808,8812,8815,9023,9027,9030,9180,9184,9187,9201,9204,9485,9488,9491,9495,9498,9509,9513,9516,9527,9530,9543,9547,9550,9561,9564,9567,9570,9584,9587,9590],[753,8415,8416],{"id":8416},"引言",[758,8418,8419],{},"在当今快节奏的金融市场中，分析师需要快速处理大量信息以做出明智的决策。传统的手动搜索网络、阅读文章和综合信息的方法既耗时又容易错过关键见解。这就是为什么将AI与网络搜索功能相结合可以创建一个强大的股票分析工具。",[758,8421,8422],{},"在这篇博客文章中，我将分享如何为公司的实验性股票分析项目构建一个网络搜索+AI摘要工具。这个工具帮助分析师快速收集和综合他们正在研究的股票信息，节省了大量手动工作时间，并提供了更全面的见解。",[753,8424,8426],{"id":8425},"结合网络搜索与ai的强大力量","结合网络搜索与AI的强大力量",[758,8428,8429],{},"在深入实现之前，让我们了解为什么这种组合特别强大：",[772,8431,8432,8438,8444,8450],{},[775,8433,8434,8437],{},[778,8435,8436],{},"实时信息访问","：SERP（搜索引擎结果页面）API提供来自网络的最新信息",[775,8439,8440,8443],{},[778,8441,8442],{},"上下文理解","：像GPT-4这样的大型语言模型可以理解信息的上下文和相关性",[775,8445,8446,8449],{},[778,8447,8448],{},"综合能力","：AI可以总结、提取关键点，并识别多个来源的趋势",[775,8451,8452,8455],{},[778,8453,8454],{},"可定制分析","：系统可以定制为关注股票分析的特定方面（财务、新闻情绪、市场趋势）",[758,8457,8458],{},"对于股票分析而言，这种组合使分析师能够：",[804,8460,8461,8464,8467,8470],{},[775,8462,8463],{},"快速收集有关公司的最新新闻和发展",[775,8465,8466],{},"分析多个来源的市场情绪",[775,8468,8469],{},"识别可能隐藏在各种文章中的潜在风险或机会",[775,8471,8472],{},"在几秒钟而不是几小时内生成全面的研究摘要",[753,8474,8476],{"id":8475},"实现构建工具","实现：构建工具",[758,8478,8479],{},"让我们一步步地了解如何构建这个工具。",[826,8481,8483],{"id":8482},"_1-设置环境","1. 设置环境",[758,8485,8486],{},"首先，我们需要设置环境并安装必要的依赖项：",[834,8488,8490],{"className":836,"code":8489,"language":838,"meta":839,"style":839},"// 安装所需的包\nnpm install axios openai dotenv\n\n// 创建.env文件存储API密钥\nOPENAI_API_KEY=你的openai_api密钥\nSERP_API_KEY=你的serp_api密钥\n",[841,8491,8492,8497,8501,8505,8510,8515],{"__ignoreMap":839},[844,8493,8494],{"class":846,"line":4},[844,8495,8496],{},"// 安装所需的包\n",[844,8498,8499],{"class":846,"line":708},[844,8500,854],{},[844,8502,8503],{"class":846,"line":110},[844,8504,860],{"emptyLinePlaceholder":859},[844,8506,8507],{"class":846,"line":69},[844,8508,8509],{},"// 创建.env文件存储API密钥\n",[844,8511,8512],{"class":846,"line":132},[844,8513,8514],{},"OPENAI_API_KEY=你的openai_api密钥\n",[844,8516,8517],{"class":846,"line":78},[844,8518,8519],{},"SERP_API_KEY=你的serp_api密钥\n",[826,8521,8523],{"id":8522},"_2-配置serp-api","2. 配置SERP API",[758,8525,8526],{},"我们将使用SERP API获取搜索结果。有几个提供商可用，但对于这个项目，我使用了SerpAPI，它提供来自搜索引擎的结构化数据：",[834,8528,8530],{"className":836,"code":8529,"language":838,"meta":839,"style":839},"const axios = require('axios')\nrequire('dotenv').config()\n\nasync function searchWeb(query, numResults = 5) {\n  try {\n    const response = await axios.get('https://serpapi.com/search', {\n      params: {\n        q: query,\n        api_key: process.env.SERP_API_KEY,\n        num: numResults,\n      },\n    })\n\n    // 提取有机搜索结果\n    const searchResults = response.data.organic_results.map((result) => ({\n      title: result.title,\n      link: result.link,\n      snippet: result.snippet,\n    }))\n\n    return searchResults\n  } catch (error) {\n    console.error('网络搜索错误:', error)\n    throw error\n  }\n}\n",[841,8531,8532,8536,8540,8544,8548,8552,8556,8560,8564,8568,8572,8576,8580,8584,8589,8593,8597,8601,8605,8609,8613,8617,8621,8626,8630,8634],{"__ignoreMap":839},[844,8533,8534],{"class":846,"line":4},[844,8535,892],{},[844,8537,8538],{"class":846,"line":708},[844,8539,897],{},[844,8541,8542],{"class":846,"line":110},[844,8543,860],{"emptyLinePlaceholder":859},[844,8545,8546],{"class":846,"line":69},[844,8547,906],{},[844,8549,8550],{"class":846,"line":132},[844,8551,911],{},[844,8553,8554],{"class":846,"line":78},[844,8555,916],{},[844,8557,8558],{"class":846,"line":162},[844,8559,921],{},[844,8561,8562],{"class":846,"line":66},[844,8563,926],{},[844,8565,8566],{"class":846,"line":63},[844,8567,931],{},[844,8569,8570],{"class":846,"line":91},[844,8571,936],{},[844,8573,8574],{"class":846,"line":939},[844,8575,942],{},[844,8577,8578],{"class":846,"line":945},[844,8579,948],{},[844,8581,8582],{"class":846,"line":388},[844,8583,860],{"emptyLinePlaceholder":859},[844,8585,8586],{"class":846,"line":204},[844,8587,8588],{},"    // 提取有机搜索结果\n",[844,8590,8591],{"class":846,"line":960},[844,8592,963],{},[844,8594,8595],{"class":846,"line":60},[844,8596,968],{},[844,8598,8599],{"class":846,"line":312},[844,8600,973],{},[844,8602,8603],{"class":846,"line":207},[844,8604,978],{},[844,8606,8607],{"class":846,"line":337},[844,8608,983],{},[844,8610,8611],{"class":846,"line":12},[844,8612,860],{"emptyLinePlaceholder":859},[844,8614,8615],{"class":846,"line":180},[844,8616,992],{},[844,8618,8619],{"class":846,"line":174},[844,8620,997],{},[844,8622,8623],{"class":846,"line":304},[844,8624,8625],{},"    console.error('网络搜索错误:', error)\n",[844,8627,8628],{"class":846,"line":155},[844,8629,1007],{},[844,8631,8632],{"class":846,"line":21},[844,8633,1012],{},[844,8635,8636],{"class":846,"line":23},[844,8637,1017],{},[826,8639,8641],{"id":8640},"_3-从搜索结果中获取内容","3. 从搜索结果中获取内容",[758,8643,8644],{},"一旦我们有了搜索结果，我们需要从网页中获取实际内容：",[834,8646,8648],{"className":836,"code":8647,"language":838,"meta":839,"style":839},"const axios = require('axios')\nconst cheerio = require('cheerio')\n\nasync function fetchContent(url) {\n  try {\n    const response = await axios.get(url)\n    const $ = cheerio.load(response.data)\n\n    // 移除脚本标签、样式标签和其他非内容元素\n    $('script, style, meta, link').remove()\n\n    // 提取主要内容（这是一种简化的方法）\n    // 对于生产环境，你可能想使用更复杂的内容提取方法\n    const content = $('body').text().replace(/\\s+/g, ' ').trim()\n\n    return content\n  } catch (error) {\n    console.error(`从${url}获取内容时出错:`, error)\n    return '' // 如果无法获取内容，则返回空字符串\n  }\n}\n\nasync function fetchContentsFromSearchResults(searchResults) {\n  const contents = []\n\n  for (const result of searchResults) {\n    const content = await fetchContent(result.link)\n    if (content) {\n      contents.push({\n        title: result.title,\n        url: result.link,\n        content: content.substring(0, 8000), // 限制内容长度\n      })\n    }\n  }\n\n  return contents\n}\n",[841,8649,8650,8654,8658,8662,8666,8670,8674,8678,8682,8687,8691,8695,8700,8705,8709,8713,8717,8721,8726,8731,8735,8739,8743,8747,8751,8755,8759,8763,8767,8771,8775,8779,8784,8788,8792,8796,8800,8804],{"__ignoreMap":839},[844,8651,8652],{"class":846,"line":4},[844,8653,892],{},[844,8655,8656],{"class":846,"line":708},[844,8657,1038],{},[844,8659,8660],{"class":846,"line":110},[844,8661,860],{"emptyLinePlaceholder":859},[844,8663,8664],{"class":846,"line":69},[844,8665,1047],{},[844,8667,8668],{"class":846,"line":132},[844,8669,911],{},[844,8671,8672],{"class":846,"line":78},[844,8673,1056],{},[844,8675,8676],{"class":846,"line":162},[844,8677,1061],{},[844,8679,8680],{"class":846,"line":66},[844,8681,860],{"emptyLinePlaceholder":859},[844,8683,8684],{"class":846,"line":63},[844,8685,8686],{},"    // 移除脚本标签、样式标签和其他非内容元素\n",[844,8688,8689],{"class":846,"line":91},[844,8690,1075],{},[844,8692,8693],{"class":846,"line":939},[844,8694,860],{"emptyLinePlaceholder":859},[844,8696,8697],{"class":846,"line":945},[844,8698,8699],{},"    // 提取主要内容（这是一种简化的方法）\n",[844,8701,8702],{"class":846,"line":388},[844,8703,8704],{},"    // 对于生产环境，你可能想使用更复杂的内容提取方法\n",[844,8706,8707],{"class":846,"line":204},[844,8708,1094],{},[844,8710,8711],{"class":846,"line":960},[844,8712,860],{"emptyLinePlaceholder":859},[844,8714,8715],{"class":846,"line":60},[844,8716,1103],{},[844,8718,8719],{"class":846,"line":312},[844,8720,997],{},[844,8722,8723],{"class":846,"line":207},[844,8724,8725],{},"    console.error(`从${url}获取内容时出错:`, error)\n",[844,8727,8728],{"class":846,"line":337},[844,8729,8730],{},"    return '' // 如果无法获取内容，则返回空字符串\n",[844,8732,8733],{"class":846,"line":12},[844,8734,1012],{},[844,8736,8737],{"class":846,"line":180},[844,8738,1017],{},[844,8740,8741],{"class":846,"line":174},[844,8742,860],{"emptyLinePlaceholder":859},[844,8744,8745],{"class":846,"line":304},[844,8746,1134],{},[844,8748,8749],{"class":846,"line":155},[844,8750,1139],{},[844,8752,8753],{"class":846,"line":21},[844,8754,860],{"emptyLinePlaceholder":859},[844,8756,8757],{"class":846,"line":23},[844,8758,1148],{},[844,8760,8761],{"class":846,"line":136},[844,8762,1153],{},[844,8764,8765],{"class":846,"line":22},[844,8766,1158],{},[844,8768,8769],{"class":846,"line":1161},[844,8770,1164],{},[844,8772,8773],{"class":846,"line":24},[844,8774,1169],{},[844,8776,8777],{"class":846,"line":297},[844,8778,1174],{},[844,8780,8781],{"class":846,"line":57},[844,8782,8783],{},"        content: content.substring(0, 8000), // 限制内容长度\n",[844,8785,8786],{"class":846,"line":94},[844,8787,1184],{},[844,8789,8790],{"class":846,"line":72},[844,8791,1189],{},[844,8793,8794],{"class":846,"line":100},[844,8795,1012],{},[844,8797,8798],{"class":846,"line":54},[844,8799,860],{"emptyLinePlaceholder":859},[844,8801,8802],{"class":846,"line":25},[844,8803,1202],{},[844,8805,8806],{"class":846,"line":1205},[844,8807,1017],{},[826,8809,8811],{"id":8810},"_4-集成openai-api","4. 集成OpenAI API",[758,8813,8814],{},"现在，我们将使用OpenAI的API来总结和分析内容：",[834,8816,8818],{"className":836,"code":8817,"language":838,"meta":839,"style":839},"const { OpenAI } = require('openai')\nrequire('dotenv').config()\n\nconst openai = new OpenAI({\n  apiKey: process.env.OPENAI_API_KEY,\n})\n\nasync function summarizeWithAI(stockSymbol, contents) {\n  // 为AI准备内容\n  const contentText = contents\n    .map((item) => `来源: ${item.title} (${item.url})\\n${item.content}\\n\\n`)\n    .join('')\n\n  // 为AI创建提示\n  const prompt = `\n    你是一位金融分析师助手。以下是关于股票${stockSymbol}的网络搜索结果。\n    请分析这些结果并提供：\n    \n    1. 最近关键发展的摘要\n    2. 市场情绪分析（积极、消极、中性）\n    3. 对股价的潜在影响\n    4. 提到的关键财务指标\n    5. 识别出的任何风险或机会\n    \n    使你的分析简洁、事实性强，并专注于对投资者有价值的信息。\n    \n    搜索结果:\n    ${contentText}\n  `\n\n  try {\n    const response = await openai.chat.completions.create({\n      model: 'gpt-4',\n      messages: [\n        { role: 'system', content: '你是一位金融分析师助手，帮助分析来自网络搜索结果的股票信息。' },\n        { role: 'user', content: prompt },\n      ],\n      temperature: 0.2, // 较低的温度以获得更事实性的回应\n      max_tokens: 1500,\n    })\n\n    return response.choices[0].message.content\n  } catch (error) {\n    console.error('生成AI摘要时出错:', error)\n    throw error\n  }\n}\n",[841,8819,8820,8824,8828,8832,8836,8840,8844,8848,8852,8857,8861,8866,8870,8874,8879,8883,8888,8893,8897,8902,8907,8912,8917,8922,8926,8931,8935,8940,8944,8948,8952,8956,8960,8964,8968,8973,8977,8981,8986,8990,8994,8998,9002,9006,9011,9015,9019],{"__ignoreMap":839},[844,8821,8822],{"class":846,"line":4},[844,8823,1224],{},[844,8825,8826],{"class":846,"line":708},[844,8827,897],{},[844,8829,8830],{"class":846,"line":110},[844,8831,860],{"emptyLinePlaceholder":859},[844,8833,8834],{"class":846,"line":69},[844,8835,1237],{},[844,8837,8838],{"class":846,"line":132},[844,8839,1242],{},[844,8841,8842],{"class":846,"line":78},[844,8843,1247],{},[844,8845,8846],{"class":846,"line":162},[844,8847,860],{"emptyLinePlaceholder":859},[844,8849,8850],{"class":846,"line":66},[844,8851,1256],{},[844,8853,8854],{"class":846,"line":63},[844,8855,8856],{},"  // 为AI准备内容\n",[844,8858,8859],{"class":846,"line":91},[844,8860,1266],{},[844,8862,8863],{"class":846,"line":939},[844,8864,8865],{},"    .map((item) => `来源: ${item.title} (${item.url})\\n${item.content}\\n\\n`)\n",[844,8867,8868],{"class":846,"line":945},[844,8869,1276],{},[844,8871,8872],{"class":846,"line":388},[844,8873,860],{"emptyLinePlaceholder":859},[844,8875,8876],{"class":846,"line":204},[844,8877,8878],{},"  // 为AI创建提示\n",[844,8880,8881],{"class":846,"line":960},[844,8882,1290],{},[844,8884,8885],{"class":846,"line":60},[844,8886,8887],{},"    你是一位金融分析师助手。以下是关于股票${stockSymbol}的网络搜索结果。\n",[844,8889,8890],{"class":846,"line":312},[844,8891,8892],{},"    请分析这些结果并提供：\n",[844,8894,8895],{"class":846,"line":207},[844,8896,1305],{},[844,8898,8899],{"class":846,"line":337},[844,8900,8901],{},"    1. 最近关键发展的摘要\n",[844,8903,8904],{"class":846,"line":12},[844,8905,8906],{},"    2. 市场情绪分析（积极、消极、中性）\n",[844,8908,8909],{"class":846,"line":180},[844,8910,8911],{},"    3. 对股价的潜在影响\n",[844,8913,8914],{"class":846,"line":174},[844,8915,8916],{},"    4. 提到的关键财务指标\n",[844,8918,8919],{"class":846,"line":304},[844,8920,8921],{},"    5. 识别出的任何风险或机会\n",[844,8923,8924],{"class":846,"line":155},[844,8925,1305],{},[844,8927,8928],{"class":846,"line":21},[844,8929,8930],{},"    使你的分析简洁、事实性强，并专注于对投资者有价值的信息。\n",[844,8932,8933],{"class":846,"line":23},[844,8934,1305],{},[844,8936,8937],{"class":846,"line":136},[844,8938,8939],{},"    搜索结果:\n",[844,8941,8942],{"class":846,"line":22},[844,8943,1353],{},[844,8945,8946],{"class":846,"line":1161},[844,8947,1358],{},[844,8949,8950],{"class":846,"line":24},[844,8951,860],{"emptyLinePlaceholder":859},[844,8953,8954],{"class":846,"line":297},[844,8955,911],{},[844,8957,8958],{"class":846,"line":57},[844,8959,1371],{},[844,8961,8962],{"class":846,"line":94},[844,8963,1376],{},[844,8965,8966],{"class":846,"line":72},[844,8967,1381],{},[844,8969,8970],{"class":846,"line":100},[844,8971,8972],{},"        { role: 'system', content: '你是一位金融分析师助手，帮助分析来自网络搜索结果的股票信息。' },\n",[844,8974,8975],{"class":846,"line":54},[844,8976,1411],{},[844,8978,8979],{"class":846,"line":25},[844,8980,1416],{},[844,8982,8983],{"class":846,"line":1205},[844,8984,8985],{},"      temperature: 0.2, // 较低的温度以获得更事实性的回应\n",[844,8987,8988],{"class":846,"line":51},[844,8989,1426],{},[844,8991,8992],{"class":846,"line":341},[844,8993,948],{},[844,8995,8996],{"class":846,"line":172},[844,8997,860],{"emptyLinePlaceholder":859},[844,8999,9000],{"class":846,"line":290},[844,9001,1440],{},[844,9003,9004],{"class":846,"line":166},[844,9005,997],{},[844,9007,9008],{"class":846,"line":278},[844,9009,9010],{},"    console.error('生成AI摘要时出错:', error)\n",[844,9012,9013],{"class":846,"line":264},[844,9014,1007],{},[844,9016,9017],{"class":846,"line":1437},[844,9018,1012],{},[844,9020,9021],{"class":846,"line":254},[844,9022,1017],{},[826,9024,9026],{"id":9025},"_5-整合所有内容","5. 整合所有内容",[758,9028,9029],{},"最后，让我们创建将所有内容整合在一起的主函数：",[834,9031,9033],{"className":836,"code":9032,"language":838,"meta":839,"style":839},"async function analyzeStock(stockSymbol) {\n  try {\n    console.log(`分析股票: ${stockSymbol}...`)\n\n    // 步骤1：搜索有关股票的最新信息\n    const searchQuery = `${stockSymbol} 股票 新闻 财务分析 最新发展`\n    const searchResults = await searchWeb(searchQuery, 8)\n\n    // 步骤2：从搜索结果中获取内容\n    const contents = await fetchContentsFromSearchResults(searchResults)\n\n    // 步骤3：生成AI摘要和分析\n    const analysis = await summarizeWithAI(stockSymbol, contents)\n\n    return {\n      stockSymbol,\n      searchResults,\n      analysis,\n    }\n  } catch (error) {\n    console.error(`分析股票${stockSymbol}时出错:`, error)\n    throw error\n  }\n}\n\n// 使用示例\nanalyzeStock('AAPL')\n  .then((result) => {\n    console.log('分析完成:')\n    console.log(result.analysis)\n  })\n  .catch((error) => {\n    console.error('分析失败:', error)\n  })\n",[841,9034,9035,9039,9043,9048,9052,9057,9062,9066,9070,9075,9079,9083,9088,9092,9096,9100,9104,9108,9112,9116,9120,9125,9129,9133,9137,9141,9146,9150,9154,9159,9163,9167,9171,9176],{"__ignoreMap":839},[844,9036,9037],{"class":846,"line":4},[844,9038,1481],{},[844,9040,9041],{"class":846,"line":708},[844,9042,911],{},[844,9044,9045],{"class":846,"line":110},[844,9046,9047],{},"    console.log(`分析股票: ${stockSymbol}...`)\n",[844,9049,9050],{"class":846,"line":69},[844,9051,860],{"emptyLinePlaceholder":859},[844,9053,9054],{"class":846,"line":132},[844,9055,9056],{},"    // 步骤1：搜索有关股票的最新信息\n",[844,9058,9059],{"class":846,"line":78},[844,9060,9061],{},"    const searchQuery = `${stockSymbol} 股票 新闻 财务分析 最新发展`\n",[844,9063,9064],{"class":846,"line":162},[844,9065,1509],{},[844,9067,9068],{"class":846,"line":66},[844,9069,860],{"emptyLinePlaceholder":859},[844,9071,9072],{"class":846,"line":63},[844,9073,9074],{},"    // 步骤2：从搜索结果中获取内容\n",[844,9076,9077],{"class":846,"line":91},[844,9078,1523],{},[844,9080,9081],{"class":846,"line":939},[844,9082,860],{"emptyLinePlaceholder":859},[844,9084,9085],{"class":846,"line":945},[844,9086,9087],{},"    // 步骤3：生成AI摘要和分析\n",[844,9089,9090],{"class":846,"line":388},[844,9091,1537],{},[844,9093,9094],{"class":846,"line":204},[844,9095,860],{"emptyLinePlaceholder":859},[844,9097,9098],{"class":846,"line":960},[844,9099,1546],{},[844,9101,9102],{"class":846,"line":60},[844,9103,1551],{},[844,9105,9106],{"class":846,"line":312},[844,9107,1556],{},[844,9109,9110],{"class":846,"line":207},[844,9111,1561],{},[844,9113,9114],{"class":846,"line":337},[844,9115,1189],{},[844,9117,9118],{"class":846,"line":12},[844,9119,997],{},[844,9121,9122],{"class":846,"line":180},[844,9123,9124],{},"    console.error(`分析股票${stockSymbol}时出错:`, error)\n",[844,9126,9127],{"class":846,"line":174},[844,9128,1007],{},[844,9130,9131],{"class":846,"line":304},[844,9132,1012],{},[844,9134,9135],{"class":846,"line":155},[844,9136,1017],{},[844,9138,9139],{"class":846,"line":21},[844,9140,860],{"emptyLinePlaceholder":859},[844,9142,9143],{"class":846,"line":23},[844,9144,9145],{},"// 使用示例\n",[844,9147,9148],{"class":846,"line":136},[844,9149,1600],{},[844,9151,9152],{"class":846,"line":22},[844,9153,1605],{},[844,9155,9156],{"class":846,"line":1161},[844,9157,9158],{},"    console.log('分析完成:')\n",[844,9160,9161],{"class":846,"line":24},[844,9162,1615],{},[844,9164,9165],{"class":846,"line":297},[844,9166,1620],{},[844,9168,9169],{"class":846,"line":57},[844,9170,1625],{},[844,9172,9173],{"class":846,"line":94},[844,9174,9175],{},"    console.error('分析失败:', error)\n",[844,9177,9178],{"class":846,"line":72},[844,9179,1620],{},[753,9181,9183],{"id":9182},"实际应用股票分析仪表板","实际应用：股票分析仪表板",[758,9185,9186],{},"对于我们公司的实验性项目，我们将此功能集成到一个仪表板中，使分析师能够：",[772,9188,9189,9192,9195,9198],{},[775,9190,9191],{},"输入多个股票代码进行分析",[775,9193,9194],{},"自定义搜索参数（时间范围、关注领域）",[775,9196,9197],{},"并排比较AI生成的摘要",[775,9199,9200],{},"保存并跟踪分析结果，以识别趋势",[758,9202,9203],{},"仪表板看起来像这样：",[834,9205,9207],{"className":836,"code":9206,"language":838,"meta":839,"style":839},"// React组件示例（简化版）\nfunction StockAnalysisDashboard() {\n  const [stocks, setStocks] = useState([])\n  const [loading, setLoading] = useState({})\n  const [analyses, setAnalyses] = useState({})\n\n  const addStock = (symbol) => {\n    if (!stocks.includes(symbol)) {\n      setStocks([...stocks, symbol])\n      analyzeStockAndUpdateState(symbol)\n    }\n  }\n\n  const analyzeStockAndUpdateState = async (symbol) => {\n    setLoading((prev) => ({ ...prev, [symbol]: true }))\n    try {\n      const result = await analyzeStock(symbol)\n      setAnalyses((prev) => ({ ...prev, [symbol]: result }))\n    } catch (error) {\n      console.error(`分析${symbol}时出错:`, error)\n    } finally {\n      setLoading((prev) => ({ ...prev, [symbol]: false }))\n    }\n  }\n\n  return (\n    \u003Cdiv className=\"dashboard\">\n      \u003Ch1>股票分析仪表板\u003C/h1>\n\n      \u003Cdiv className=\"stock-input\">\n        \u003Cinput\n          type=\"text\"\n          placeholder=\"输入股票代码（例如，AAPL）\"\n          onKeyPress={(e) => e.key === 'Enter' && addStock(e.target.value)}\n        />\n      \u003C/div>\n\n      \u003Cdiv className=\"stock-analyses\">\n        {stocks.map((symbol) => (\n          \u003Cdiv key={symbol} className=\"stock-card\">\n            \u003Ch2>{symbol}\u003C/h2>\n            {loading[symbol] ? (\n              \u003Cp>加载分析中...\u003C/p>\n            ) : analyses[symbol] ? (\n              \u003Cdiv>\n                \u003Ch3>AI分析\u003C/h3>\n                \u003Cdiv className=\"analysis-content\">{analyses[symbol].analysis}\u003C/div>\n                \u003Ch3>来源\u003C/h3>\n                \u003Cul>\n                  {analyses[symbol].searchResults.map((result, i) => (\n                    \u003Cli key={i}>\n                      \u003Ca href={result.link} target=\"_blank\" rel=\"noopener noreferrer\">\n                        {result.title}\n                      \u003C/a>\n                    \u003C/li>\n                  ))}\n                \u003C/ul>\n              \u003C/div>\n            ) : (\n              \u003Cp>没有可用的分析\u003C/p>\n            )}\n          \u003C/div>\n        ))}\n      \u003C/div>\n    \u003C/div>\n  )\n}\n",[841,9208,9209,9214,9218,9222,9226,9230,9234,9238,9242,9246,9250,9254,9258,9262,9266,9270,9274,9278,9282,9286,9291,9295,9299,9303,9307,9311,9315,9319,9324,9328,9332,9336,9340,9345,9349,9353,9357,9361,9365,9369,9373,9377,9381,9386,9390,9394,9399,9403,9408,9412,9416,9420,9424,9428,9432,9436,9440,9444,9448,9452,9457,9461,9465,9469,9473,9477,9481],{"__ignoreMap":839},[844,9210,9211],{"class":846,"line":4},[844,9212,9213],{},"// React组件示例（简化版）\n",[844,9215,9216],{"class":846,"line":708},[844,9217,1673],{},[844,9219,9220],{"class":846,"line":110},[844,9221,1678],{},[844,9223,9224],{"class":846,"line":69},[844,9225,1683],{},[844,9227,9228],{"class":846,"line":132},[844,9229,1688],{},[844,9231,9232],{"class":846,"line":78},[844,9233,860],{"emptyLinePlaceholder":859},[844,9235,9236],{"class":846,"line":162},[844,9237,1697],{},[844,9239,9240],{"class":846,"line":66},[844,9241,1702],{},[844,9243,9244],{"class":846,"line":63},[844,9245,1707],{},[844,9247,9248],{"class":846,"line":91},[844,9249,1712],{},[844,9251,9252],{"class":846,"line":939},[844,9253,1189],{},[844,9255,9256],{"class":846,"line":945},[844,9257,1012],{},[844,9259,9260],{"class":846,"line":388},[844,9261,860],{"emptyLinePlaceholder":859},[844,9263,9264],{"class":846,"line":204},[844,9265,1729],{},[844,9267,9268],{"class":846,"line":960},[844,9269,1734],{},[844,9271,9272],{"class":846,"line":60},[844,9273,1739],{},[844,9275,9276],{"class":846,"line":312},[844,9277,1744],{},[844,9279,9280],{"class":846,"line":207},[844,9281,1749],{},[844,9283,9284],{"class":846,"line":337},[844,9285,1754],{},[844,9287,9288],{"class":846,"line":12},[844,9289,9290],{},"      console.error(`分析${symbol}时出错:`, error)\n",[844,9292,9293],{"class":846,"line":180},[844,9294,1764],{},[844,9296,9297],{"class":846,"line":174},[844,9298,1769],{},[844,9300,9301],{"class":846,"line":304},[844,9302,1189],{},[844,9304,9305],{"class":846,"line":155},[844,9306,1012],{},[844,9308,9309],{"class":846,"line":21},[844,9310,860],{"emptyLinePlaceholder":859},[844,9312,9313],{"class":846,"line":23},[844,9314,1786],{},[844,9316,9317],{"class":846,"line":136},[844,9318,1791],{},[844,9320,9321],{"class":846,"line":22},[844,9322,9323],{},"      \u003Ch1>股票分析仪表板\u003C/h1>\n",[844,9325,9326],{"class":846,"line":1161},[844,9327,860],{"emptyLinePlaceholder":859},[844,9329,9330],{"class":846,"line":24},[844,9331,1805],{},[844,9333,9334],{"class":846,"line":297},[844,9335,1810],{},[844,9337,9338],{"class":846,"line":57},[844,9339,1815],{},[844,9341,9342],{"class":846,"line":94},[844,9343,9344],{},"          placeholder=\"输入股票代码（例如，AAPL）\"\n",[844,9346,9347],{"class":846,"line":72},[844,9348,1825],{},[844,9350,9351],{"class":846,"line":100},[844,9352,1830],{},[844,9354,9355],{"class":846,"line":54},[844,9356,1835],{},[844,9358,9359],{"class":846,"line":25},[844,9360,860],{"emptyLinePlaceholder":859},[844,9362,9363],{"class":846,"line":1205},[844,9364,1844],{},[844,9366,9367],{"class":846,"line":51},[844,9368,1849],{},[844,9370,9371],{"class":846,"line":341},[844,9372,1854],{},[844,9374,9375],{"class":846,"line":172},[844,9376,1859],{},[844,9378,9379],{"class":846,"line":290},[844,9380,1864],{},[844,9382,9383],{"class":846,"line":166},[844,9384,9385],{},"              \u003Cp>加载分析中...\u003C/p>\n",[844,9387,9388],{"class":846,"line":278},[844,9389,1874],{},[844,9391,9392],{"class":846,"line":264},[844,9393,1879],{},[844,9395,9396],{"class":846,"line":1437},[844,9397,9398],{},"                \u003Ch3>AI分析\u003C/h3>\n",[844,9400,9401],{"class":846,"line":254},[844,9402,1889],{},[844,9404,9405],{"class":846,"line":1447},[844,9406,9407],{},"                \u003Ch3>来源\u003C/h3>\n",[844,9409,9410],{"class":846,"line":1453},[844,9411,1899],{},[844,9413,9414],{"class":846,"line":599},[844,9415,1904],{},[844,9417,9418],{"class":846,"line":1462},[844,9419,1909],{},[844,9421,9422],{"class":846,"line":159},[844,9423,1914],{},[844,9425,9426],{"class":846,"line":459},[844,9427,1919],{},[844,9429,9430],{"class":846,"line":269},[844,9431,1924],{},[844,9433,9434],{"class":846,"line":1927},[844,9435,1930],{},[844,9437,9438],{"class":846,"line":235},[844,9439,1935],{},[844,9441,9442],{"class":846,"line":408},[844,9443,1940],{},[844,9445,9446],{"class":846,"line":1943},[844,9447,1946],{},[844,9449,9450],{"class":846,"line":1949},[844,9451,1952],{},[844,9453,9454],{"class":846,"line":320},[844,9455,9456],{},"              \u003Cp>没有可用的分析\u003C/p>\n",[844,9458,9459],{"class":846,"line":1960},[844,9460,1963],{},[844,9462,9463],{"class":846,"line":1966},[844,9464,1969],{},[844,9466,9467],{"class":846,"line":250},[844,9468,1974],{},[844,9470,9471],{"class":846,"line":301},[844,9472,1835],{},[844,9474,9475],{"class":846,"line":1981},[844,9476,1984],{},[844,9478,9479],{"class":846,"line":85},[844,9480,1989],{},[844,9482,9483],{"class":846,"line":1992},[844,9484,1017],{},[753,9486,9487],{"id":9487},"挑战与考虑因素",[758,9489,9490],{},"在构建这个工具的过程中，我们遇到了几个值得注意的挑战：",[826,9492,9494],{"id":9493},"_1-api速率限制和成本","1. API速率限制和成本",[758,9496,9497],{},"SERP API和OpenAI的API都有速率限制和使用成本。对于生产系统，你需要：",[804,9499,9500,9503,9506],{},[775,9501,9502],{},"实现缓存以避免重复搜索",[775,9504,9505],{},"设置使用监控和警报",[775,9507,9508],{},"考虑多个股票的批处理",[826,9510,9512],{"id":9511},"_2-内容提取质量","2. 内容提取质量",[758,9514,9515],{},"从网页中提取有意义的内容可能具有挑战性，原因如下：",[804,9517,9518,9521,9524],{},[775,9519,9520],{},"金融新闻网站的付费墙",[775,9522,9523],{},"通过JavaScript加载的动态内容",[775,9525,9526],{},"不同网站之间的页面结构各异",[758,9528,9529],{},"我们通过以下方式改进了内容提取：",[804,9531,9532,9537,9540],{},[775,9533,9534,9535],{},"使用更复杂的库，如",[841,9536,2048],{},[775,9538,9539],{},"为常见的金融新闻来源实现特定网站的提取器",[775,9541,9542],{},"当无法获取完整内容时，回退到元描述",[826,9544,9546],{"id":9545},"_3-确保分析质量","3. 确保分析质量",[758,9548,9549],{},"为了提高AI生成分析的质量：",[804,9551,9552,9555,9558],{},[775,9553,9554],{},"我们根据金融分析师的反馈微调了提示",[775,9556,9557],{},"通过交叉引用关键主张实现事实检查",[775,9559,9560],{},"添加来源归属，以明确信息来源",[753,9562,9563],{"id":9563},"结论",[758,9565,9566],{},"将网络搜索功能与AI摘要相结合，创建了一个强大的股票分析工具，可以节省数小时的研究时间，并提供更全面的见解。我们在这里概述的实现只是一个起点——有很多方法可以扩展和改进这个系统。",[758,9568,9569],{},"一些潜在的扩展包括：",[804,9571,9572,9575,9578,9581],{},[775,9573,9574],{},"添加专门针对金融新闻的情感分析",[775,9576,9577],{},"纳入历史股价数据进行相关性分析",[775,9579,9580],{},"扩展到包括来自Twitter/X等平台的社交媒体情绪",[775,9582,9583],{},"创建可能影响股价的重大新闻警报",[758,9585,9586],{},"随着AI能力的不断进步，像这样的工具将变得越来越复杂和有价值，用于金融分析和决策。",[758,9588,9589],{},"你是否构建过类似的工具或有改进的想法？我很想在评论中听到你的经验！",[2104,9591,2106],{},{"title":839,"searchDepth":708,"depth":708,"links":9593},[9594,9595,9596,9603,9604,9609],{"id":8416,"depth":708,"text":8416},{"id":8425,"depth":708,"text":8426},{"id":8475,"depth":708,"text":8476,"children":9597},[9598,9599,9600,9601,9602],{"id":8482,"depth":110,"text":8483},{"id":8522,"depth":110,"text":8523},{"id":8640,"depth":110,"text":8641},{"id":8810,"depth":110,"text":8811},{"id":9025,"depth":110,"text":9026},{"id":9182,"depth":708,"text":9183},{"id":9487,"depth":708,"text":9487,"children":9605},[9606,9607,9608],{"id":9493,"depth":110,"text":9494},{"id":9511,"depth":110,"text":9512},{"id":9545,"depth":110,"text":9546},{"id":9563,"depth":708,"text":9563},"如何结合OpenAI和SERP API创建强大的网络搜索和AI摘要工具，用于股票分析和研究",{"date":2128,"image":7962,"alt":9612,"tags":9613,"category":2132,"topics":9614,"published":859,"featured":859},"OpenAI和SERP API集成",[2130,2131],[2134,2135],"/blogs/zh/openai-serp-api-web-search-ai-summary",{"title":8411,"description":9610},"blogs/zh/10.openai-serp-api-web-search-ai-summary","3dnvz0LExYWntz9aav6FJd1X0m0ATWD77_80kMtWlK4",{"id":9620,"title":9621,"body":9622,"description":9898,"extension":2126,"meta":9899,"navigation":859,"ogImage":2459,"path":9903,"seo":9904,"stem":9905,"__hash__":9906},"zh/blogs/zh/11.cloud-based-scraping-solutions.md","网页抓取迁移到云端",{"type":750,"value":9623,"toc":9883},[9624,9627,9630,9636,9639,9641,9644,9648,9655,9662,9665,9669,9676,9679,9682,9686,9692,9695,9721,9724,9727,9730,9764,9766,9770,9773,9777,9780,9786,9790,9796,9799,9801,9804,9807,9847,9850,9852,9855,9858,9875,9878,9881],[753,9625,9626],{"id":9626},"我尝试解决的问题",[758,9628,9629],{},"上个月，我在做一个需要从多个JavaScript密集型网站抓取数据的项目。我一开始使用了常见的工具——本地运行Puppeteer和Playwright——但很快就遇到了瓶颈。我的笔记本风扇不停地转，内存使用率飙升，而且扩展到几个并发请求以上似乎是不可能的。",[758,9631,9632,9633],{},"我心想：",[2156,9634,9635],{},"\"一定有更好的方法可以做到这一点，而不会把我的MacBook变成一个暖气片。\"",[758,9637,9638],{},"于是我开始研究基于云的替代方案，这些方案可以处理浏览器自动化的重任，同时保持我的本地环境轻量化。以下是我花了一周时间测试不同解决方案后的发现。",[2163,9640],{},[753,9642,9643],{"id":9643},"我测试的基于云的无头浏览器服务",[826,9645,9647],{"id":9646},"apify一站式平台","Apify：一站式平台",[758,9649,9650,9651,9654],{},"我首先尝试了",[2177,9652,2182],{"href":2179,"rel":9653},[2181],"，它感觉就像是\"网页抓取的AWS\"。他们的云平台提供预构建的\"Actors\"（可以理解为抓取用的无服务器函数），内置了代理、调度和存储等所有功能。",[758,9656,9657,9658,9661],{},"我最喜欢的是，我可以拿我现有的Node.js爬虫代码，做最小的修改，然后将其部署为无服务器Actor。他们提供的",[2177,9659,2191],{"href":2189,"rel":9660},[2181],"库使这个转换过程出奇地顺利。",[758,9663,9664],{},"仪表板让你可以查看运行情况、内存使用和日志——这比我在本地使用console.log进行调试的方式要先进得多。",[826,9666,9668],{"id":9667},"browserlessio当你只需要浏览器时","Browserless.io：当你只需要浏览器时",[758,9670,9671,9672,9675],{},"接下来，我尝试了",[2177,9673,2207],{"href":2205,"rel":9674},[2181],"，它采用了更加专注的方法。它本质上是一个托管的无头Chrome和Playwright环境，你可以通过REST API访问。",[758,9677,9678],{},"最棒的部分？我可以保留大部分现有的Puppeteer代码，只需将其指向他们的服务，而不是启动本地浏览器。他们的API在后台处理所有的代理轮换和CAPTCHA解决方案。",[758,9680,9681],{},"对于已经投入使用Puppeteer/Playwright工作流的团队来说，这感觉是阻力最小的路径。",[826,9683,9685],{"id":9684},"scrapingbee一行代码解决方案","ScrapingBee：一行代码解决方案",[758,9687,9688,9691],{},[2177,9689,2225],{"href":2223,"rel":9690},[2181],"采用了一种不同的方法，在我只想要结果而不想费心处理浏览器自动化代码的日子里，我真的很欣赏这种方法。",[758,9693,9694],{},"他们的API非常简单——发送一个带有目标URL的HTTP请求，然后获取完全渲染的HTML。不需要管理浏览器实例，处理JavaScript执行，或担心被封禁。",[834,9696,9698],{"className":836,"code":9697,"language":838,"meta":839,"style":839},"// 这真的就是全部所需的代码\nconst response = await fetch(\n  `https://app.scrapingbee.com/api/v1/?api_key=${apiKey}&url=${targetUrl}&render_js=true`,\n)\nconst html = await response.text()\n",[841,9699,9700,9705,9709,9713,9717],{"__ignoreMap":839},[844,9701,9702],{"class":846,"line":4},[844,9703,9704],{},"// 这真的就是全部所需的代码\n",[844,9706,9707],{"class":846,"line":708},[844,9708,2244],{},[844,9710,9711],{"class":846,"line":110},[844,9712,2249],{},[844,9714,9715],{"class":846,"line":69},[844,9716,2254],{},[844,9718,9719],{"class":846,"line":132},[844,9720,2259],{},[758,9722,9723],{},"当我需要快速结果而不想承担浏览器自动化的认知负担时，我发现自己会选择这种方式。",[826,9725,9726],{"id":9726},"其他值得一提的服务",[758,9728,9729],{},"我还简单测试了：",[804,9731,9732,9740,9748,9756],{},[775,9733,9734,9739],{},[778,9735,9736],{},[2177,9737,2280],{"href":2278,"rel":9738},[2181],"（前身是Splash）：凭借其反封禁技术，非常适合JavaScript密集型页面",[775,9741,9742,9747],{},[778,9743,9744],{},[2177,9745,2290],{"href":2288,"rel":9746},[2181],"：如果你需要自动化抓取之外的工作流程，这是一个很好的选择",[775,9749,9750,9755],{},[778,9751,9752],{},[2177,9753,2300],{"href":2298,"rel":9754},[2181],"：使用真实浏览器规避检测的可靠选项",[775,9757,9758,9763],{},[778,9759,9760],{},[2177,9761,2310],{"href":2308,"rel":9762},[2181],"：如果你已经在使用Scrapy，这是完美的选择",[2163,9765],{},[753,9767,9769],{"id":9768},"无服务器路线diy云抓取","无服务器路线：DIY云抓取",[758,9771,9772],{},"对于某些项目，我想要更多地控制执行环境，同时仍然避免本地资源限制。这时我开始探索在无服务器函数中部署无头浏览器：",[826,9774,9776],{"id":9775},"aws-lambda实验","AWS Lambda实验",[758,9778,9779],{},"我在Lambda Layers中打包了带有无头Chrome的Puppeteer，令人惊讶的是它工作得很好。设置比使用专用服务更复杂，但对于我偶尔的抓取需求来说，按执行时间计费的模式很合理。",[758,9781,9782,9783,9785],{},"关键的发现是使用",[841,9784,2333],{},"包，它提供了一个适当编译的Chromium版本，保持在Lambda的大小限制之内。",[826,9787,9789],{"id":9788},"google-cloud-functions体验","Google Cloud Functions体验",[758,9791,9792,9793,9795],{},"Google Cloud Functions（第二代）设置起来更容易，因为它已经包含了无头Chrome所需的许多系统包。我只需要使用",[841,9794,2344],{},"来保持部署大小可控。",[758,9797,9798],{},"这种方法为我提供了最佳的控制和可扩展性平衡，特别是在那些需要围绕抓取过程进行自定义逻辑的项目中。",[2163,9800],{},[753,9802,9803],{"id":9803},"我学到的经验和建议",[758,9805,9806],{},"经过几周的测试，如果你面临类似的挑战，以下是我的实用建议：",[804,9808,9809,9818,9827,9835],{},[775,9810,9811,9814,9815,9817],{},[778,9812,9813],{},"如果你刚刚开始","：先尝试",[778,9816,2225],{},"。单个API调用的简单性将为你节省数小时的设置时间。",[775,9819,9820,9823,9824,9826],{},[778,9821,9822],{},"如果你有现有的Puppeteer/Playwright代码","：",[778,9825,2207],{},"提供了最平滑的过渡，只需最少的代码更改。",[775,9828,9829,9823,9832,9834],{},[778,9830,9831],{},"对于大规模、复杂的项目",[778,9833,2182],{},"提供了最完整的生态系统，内置了调度、存储和代理管理。",[775,9836,9837,9840,9841,9843,9844,9846],{},[778,9838,9839],{},"如果你熟悉云服务","：部署到",[778,9842,2394],{},"或",[778,9845,2398],{},"可以给你最大的控制权，潜在成本更低，特别是对于间歇性工作负载。",[758,9848,9849],{},"我学到的最大教训？**不要让基础设施问题限制你的数据收集项目。**这些云解决方案已经成熟到了这样一个程度：几乎没有什么好理由在本地机器上运行资源密集型浏览器。",[2163,9851],{},[753,9853,9854],{"id":9854},"我的抓取项目下一步计划",[758,9856,9857],{},"目前，我采用了一种混合方法：",[804,9859,9860,9865,9870],{},[775,9861,9862,9864],{},[778,9863,2225],{},"用于快速、一次性的抓取任务",[775,9866,9867,9869],{},[778,9868,2182],{},"用于定期运行的计划数据收集作业",[775,9871,9872,9874],{},[778,9873,2394],{},"用于需要自定义逻辑的专门抓取器",[758,9876,9877],{},"在我的下一篇文章中，我将分享一些实际的代码示例，展示我如何将这些服务集成到一个统一的数据管道中，为我的分析系统提供数据。",[758,9879,9880],{},"你尝试过这些服务吗？或者你有其他推荐的云抓取解决方案吗？请在评论中告诉我！",[2104,9882,2106],{},{"title":839,"searchDepth":708,"depth":708,"links":9884},[9885,9886,9892,9896,9897],{"id":9626,"depth":708,"text":9626},{"id":9643,"depth":708,"text":9643,"children":9887},[9888,9889,9890,9891],{"id":9646,"depth":110,"text":9647},{"id":9667,"depth":110,"text":9668},{"id":9684,"depth":110,"text":9685},{"id":9726,"depth":110,"text":9726},{"id":9768,"depth":708,"text":9769,"children":9893},[9894,9895],{"id":9775,"depth":110,"text":9776},{"id":9788,"depth":110,"text":9789},{"id":9803,"depth":708,"text":9803},{"id":9854,"depth":708,"text":9854},"我探索了不需要在本地运行Puppeteer或Playwright等重型服务的基于云的无头浏览器抓取替代方案",{"date":2458,"image":2459,"alt":9900,"tags":9901,"category":2463,"topics":9902,"published":859},"基于云的网页抓取解决方案",[2131,2462],[2135,2134],"/blogs/zh/cloud-based-scraping-solutions",{"title":9621,"description":9898},"blogs/zh/11.cloud-based-scraping-solutions","uM9QzpKp_NhPukz5GZY1gr77cNsjvKMMxPT4VW0tsZQ",{"id":9908,"title":9909,"body":9910,"description":10429,"extension":2126,"meta":10430,"navigation":859,"ogImage":3011,"path":10434,"seo":10435,"stem":10436,"__hash__":10437},"zh/blogs/zh/12.naval-ravikant-specific-knowledge-responsibility-assets.md","从《纳瓦尔宝典》中学到的：专属知识、责任和资产",{"type":750,"value":9911,"toc":10405},[9912,9915,9918,9921,9926,9929,9932,9935,9938,9958,9961,9964,9968,9971,9974,9977,9991,9994,9998,10001,10004,10010,10014,10017,10023,10049,10052,10055,10058,10061,10064,10068,10071,10074,10085,10088,10091,10094,10097,10100,10107,10110,10113,10140,10143,10146,10149,10154,10157,10161,10164,10167,10170,10173,10199,10205,10209,10212,10215,10229,10232,10235,10260,10263,10266,10277,10280,10283,10286,10290,10293,10296,10307,10310,10314,10317,10320,10331,10334,10338,10341,10346,10349,10352,10356,10359,10362,10365,10376,10379,10382,10385,10388,10399,10402],[758,9913,9914],{},"最近我一直在思考杠杆——不是\"更聪明地工作\"那种，而是结构性的杠杆。那种能让你在不燃尽自己的情况下创造不成比例价值的杠杆。这促使我重新阅读了《纳瓦尔宝典》。",[758,9916,9917],{},"我终于读完了这本书，现在我明白为什么它会出现在那么多\"每年必读\"的书单上。这本书不是典型的商业手册。它更像是一套关于如何好好生活的心智模型——以及如何在不出卖灵魂（或整个日程表）的情况下积累财富。",[758,9919,9920],{},"书中有很多令人难忘的想法：杠杆、判断力、复利、平和、阅读、长期游戏。但如果我把个人收获压缩成一句话，那就是：",[758,9922,9923],{},[778,9924,9925],{},"建立我的专属知识，承担更多责任，构建资产。",[758,9927,9928],{},"这三者感觉像是一个实用的路线图——不是励志的，不是模糊的——只是一个我可以执行多年的清晰循环。",[753,9930,9931],{"id":9931},"为什么这个三元组对我很重要",[758,9933,9934],{},"过去，我经常像清单一样对待成长：学习一个新框架，交付另一个项目，读另一本书，做另一件\"有成效\"的事情。这有效……直到它不再有效。你最终会有动作但没有动力。",[758,9936,9937],{},"这本书促使我审视动力背后更深层的结构：",[804,9939,9940,9946,9952],{},[775,9941,9942,9945],{},[778,9943,9944],{},"专属知识","给你一个不容易被复制的优势。",[775,9947,9948,9951],{},[778,9949,9950],{},"责任","给你做决定和获取收益的权利。",[775,9953,9954,9957],{},[778,9955,9956],{},"资产","给你可以在你睡觉时复利的杠杆。",[758,9959,9960],{},"这不是三个独立的技巧。这是一个系统。",[758,9962,9963],{},"你可以有技能但没有自主权。你可以有自主权但没有可扩展的东西。你可以构建东西但从不培养品味和判断力。这个三元组闭合了循环。",[753,9965,9967],{"id":9966},"_1-建立专属知识无法克隆的那种","1) 建立专属知识：无法克隆的那种",[758,9969,9970],{},"\"专属知识\"这个短语听起来像是一个生产力流行语，直到你真正深入思考它。",[758,9972,9973],{},"专属知识不是我们大多数人所说的\"聪明\"。它不是证书。它不是擅长面试。它甚至不是泛泛意义上的\"优秀工程师\"。",[758,9975,9976],{},"专属知识是以下的组合：",[804,9978,9979,9982,9985,9988],{},[775,9980,9981],{},"通过痴迷建立的真正专业知识，",[775,9983,9984],{},"生活经验，",[775,9986,9987],{},"品味和判断力，",[775,9989,9990],{},"以及在一个人身上独特堆叠的一套技能。",[758,9992,9993],{},"它是让人们说的知识：\"我不知道你是怎么做到的，但你让它看起来很明显。\"",[826,9995,9997],{"id":9996},"不舒服的真相它无法快速获得","不舒服的真相：它无法快速获得",[758,9999,10000],{},"这是我需要听到的部分：专属知识是赚来的，不是获得的。",[758,10002,10003],{},"你不能把它作为课程下载。\n你不能在一个周末\"赶上\"。\n你通过多年的重复和好奇心来建立它——通常是通过追随真正让你感兴趣的东西。",[758,10005,10006,10007],{},"这有一个战略含义：",[778,10008,10009],{},"我应该停止为流行的东西优化，开始为我可以复利的东西优化。",[826,10011,10013],{"id":10012},"专属知识在实践中是什么样子","\"专属知识\"在实践中是什么样子",[758,10015,10016],{},"对我来说，专属知识不是\"编码\"。不是\"产品管理\"。不是\"设计\"。",[758,10018,10019,10020,9823],{},"它是多种技能的",[778,10021,10022],{},"交集",[804,10024,10025,10031,10037,10043],{},[775,10026,10027,10030],{},[778,10028,10029],{},"技术执行"," - 我可以端到端构建，而不仅仅是规范",[775,10032,10033,10036],{},[778,10034,10035],{},"产品品味"," - 我在构建之前就知道\"好\"是什么样子",[775,10038,10039,10042],{},[778,10040,10041],{},"简化"," - 我可以让复杂的东西感觉显而易见",[775,10044,10045,10048],{},[778,10046,10047],{},"执行领导力"," - 我可以交付，而不仅仅是提议",[758,10050,10051],{},"这种组合比任何单一技能都更难替代。如果我继续堆叠它，它会随着时间的推移变得更有价值。",[826,10053,10054],{"id":10054},"我的目标标准",[758,10056,10057],{},"我采用的一个有用测试：",[758,10059,10060],{},"如果我消失一个月，我的团队还会交付吗？\n如果我消失一年，我的产品还会进化吗？\n如果我永远消失，会有什么东西继续创造价值吗？",[758,10062,10063],{},"专属知识帮助我今天做出贡献。但它也帮助我设计能在我之后存续的东西。",[753,10065,10067],{"id":10066},"_2-承担更多责任赢得结果的权利","2) 承担更多责任：赢得结果的权利",[758,10069,10070],{},"这可能是最\"简单但不容易\"的教训。",[758,10072,10073],{},"承担责任不意味着\"更努力工作\"。它意味着：",[804,10075,10076,10079,10082],{},[775,10077,10078],{},"拥有结果而不是任务，",[775,10080,10081],{},"成为会解决问题的人，",[775,10083,10084],{},"并接受如果失败了，责任在你。",[758,10086,10087],{},"这很可怕。它也改变了一切。",[758,10089,10090],{},"因为责任是创造信任的东西。信任创造自主权。自主权创造决策权。而决策权是杠杆开始的地方。",[826,10092,10093],{"id":10093},"责任是你在不请求许可的情况下获得杠杆的方式",[758,10095,10096],{},"很多人想要更多自由但避免责任。\n他们想要收益但不想要责备。\n他们想要影响力但不想要问责制。",[758,10098,10099],{},"但责任是交易。你不能在不付出这个代价的情况下获得好处。",[758,10101,10102,10103,10106],{},"当我回顾我职业生涯中每一次有意义的跳跃时，不是因为我在当前范围内表现良好。而是因为我",[778,10104,10105],{},"扩大了范围","，然后成长到其中。",[826,10108,10109],{"id":10109},"我试图攀登的责任阶梯",[758,10111,10112],{},"我试图有意识地向上移动这个阶梯：",[772,10114,10115,10120,10125,10130,10135],{},[775,10116,10117],{},[778,10118,10119],{},"把任务做好",[775,10121,10122],{},[778,10123,10124],{},"端到端拥有一个项目",[775,10126,10127],{},[778,10128,10129],{},"拥有一个系统（及其随时间的可靠性）",[775,10131,10132],{},[778,10133,10134],{},"拥有一个结果（业务指标、客户价值、战略执行）",[775,10136,10137],{},[778,10138,10139],{},"拥有一个投资组合（判断力最重要的地方）",[758,10141,10142],{},"在顶部，表现不如判断力重要。你因为正确而获得报酬，而不是忙碌。",[826,10144,10145],{"id":10145},"个人规则",[758,10147,10148],{},"我采用了一个感觉非常\"纳瓦尔\"的规则：",[758,10150,10151],{},[778,10152,10153],{},"不要只是交付工作。交付责任。",[758,10155,10156],{},"如果我要花费精力，我希望它增加我的所有权表面积，而不仅仅是填满我的一天。",[753,10158,10160],{"id":10159},"_3-构建资产创造无需你参与就能复利的东西","3) 构建资产：创造无需你参与就能复利的东西",[758,10162,10163],{},"这是最实用的收获，因为它改变了我看待\"工作\"的方式。",[758,10165,10166],{},"如果我用时间换金钱，我每天早上都会重置为零。\n如果我构建资产，我就创造了复利。",[758,10168,10169],{},"资产是可以重复产生价值而不需要每次都付出相同努力的东西。",[758,10171,10172],{},"那可以是：",[804,10174,10175,10178,10181,10184,10187,10190,10193,10196],{},[775,10176,10177],{},"企业中的股权，",[775,10179,10180],{},"运行并服务用户的代码，",[775,10182,10183],{},"吸引受众的内容，",[775,10185,10186],{},"数据集，",[775,10188,10189],{},"分发渠道，",[775,10191,10192],{},"品牌，",[775,10194,10195],{},"可重用的系统或平台，",[775,10197,10198],{},"甚至创造长期机会的关系。",[758,10200,10201,10202],{},"关键不是什么类型的资产。关键是：",[778,10203,10204],{},"它在我停止后还能继续工作吗？",[826,10206,10208],{"id":10207},"转变从项目到工厂","转变：从\"项目\"到\"工厂\"",[758,10210,10211],{},"我试图将我的思维方式从构建一次性项目转变为构建小型工厂。",[758,10213,10214],{},"工厂可以重复产生结果：",[804,10216,10217,10220,10223,10226],{},[775,10218,10219],{},"持续吸引机会的个人网站，",[775,10221,10222],{},"解决重复痛点的产品，",[775,10224,10225],{},"加速未来工作的库或工具，",[775,10227,10228],{},"使想法可被发现的内容引擎。",[758,10230,10231],{},"项目结束。工厂复利。",[758,10233,10234],{},"对我来说，这意味着：",[804,10236,10237,10243,10248,10254],{},[775,10238,10239,10242],{},[778,10240,10241],{},"这个博客"," - 吸引机会并建立声誉的内容",[775,10244,10245,10247],{},[778,10246,2819],{}," - 在我睡觉时为用户服务的产品",[775,10249,10250,10253],{},[778,10251,10252],{},"可重用的代码模式"," - 加速未来项目的系统",[775,10255,10256,10259],{},[778,10257,10258],{},"分发渠道"," - RSS、社交、放大影响力的网络",[826,10261,10262],{"id":10262},"资产需要杠杆",[758,10264,10265],{},"这就是三元组连接的地方：",[804,10267,10268,10271,10274],{},[775,10269,10270],{},"专属知识帮助你构建差异化的东西。",[775,10272,10273],{},"责任帮助你做决定并快速行动。",[775,10275,10276],{},"资产帮助你将产出扩展到你的工作时间之外。",[758,10278,10279],{},"这是一个飞轮。",[753,10281,10282],{"id":10282},"如何将其转化为行动",[758,10284,10285],{},"阅读很容易。整合很难。所以我写下了一个我实际上可以遵循的简单操作计划。",[826,10287,10289],{"id":10288},"步骤-1选择一个专属知识赛道今年深化","步骤 1：选择一个\"专属知识赛道\"今年深化",[758,10291,10292],{},"与其分散，我选择一个我可以复利的赛道：",[758,10294,10295],{},"一个赛道应该是：",[804,10297,10298,10301,10304],{},[775,10299,10300],{},"足够有趣，我会坚持下去，",[775,10302,10303],{},"足够有价值，市场会奖励它，",[775,10305,10306],{},"并且足够稀有，它不会立即商品化。",[758,10308,10309],{},"然后我问：这里的\"10,000 次重复\"是什么样子？",[826,10311,10313],{"id":10312},"步骤-2有意识地增加责任不燃尽","步骤 2：有意识地增加责任（不燃尽）",[758,10315,10316],{},"这不是对所有事情说是。这是对正确的事情说是。",[758,10318,10319],{},"我在寻找以下责任：",[804,10321,10322,10325,10328],{},[775,10323,10324],{},"增加我的决策权，",[775,10326,10327],{},"教我判断力，",[775,10329,10330],{},"并直接连接到结果。",[758,10332,10333],{},"如果某件事只增加我的忙碌，我想拒绝它——无论它看起来多么\"有成效\"。",[826,10335,10337],{"id":10336},"步骤-3每周将努力转化为资产","步骤 3：每周将努力转化为资产",[758,10339,10340],{},"我采用了一个每周问题：",[758,10342,10343],{},[778,10344,10345],{},"我这周构建了什么资产？",[758,10347,10348],{},"不是\"我做了什么\"。不是\"我参加了什么会议\"。\n我构建了什么会继续回报我的东西？",[758,10350,10351],{},"有时那个资产将是外部的（内容、产品）。有时是内部的（系统、模板、可重用组件）。但目标是继续将努力转化为复利价值。",[753,10353,10355],{"id":10354},"更深层的教训与长期伙伴玩长期游戏","更深层的教训：与长期伙伴玩长期游戏",[758,10357,10358],{},"在一切之下，我认为这本书指向一个生活策略：",[758,10360,10361],{},"选择长期游戏。\n与你信任的人一起玩。\n保持头脑清醒。\n道德地构建杠杆。\n让复利做繁重的工作。",[758,10363,10364],{},"\"财富\"部分只是一个输出。更大的胜利是这个框架也产生了更平静的生活：",[804,10366,10367,10370,10373],{},[775,10368,10369],{},"当你建立专属知识时，你停止追逐趋势。",[775,10371,10372],{},"当你承担责任时，你停止等待许可。",[775,10374,10375],{},"当你构建资产时，你停止感觉时间总是不够用。",[758,10377,10378],{},"你变得更难被替代——并且对被替代感到不那么焦虑。",[753,10380,10381],{"id":10381},"结语",[758,10383,10384],{},"如果我必须总结我从《纳瓦尔宝典》中带走的东西，那不是一句引语。而是一个方向：",[758,10386,10387],{},"我想成为那种可以：",[804,10389,10390,10393,10396],{},[775,10391,10392],{},"发展稀有、真实技能，",[775,10394,10395],{},"承担真正责任，",[775,10397,10398],{},"并构建复利资产的人。",[758,10400,10401],{},"不是作为一个奋斗故事。而是作为一个生活设计。",[758,10403,10404],{},"现在真正的工作开始了：长期、安静、持续地做这件事。",{"title":839,"searchDepth":708,"depth":708,"links":10406},[10407,10408,10413,10418,10422,10427,10428],{"id":9931,"depth":708,"text":9931},{"id":9966,"depth":708,"text":9967,"children":10409},[10410,10411,10412],{"id":9996,"depth":110,"text":9997},{"id":10012,"depth":110,"text":10013},{"id":10054,"depth":110,"text":10054},{"id":10066,"depth":708,"text":10067,"children":10414},[10415,10416,10417],{"id":10093,"depth":110,"text":10093},{"id":10109,"depth":110,"text":10109},{"id":10145,"depth":110,"text":10145},{"id":10159,"depth":708,"text":10160,"children":10419},[10420,10421],{"id":10207,"depth":110,"text":10208},{"id":10262,"depth":110,"text":10262},{"id":10282,"depth":708,"text":10282,"children":10423},[10424,10425,10426],{"id":10288,"depth":110,"text":10289},{"id":10312,"depth":110,"text":10313},{"id":10336,"depth":110,"text":10337},{"id":10354,"depth":708,"text":10355},{"id":10381,"depth":708,"text":10381},"建立专属知识，承担更多责任，构建资产。一个关于长期杠杆和复利价值的实用框架。",{"date":3010,"image":3011,"alt":10431,"tags":10432,"category":3016,"topics":10433,"published":859,"featured":859},"纳瓦尔宝典 - 专属知识、责任和资产",[3014,3015],[2135],"/blogs/zh/naval-ravikant-specific-knowledge-responsibility-assets",{"title":9909,"description":10429},"blogs/zh/12.naval-ravikant-specific-knowledge-responsibility-assets","E66RMbqbxuJZAax2iLJExX1EObMLWn4c3vLVLlyowxU",{"id":10439,"title":10440,"body":10441,"description":10582,"extension":2126,"meta":10583,"navigation":859,"ogImage":3178,"path":10587,"seo":10588,"stem":10589,"__hash__":10590},"zh/blogs/zh/13.start-my-ai-native-journey.md","开启我的 AI Native 之旅",{"type":750,"value":10442,"toc":10575},[10443,10446,10449,10454,10458,10461,10464,10469,10472,10475,10478,10481,10484,10487,10492,10496,10499,10505,10511,10517,10522,10525,10528,10531,10534,10537,10540,10560,10567,10570],[758,10444,10445],{},"我在科技行业已经待了不少年了——做产品、写代码、带团队。我见过不少\"下一个大趋势\"来了又去。但这一次感觉不一样。AI 不仅仅是工具箱里的又一个工具，它正在改变我思考的方式、构建的方式，说实话，也改变了我一天能完成多少事情。",[758,10447,10448],{},"所以我决定全力投入。这是我 AI Native 之旅的起点，我想分享一下我到目前为止学到的东西。",[758,10450,10451],{},[3036,10452],{"alt":10453,"src":3039},"AI Native 之旅启程 — 踏入充满可能的新世界",[753,10455,10457],{"id":10456},"什么是-ai-native","什么是 \"AI Native\"？",[758,10459,10460],{},"对我来说，AI Native 意味着 AI 不是附加品——它是起点。每次我坐下来工作，AI 就已经是我工作流程的一部分了。写代码？AI 就在旁边。研究一个新领域？AI 帮我更快地学习。起草产品规格？同样如此。",[758,10462,10463],{},"这就像从桌面端到移动优先的转变。一旦你走上了 Native 的路，就回不去了。",[758,10465,10466],{},[3036,10467],{"alt":10468,"src":3055},"AI Native 与传统方式 — 从附加品到起点",[753,10470,10471],{"id":10471},"顿悟的那一刻",[758,10473,10474],{},"我是一个仍然在写代码的产品负责人。我喜欢亲手构建东西——自己设计架构并实现一个功能，然后上线，那种满足感是无与伦比的。",[758,10476,10477],{},"但说实话：时间永远是瓶颈。在产品策略、团队协调和实际写代码之间，永远没有足够的时间。这就是 AI 改变游戏规则的地方。",[758,10479,10480],{},"几个月前，我在处理一个项目，正常情况下需要好几天的专注编码。我决定以 AI 优先的方式来做——不仅仅是用自动补全，而是在整个过程中真正与 AI 协作。架构决策、实现、调试、测试。",[758,10482,10483],{},"原本需要好几天的工作，几个小时就完成了。而且我说的不是那种马马虎虎、\"差不多就行\"的产出。质量是有保证的。代码是干净的。我有信心地把它上线了。",[758,10485,10486],{},"那就是我的顿悟时刻。我意识到这不是一个效率技巧——而是一个人能完成什么的根本性转变。",[758,10488,10489],{},[3036,10490],{"alt":10491,"src":3080},"突破时刻 — 从几天到几小时",[753,10493,10495],{"id":10494},"ai-实际上改变了我日常工作的什么","AI 实际上改变了我日常工作的什么",[758,10497,10498],{},"以下是 AI Native 在实践中的样子：",[758,10500,10501,10504],{},[778,10502,10503],{},"编码速度提升 2-3 倍。"," 我没有夸张。当我与 AI 结对时，我的实现速度更快了，因为我在样板代码、查找 API 和无聊的部分上花的时间更少了。我可以把精力集中在有趣的决策上——架构、用户体验、权衡取舍。",[758,10506,10507,10510],{},[778,10508,10509],{},"几小时学会新东西，而不是几周。"," 作为产品负责人，我需要了解广泛的技术和领域。AI 已经成为我首选的学习伙伴。我可以快速深入一个不熟悉的代码库、一个新框架或一个复杂的技术概念，学习速度大幅提升。就像有一个耐心、知识渊博的同事 24/7 全天候待命。",[758,10512,10513,10516],{},[778,10514,10515],{},"构建以前不敢尝试的东西。"," 这是最让我兴奋的部分。有些项目我以前会觉得\"一个人做太费劲了\"而放弃。现在我看着同样的项目会想——为什么不呢？构建的门槛已经大幅降低，限制因素不再是时间或技术能力，而是想象力。",[758,10518,10519],{},[3036,10520],{"alt":10521,"src":3111},"AI Native 的三大好处",[753,10523,10524],{"id":10524},"为什么要写这些",[758,10526,10527],{},"我写这篇文章不是为了说服谁 AI 是未来——如果你在读这篇文章，你可能已经感觉到了。我写这篇文章是因为我想记录下，作为一个每天都在做产品和写代码的人，走上 AI Native 之路到底是什么样子。",[758,10529,10530],{},"外面有很多炒作。也有很多质疑。但缺少的是那种诚实的、实用的、来自一线的视角。什么有用？什么没用？AI 在哪里真正帮你节省时间，又在哪里反而拖慢你？",[758,10532,10533],{},"这就是我想在这个系列中探讨的。",[753,10535,10536],{"id":10536},"接下来",[758,10538,10539],{},"这是我称之为 AI Native 之旅系列的第一篇。接下来，我计划分享：",[804,10541,10542,10548,10554],{},[775,10543,10544,10547],{},[778,10545,10546],{},"我用 AI 构建的真实项目"," — 好的、坏的、和丑的",[775,10549,10550,10553],{},[778,10551,10552],{},"真正有效的工作流程和工具"," — 不是理论，而是我每天在用的",[775,10555,10556,10559],{},[778,10557,10558],{},"经验教训"," — 包括犯过的错误，因为那才是真正学习发生的地方",[758,10561,10562,10563,10566],{},"我坚信我们正处于一场巨大变革的开端。AI 不仅仅让我们更快——它扩展了个人和小团队能做到的事情的边界。那些早早拥抱这个转变、学会与 AI ",[2156,10564,10565],{},"协作","的人，将拥有巨大的优势。",[758,10568,10569],{},"我决定全力投入。一起来吧。",[758,10571,10572],{},[3036,10573],{"alt":10574,"src":3167},"前方的旅程 — 向着日出前行",{"title":839,"searchDepth":708,"depth":708,"links":10576},[10577,10578,10579,10580,10581],{"id":10456,"depth":708,"text":10457},{"id":10471,"depth":708,"text":10471},{"id":10494,"depth":708,"text":10495},{"id":10524,"depth":708,"text":10524},{"id":10536,"depth":708,"text":10536},"我决定全力拥抱 AI Native。这篇文章分享了它的含义、我为什么这样做，以及作为一个仍在写代码的产品负责人，我到目前为止学到了什么。",{"date":3177,"image":3178,"alt":10584,"tags":10585,"category":2132,"topics":10586,"published":859,"featured":859},"AI Native 之旅 - 一个人开启 AI 旅程的手绘插图",[2130,2462],[2134,3182],"/blogs/zh/start-my-ai-native-journey",{"title":10440,"description":10582},"blogs/zh/13.start-my-ai-native-journey","T0lLEL9RjDqVYSQ8QLC75tHKUWUrK53CVmVdRdrgIU8",{"id":10592,"title":10593,"body":10594,"description":10807,"extension":2126,"meta":10808,"navigation":859,"ogImage":3414,"path":10812,"seo":10813,"stem":10814,"__hash__":10815},"zh/blogs/zh/14.marriott-timeshare-las-vegas.md","在拉斯维加斯的万豪分时度假推销中，我学到了关于销售文化崩坏的一课",{"type":750,"value":10595,"toc":10799},[10596,10599,10603,10606,10609,10629,10632,10637,10640,10643,10646,10649,10652,10657,10661,10664,10669,10695,10698,10703,10706,10709,10715,10721,10727,10732,10736,10739,10742,10745,10748,10751,10756,10759,10762,10765,10785,10792,10794],[758,10597,10598],{},"最近我和太太去了拉斯维加斯，参加了万豪假期俱乐部的推介会。我们知道这是一场分时度假的推销——他们对此很坦率。但我们没想到的是，这次经历会让我们整个住宿体验都蒙上阴影，也让我深入思考：当一家公司的销售文化与服务品牌背道而驰时，会发生什么。",[753,10600,10602],{"id":10601},"产品介绍积分费用和细则","产品介绍：积分、费用和细则",[758,10604,10605],{},"万豪的分时度假系统是这样运作的：你需要预付一大笔钱来购买\"积分\"，这些积分可以兑换全球万豪旗下酒店的住宿。听起来还算合理。",[758,10607,10608],{},"但复杂的地方在于：",[804,10610,10611,10617,10623],{},[775,10612,10613,10616],{},[778,10614,10615],{},"年度维护费。"," 即使你已经支付了预付费用，你每年仍然需要缴纳维护费——而且这些费用往往逐年递增。你在为使用自己已经买过的东西继续付费。",[775,10618,10619,10622],{},[778,10620,10621],{},"没有清晰的退出机制。"," 这是我最大的警觉信号。一旦加入，想要退出就变得非常模糊。没有简单直接的方式把积分卖回去或者干净利落地退出。",[775,10624,10625,10628],{},[778,10626,10627],{},"转让困难。"," 想把积分卖给别人？并不容易。分时度假的二手市场对卖家极为不利。",[758,10630,10631],{},"他们展示的签约套餐表面上看很诱人——额外积分、折扣房价、专属福利。但当我开始问\"如果我们想退出怎么办\"的时候，答案变得含糊其辞。这从来都不是好兆头。",[758,10633,10634],{},[3036,10635],{"alt":10636,"src":3233},"分时度假模式——钱进得来，却没有明确的出路",[753,10638,10639],{"id":10639},"一切改变的那一刻",[758,10641,10642],{},"我能接受强势推销。在商业领域待了这么久，我理解销售团队有业绩指标要完成。但我无法接受的是不尊重。",[758,10644,10645],{},"当我们告诉销售代表我们需要时间考虑一下是否购买时，他的反应既迅速又令人不适。他说 30 分钟\"足够\"我们夫妻俩商量了，催促我们当场做决定。",[758,10647,10648],{},"当我们坚持立场——我们想把资料带回去仔细考虑——他的态度完全变了。热情消失了，微笑不见了。他虽然没有直说，但明确表达了一个意思：我们在浪费他的时间。",[758,10650,10651],{},"这可是万豪的酒店。一个以热情好客和优质服务著称的品牌。万豪所代表的一切和我们在那个房间里经历的一切，形成了刺眼的反差。",[758,10653,10654],{},[3036,10655],{"alt":10656,"src":3255},"施压时刻——30分钟决定你的财务未来",[753,10658,10660],{"id":10659},"商业模式利润丰厚但代价是什么","商业模式：利润丰厚，但代价是什么？",[758,10662,10663],{},"暂且抛开情绪，来看看这个生意到底是怎么运作的。",[758,10665,10666],{},[778,10667,10668],{},"盈利引擎简单而高效：",[772,10670,10671,10677,10683,10689],{},[775,10672,10673,10676],{},[778,10674,10675],{},"预付销售"," 产生大额一次性收入。单笔分时度假购买的金额从 2 万美元到 10 万美元以上不等。",[775,10678,10679,10682],{},[778,10680,10681],{},"经常性维护费"," 创造稳定的年度收入流，且逐年增长——无论业主是否使用了积分。",[775,10684,10685,10688],{},[778,10686,10687],{},"融资贷款"," 带来利息收入。许多买家选择分期付款，意味着万豪还能从贷款中获利。",[775,10690,10691,10694],{},[778,10692,10693],{},"高退出门槛"," 锁住业主。没有清晰的退出路径，大多数业主年复一年地继续缴纳维护费——即使他们早已不再使用这个产品。",[758,10696,10697],{},"这是一个为获客和通过摩擦力留存而优化的商业模式，而非通过价值留存。这就是根本问题所在。",[758,10699,10700],{},[3036,10701],{"alt":10702,"src":3302},"分时度假商业模式的四大盈利引擎",[753,10704,10705],{"id":10705},"销售模式的问题",[758,10707,10708],{},"高压销售策略在分时度假行业并不新鲜。但它们造成了一系列具体的问题：",[758,10710,10711,10714],{},[778,10712,10713],{},"短期收益，长期损害。"," 一个逼迫夫妻当天做决定的销售人员可能拿下了这单。但这对夫妻会永远将这个品牌与负面体验联系在一起。在网络评价和社交媒体的时代，一次糟糕的体验传播得很远。",[758,10716,10717,10720],{},[778,10718,10719],{},"激励错位。"," 当销售人员的薪酬主要基于成交率和交易金额时，客户的最佳利益就退居其次了。30 分钟的施压策略不是为了给客户思考时间——而是为了阻止他们清醒地思考。",[758,10722,10723,10726],{},[778,10724,10725],{},"品牌侵蚀。"," 万豪花了几十年通过一致的服务品质建立信任。每一次高压分时度假推销都在一点点侵蚀这份信任。假期俱乐部部门可能创造了亮眼的营收，但付出的品牌代价是什么？",[758,10728,10729],{},[3036,10730],{"alt":10731,"src":3333},"短期销售胜利 vs 长期品牌损害",[753,10733,10735],{"id":10734},"更深层的思考当销售与服务背道而驰","更深层的思考：当销售与服务背道而驰",[758,10737,10738],{},"这次经历让我开始思考一个存在于许多服务型企业中的矛盾：服务品牌与销售文化之间的鸿沟。",[758,10740,10741],{},"万豪的酒店运营建立在一个简单的前提上——让客人感到受欢迎、舒适、被重视。这是有效的。我在万豪旗下的酒店住过很多次，通常都有很好的体验。",[758,10743,10744],{},"但走进他们的假期俱乐部销售间，你就进入了一个完全不同的世界。优先级翻转了。这里不再是让你感到被重视——而是要在你离开这栋楼之前把你签下来。",[758,10746,10747],{},"这不是万豪独有的问题。你在银行、保险、电信以及无数其他行业都能看到同样的现象。服务团队建立信任；销售团队消耗信任。",[758,10749,10750],{},"真正做对这一点的公司——那些真正赢得长期忠诚的公司——是那些销售与服务相统一的公司。在这些公司里，销售过程本身就是服务体验的延伸，而不是对服务体验的否定。",[758,10752,10753],{},[3036,10754],{"alt":10755,"src":3358},"服务文化 vs 销售文化——打破信任的鸿沟",[753,10757,10758],{"id":10758},"我的收获",[758,10760,10761],{},"我离开拉斯维加斯时没有购买分时度假。同时离开的，还有我对一个曾经尊重的品牌的好感。",[758,10763,10764],{},"这次经历中，有几件事我会一直记住：",[804,10766,10767,10773,10779],{},[775,10768,10769,10772],{},[778,10770,10771],{},"任何让你容易购买却难以离开的产品，都应该极度谨慎对待。"," 如果退出条款不清晰，说明这家公司赌的是你会因为惯性而留下，而非因为满意。",[775,10774,10775,10778],{},[778,10776,10777],{},"催促你快速决定，几乎从来都不是为了买家的利益。"," 一个好产品应该经得起一周的考虑。如果经不起，问问自己为什么。",[775,10780,10781,10784],{},[778,10782,10783],{},"一家公司的真正价值观，体现在最糟糕的时刻，而非最光鲜的时刻。"," 当客户说\"我需要考虑一下\"时，对方的回应揭示了一切。",[758,10786,10787,10788,10791],{},"对服务行业的从业者来说：你的销售体验",[2156,10789,10790],{},"就是","你的服务体验。客户不会把两者分开。每一个接触点，要么在建立信任，要么在摧毁信任，没有中间地带。",[2163,10793],{},[758,10795,10796],{},[2156,10797,10798],{},"你有过类似的分时度假推销经历吗？我很想听听你的看法。欢迎联系我，分享你的故事。",{"title":839,"searchDepth":708,"depth":708,"links":10800},[10801,10802,10803,10804,10805,10806],{"id":10601,"depth":708,"text":10602},{"id":10639,"depth":708,"text":10639},{"id":10659,"depth":708,"text":10660},{"id":10705,"depth":708,"text":10705},{"id":10734,"depth":708,"text":10735},{"id":10758,"depth":708,"text":10758},"当万豪假期俱乐部的销售策略与品牌所代表的一切背道而驰时，它揭示了销售文化与服务文化之间的深层矛盾。",{"date":3413,"image":3414,"alt":10809,"tags":10810,"category":3417,"topics":10811,"published":859},"一对夫妻站在酒店好客与分时度假销售压力的十字路口",[3014,3015],[3182],"/blogs/zh/marriott-timeshare-las-vegas",{"title":10593,"description":10807},"blogs/zh/14.marriott-timeshare-las-vegas","W_81HlCHor9h8PUS6KtlK0S5mF7TIq5fzb3h3AaTa8s",{"id":10817,"title":10818,"body":10819,"description":11116,"extension":2126,"meta":11117,"navigation":859,"ogImage":3755,"path":11121,"seo":11122,"stem":11123,"__hash__":11124},"zh/blogs/zh/15.why-i-stopped-writing-and-built-ai-studio.md","为什么我不再写博客了（转而搭建了一个 AI 内容工作室）",{"type":750,"value":10820,"toc":11107},[10821,10824,10827,10830,10833,10838,10841,10844,10847,10850,10888,10891,10896,10900,10906,10909,10923,10926,10929,10932,10937,10941,10944,10947,10953,10959,10963,10966,10970,10973,10979,10985,10991,10997,11003,11009,11015,11020,11024,11027,11030,11036,11042,11045,11048,11051,11054,11060,11063,11077,11080,11085,11088,11091,11094,11097,11100,11102],[758,10822,10823],{},"我尝试坚持写博客已经好几年了。每次都是一腔热血地开始，写两三篇，感觉良好——然后生活就插进来了。工作变忙，周末被各种事情填满。编辑器里那篇半成品的草稿开始变得陈旧。我告诉自己下周一定继续，结果下周变成了下个月，下个月变成了沉默。",[758,10825,10826],{},"如此循环。持续了好几年。",[758,10828,10829],{},"但上个月，我发布的博客文章数量比过去两年加起来还多。不是因为我突然多出了时间，也不是因为发现了什么效率秘诀。我不再试图去\"写\"博客了，而是开始\"搭建一个系统\"，让它把我的想法变成发布的内容。",[758,10831,10832],{},"这是一个关于我如何从\"反复放弃的博客写手\"变成\"稳定输出的创作者\"的故事——靠的是用工程师的思维来解决问题，而不是靠写作的毅力。",[758,10834,10835],{},[3036,10836],{"alt":10837,"src":3453},"被遗弃的博客们——那些永远没有发布的草稿",[753,10839,10840],{"id":10840},"我反复放弃的真正原因",[758,10842,10843],{},"说实话吧。问题不在于缺少想法。想法到处都是——在拉斯维加斯遭遇的一次令人恼火的分时度假推销、关于 AI 如何改变我日常工作的思考、对产品策略的独到见解。我的备忘录里堆满了各种半成型的想法，都在等着变成文章。",[758,10845,10846],{},"问题出在想法\"之后\"的所有事情上。",[758,10848,10849],{},"想想发布一篇博客文章到底需要做多少事情：",[804,10851,10852,10858,10864,10870,10876,10882],{},[775,10853,10854,10857],{},[778,10855,10856],{},"写初稿"," —— 顺利的话 1 到 2 小时，不顺利的话更久",[775,10859,10860,10863],{},[778,10861,10862],{},"排版和编辑"," —— 整理 Markdown 格式、添加标题、让它变得好读（30 分钟）",[775,10865,10866,10869],{},[778,10867,10868],{},"制作图片"," —— 封面图、流程图、配图，不能只用库存图片（至少 1 小时）",[775,10871,10872,10875],{},[778,10873,10874],{},"翻译成中文"," —— 我面向双语读者写作，每篇都需要自然流畅的中文版本，不是 Google 翻译那种（1 到 2 小时）",[775,10877,10878,10881],{},[778,10879,10880],{},"发布上线"," —— frontmatter 元数据、图片路径、SEO 信息、部署到网站（30 分钟）",[775,10883,10884,10887],{},[778,10885,10886],{},"社交媒体"," —— 写 X 推文、LinkedIn 帖子，或许还有 YouTube 脚本（又要 30 分钟）",[758,10889,10890],{},"加起来算一下：一篇博客文章需要 5 到 6 小时的投入。作为一个还在写代码的产品负责人，有全职工作和副项目，这些时间根本不可能持续存在。数学上就算不过来。所以我只能在灵感和空闲时间碰巧重合时才写——而这几乎从不发生。",[758,10892,10893],{},[3036,10894],{"alt":10895,"src":3517},"一篇博客的隐藏成本——从想法到发布需要 6 小时",[753,10897,10899],{"id":10898},"转变如果我只需要做思考的部分呢","转变：如果我只需要做\"思考\"的部分呢？",[758,10901,10902,10903],{},"突破点不在于写得更快。而是一个完全不同的问题：",[778,10904,10905],{},"如果我只需要做那些需要动脑子的部分呢？",[758,10907,10908],{},"那个 6 小时流程中的每一步都可以归为两类：",[772,10910,10911,10917],{},[775,10912,10913,10916],{},[778,10914,10915],{},"思考"," —— 决定说什么、选什么角度、分享什么个人经验、让读者得到什么",[775,10918,10919,10922],{},[778,10920,10921],{},"执行"," —— 把这些决定变成格式化的文本、图片、翻译、元数据和已部署的页面",[758,10924,10925],{},"思考只有我能做。但执行呢？那是流程。是可重复的。是可以自动化的。",[758,10927,10928],{},"这正是大多数\"AI 写博客\"建议出错的地方。他们告诉你用 ChatGPT 来写文章。那样产出的是千篇一律的内容——AI 口水文。没人想读，我更不想发。",[758,10930,10931],{},"我不想让 AI 替我思考。我想让 AI 处理思考\"之外\"的一切。",[758,10933,10934],{},[3036,10935],{"alt":10936,"src":3563},"思考 vs. 执行——唯一需要人类的部分",[753,10938,10940],{"id":10939},"我搭建了什么一个-ai-内容工作室","我搭建了什么：一个 AI 内容工作室",[758,10942,10943],{},"于是我就真的搭了一个。我称它为我的内容工作室——一组 AI 驱动的技能，它们像链条一样串联在一起，把一个粗糙的想法变成一篇完整发布的、双语的博客文章，配有插图、社交媒体内容，甚至还有视频。",[758,10945,10946],{},"在我的系统里，一篇博客文章长这样：",[834,10948,10951],{"className":10949,"code":10950,"language":3579},[3577],"src/blogs/2026-02-14/\n├── plan.md                        ← 我的粗略想法（唯一真正需要我写的）\n├── marriott-timeshare-las-vegas.md     ← 英文博客\n├── marriott-timeshare-las-vegas-zh.md  ← 中文翻译\n├── youtube-script.md              ← 旁白脚本\n├── x-teaser.md                    ← X/Twitter 推文\n├── linkedin-post.md               ← LinkedIn 版本\n├── video.mp4                      ← YouTube 成品视频\n└── imgs/\n    ├── 00-cover.png               ← 封面图\n    ├── 01-infographic-*.png       ← 插图\n    ├── 02-scene-*.png             ← 插图\n    └── ... (5 张定制插图)\n",[841,10952,10950],{"__ignoreMap":839},[758,10954,10955,10956,10958],{},"一个输入，八个输出。而输入是什么？就是一个 ",[841,10957,3587],{},"，像这样——这是我写 Marriott 分时度假那篇文章的实际计划：",[3590,10960,10961],{},[758,10962,3594],{},[758,10964,10965],{},"就是这样。一段中英混合的想法记录。没有格式，没有结构，只有原始想法和我希望文章探讨的方向。之后的事情，流水线全包了。",[753,10967,10969],{"id":10968},"流水线它到底是怎么运作的","流水线：它到底是怎么运作的",[758,10971,10972],{},"每一步都由一个专门的 AI 技能来处理，每个技能都清楚地知道该做什么：",[758,10974,10975,10978],{},[778,10976,10977],{},"第 1 步：我写计划。"," 这是创意的部分——记录我想说什么、选什么角度、以什么个人经历为核心。有时候是一段话，有时候是几个要点，有时候是语音备忘录的转录文字。格式不重要，思考才重要。",[758,10980,10981,10984],{},[778,10982,10983],{},"第 2 步：AI 起草博客文章。"," 用我的计划作为指引，AI 写出完整的文章。但关键是——我会审阅一切。我调整语气，加入只有我才知道的细节，删掉感觉太泛的部分。草稿让我完成了 80%；我把它变成 100% 属于我的。",[758,10986,10987,10990],{},[778,10988,10989],{},"第 3 步：AI 生成插图。"," 我的文章插图技能会分析文章，找出哪里需要配图，然后用统一的视觉风格生成图片。不是库存图片——而是匹配内容的定制插图。数据用信息图，叙事用场景图，分析用对比图。",[758,10992,10993,10996],{},[778,10994,10995],{},"第 4 步：AI 翻译成中文。"," 不是逐字翻译——而是自然、流畅的简体中文，读起来像母语人士写的。技术术语在合适的地方保留英文。产出的是一篇真正的中文文章，而不是翻译腔的英文。",[758,10998,10999,11002],{},[778,11000,11001],{},"第 5 步：AI 创建分发内容。"," 一条吸引 Twitter 受众的 X 推文，一篇面向职场人士的 LinkedIn 帖子，一个为旁白设计的 YouTube 脚本。每一个都按平台特点定制，而不是简单复制粘贴。",[758,11004,11005,11008],{},[778,11006,11007],{},"第 6 步：AI 生成视频。"," YouTube 脚本变成带有过渡效果、背景音乐和合理节奏的旁白幻灯片视频，可以直接上传。",[758,11010,11011,11014],{},[778,11012,11013],{},"第 7 步：AI 发布。"," 生成 frontmatter、重写图片路径、按序编号、部署到我的 Nuxt 3 博客网站。英文和中文版本同时发布，互相关联为翻译版本。",[758,11016,11017],{},[3036,11018],{"alt":11019,"src":3652},"内容流水线——从粗略计划到 8 个已发布的产出",[753,11021,11023],{"id":11022},"为什么这不是-ai-口水文","为什么这不是 AI 口水文",[758,11025,11026],{},"我知道你可能在想什么。\"这不就是 AI 生成的内容加了几个步骤吗？\"",[758,11028,11029],{},"不是。这个区别很重要。",[758,11031,11032,11035],{},[778,11033,11034],{},"AI 生成的内容","从一个提示词开始，比如\"帮我写一篇关于分时度假的博客\"。AI 凭空创造视角、案例和观点。读起来像是所有人写的，又像是没人写的。",[758,11037,11038,11041],{},[778,11039,11040],{},"AI 制作的内容","从我的经历、我的分析、我的视角开始。那篇 Marriott 文章之所以存在，是因为我真的坐在那里经历了那次推销，而且对销售手法感到由衷的愤怒。AI Native Journey 那篇之所以存在，是因为我真的改变了工作方式，想要记录我学到的东西。任何 AI 都无法发明这些经历。",[758,11043,11044],{},"AI 负责制作——写作技巧、排版、翻译、图片。但每篇文章都始于真实的东西。这就是为什么产出不会读起来像泛泛之谈。想法有锋芒，案例是具体的，观点是我自己的。",[758,11046,11047],{},"结果就是证明：我第一次在几年里实现了持续发布。我那篇 Marriott 文章从一段沮丧的中文碎碎念开始，变成了一篇双语的分时度假商业模式分析，配有定制插图。这种从原始经历到精致多格式内容的转变——正是这条流水线所实现的。",[753,11049,11050],{"id":11050},"这对你意味着什么",[758,11052,11053],{},"我不是说每个人都需要搭建和我一模一样的系统。但我认为这里有一个更广泛的启示，适用于任何在持续创作上苦苦挣扎的人。",[758,11055,11056,11059],{},[778,11057,11058],{},"2026 年，成为一个持续创作者的障碍不是才华或想法——而是执行开销。"," 从你的想法到你的受众之间的每一个手动步骤，都是一个你可能放弃的摩擦点。步骤越多，你越可能在半路上放弃。",[758,11061,11062],{},"AI 工具已经成熟到可以自动化大部分执行开销的程度。你不需要在第一天就搭建一整套内容工作室。从那个扼杀你动力的步骤开始：",[804,11064,11065,11068,11071,11074],{},[775,11066,11067],{},"如果翻译是你的瓶颈，先自动化翻译",[775,11069,11070],{},"如果制作图片阻碍了你发布，用 AI 插图",[775,11072,11073],{},"如果发布流程很繁琐，用脚本搞定",[775,11075,11076],{},"如果为社交媒体重新编排内容让你觉得很烦，让 AI 起草初版",[758,11078,11079],{},"你从手动工作中移除的每一步，都让坚持变得更可持续。而坚持会产生复利效应——更多产出意味着更多受众、更多反馈、更多继续前进的动力。",[758,11081,11082],{},[3036,11083],{"alt":11084,"src":3719},"从小处开始——一次移除一个摩擦点",[753,11086,11087],{"id":11087},"诚实的真相",[758,11089,11090],{},"我没有变成一个更好的写作者。我的想法并不比两年前失败的时候更敏锐。改变的是，我不再把写博客当作一个写作问题，而是开始把它当作一个系统问题来对待。",[758,11092,11093],{},"你正在读的这篇博客，就是通过这条完全相同的流水线生产的。一个粗略的计划、一篇草稿、插图、翻译、发布——全部由我搭建的 AI 内容工作室处理。我的工作就是思考。系统处理了剩下的一切。",[758,11095,11096],{},"如果你也是那种有很多好想法但总是难以持续输出的人，我建议你用同样的方式审视自己的工作流程。找到那个扼杀你动力的步骤。那就是你的起点。",[758,11098,11099],{},"工具已经存在了。问题是你是否愿意去搭建这个系统。",[2163,11101],{},[758,11103,11104],{},[2156,11105,11106],{},"这是我 AI Native Journey 系列的一部分，在这个系列中我分享 AI 如何改变我构建、创造和工作的方式。如果你也在搭建自己的 AI 驱动工作流程，我很想听听你有什么心得。",{"title":839,"searchDepth":708,"depth":708,"links":11108},[11109,11110,11111,11112,11113,11114,11115],{"id":10840,"depth":708,"text":10840},{"id":10898,"depth":708,"text":10899},{"id":10939,"depth":708,"text":10940},{"id":10968,"depth":708,"text":10969},{"id":11022,"depth":708,"text":11023},{"id":11050,"depth":708,"text":11050},{"id":11087,"depth":708,"text":11087},"我尝试坚持写博客好几年，每次都半途而废。后来我搭建了一个 AI 内容工作室，能在几分钟内把一个粗略的想法变成已发布的双语博客文章，配有插图、社交媒体内容和视频。",{"date":3754,"image":3755,"alt":11118,"tags":11119,"category":3758,"topics":11120,"published":859},"AI 内容工作室流水线——将一个粗略想法转化为多种精致产出",[2130,2462],[2134,3182],"/blogs/zh/why-i-stopped-writing-and-built-ai-studio",{"title":10818,"description":11116},"blogs/zh/15.why-i-stopped-writing-and-built-ai-studio","lNL9ATFH6GOaLD_vPwVO5YEun0irkY9t1k0zTEjBADA",{"id":11126,"title":11127,"body":11128,"description":11323,"extension":2126,"meta":11324,"navigation":859,"ogImage":3967,"path":11331,"seo":11332,"stem":11333,"__hash__":11334},"zh/blogs/zh/16.mac-mini-ai-agent-content-team.md","我有一台吃灰的 Mac Mini，现在它撑起了我整个内容运营。",{"type":750,"value":11129,"toc":11317},[11130,11133,11137,11140,11143,11179,11182,11187,11191,11194,11199,11202,11205,11208,11213,11216,11219,11222,11226,11229,11232,11238,11244,11250,11256,11262,11265,11268,11271,11274,11277,11282,11285,11288,11291,11294,11297,11300,11303,11306,11311],[758,11131,11132],{},"我有一台 Mac mini 一直放在架子上吃灰。上周六，我在上面装了 OpenClaw，连上了 WhatsApp，给这个 AI 智能体取了个名字 —— GG。48 小时之内，GG 为我的内容事业做的事，比我上个月做的还多。这就是 AI 原生工作流真实的样子 —— 没有炒作，只有一台吃灰的机器和一个愿意尝试新事物的 builder。",[753,11134,11136],{"id":11135},"安装配置周六上午","安装配置（周六上午）",[758,11138,11139],{},"这台 Mac mini 已经闲置好几个月了。当时买来是为一个项目准备的，但那个项目最终没做成。一月份 OpenClaw 开始爆火 —— 甚至直接导致了 Mac mini 缺货 —— 我才意识到，大家在抢购的硬件，我手边就有一台。",[758,11141,11142],{},"整个配置大约花了两小时。以下是我实际做的事情：",[772,11144,11145,11151,11161,11167,11173],{},[775,11146,11147,11150],{},[778,11148,11149],{},"让 Mac mini 永不休眠。"," 系统设置 > 节能 > 防止自动进入睡眠。听起来很基础，但如果你跳过这步，你的智能体凌晨两点就掉线了。",[775,11152,11153,11156,11157,11160],{},[778,11154,11155],{},"把 OpenClaw 网关安装为系统服务。"," 在终端运行 ",[841,11158,11159],{},"openclaw onboard","。安装向导会引导你完成所有步骤 —— 工作空间、频道、技能配置。对于一个开源工具来说，体验出奇地流畅。",[775,11162,11163,11166],{},[778,11164,11165],{},"把 WhatsApp 设为主要通信频道。"," 这是一个关键决策。我希望和 AI 智能体的沟通方式，跟和人聊天一样 —— 通过我每天都在用的即时通讯软件。不需要专门的界面，不需要保持一个浏览器标签页。",[775,11168,11169,11172],{},[778,11170,11171],{},"设置 Tailscale 远程访问。"," 这会在我的 MacBook Pro（日常主力机）和 Mac mini 之间创建一个私有网络。现在我可以从任何地方通过 SSH 或屏幕共享连接 —— 沙发上、咖啡馆里，随时随地。",[775,11174,11175,11178],{},[778,11176,11177],{},"给智能体取名 GG。"," 这听起来可能有点傻，但其实很重要。当你给智能体一个身份，你就开始把它当作协作者，而不是搜索框。整个互动的氛围就变了。",[758,11180,11181],{},"两小时之内，我就拥有了一个 24/7 运行的 AI 智能体，它活在我的家庭网络上，通过 WhatsApp 联系，能浏览网页、读写文件、运行 shell 命令，并且记住所有跨会话的内容。",[758,11183,11184],{},[3036,11185],{"alt":11186,"src":3822},"永不停歇的 AI 中心：Mac mini 作为互联生态的核心",[753,11188,11190],{"id":11189},"第一天从启动到产出价值","第一天：从启动到产出价值",[758,11192,11193],{},"我做的第一件事是设置定时任务 —— 晨间简报、每日内容研究和晚间汇报。我希望 GG 能主动为我工作，而不只是被动地等我提问。",[758,11195,11196],{},[778,11197,11198],{},"晨间简报失败了。",[758,11200,11201],{},"gog CLI（GG 用来检查我的邮件和日历的工具）还没有完成认证。所以定时任务跑起来了，碰到了认证错误，什么都没产出。我醒来看到的是一份空白简报。",[758,11203,11204],{},"我分享这件事是因为它很重要。AI 工作流会出错。就像其他任何软件一样会出错。晨间简报失败是因为我漏掉了一个配置步骤。修复只花了五分钟 —— 认证 CLI、手动测试、确认正常、重新设置定时任务。搞定。",[758,11206,11207],{},"\"AI 炒作\"和\"AI 工作流\"的区别就在这里：出了问题不慌张，修好它然后继续。",[758,11209,11210],{},[3036,11211],{"alt":11212,"src":3849},"Builder 的迭代循环：出错、修复、发布、重复",[758,11214,11215],{},"修好简报之后，我给 GG 布置了第一个真正的任务：了解我的内容工作室工作流。我有一套串联的 AI 技能链 —— 头脑风暴、插图生成、视频制作、发布 —— 各个工具之间互相对话。我想让 GG 理解整个流程。",[758,11217,11218],{},"那就是第一个\"顿悟时刻\"。GG 不只是读文件。它浏览了我的博客，分析了技能定义，把各个工具之间的关系串联起来，然后给出了一份完整的流程总结 —— 还包括一些我没想到的改进建议。",[758,11220,11221],{},"这不是一个聊天机器人在重复我说的话。GG 有上下文。它能看我的文件，读我的代码，理解各个系统之间的关系。而且它把所有这些都记住了，留待后用。",[753,11223,11225],{"id":11224},"第二天复利效应","第二天：复利效应",[758,11227,11228],{},"第二天，一切都变了。",[758,11230,11231],{},"我醒来看到晨间简报正常运行了（第一天的修复成功了），然后 GG 和我开始了高产的一天。以下是我们一天内完成的事情：",[758,11233,11234,11237],{},[778,11235,11236],{},"深度博客分析。"," GG 爬取了 aaronguo.com，带回了具体的改进建议 —— 不是泛泛的\"多加关键词\"那种建议，而是关于导航结构、内容缺口和定位清晰度的深入反馈。它读了每一篇文章。",[758,11239,11240,11243],{},[778,11241,11242],{},"向最好的学习。"," 我让 GG 分析了五位我欣赏的创作者：Justin Welsh、Pieter Levels、Lenny Rachitsky、Sahil Bloom 和 Dickie Bush。GG 不只是总结他们的简介。它分析了他们的内容策略 —— 发布频率、话题模式、变现模型、受众互动策略 —— 并告诉我能从每个人身上学到什么。从 Justin Welsh 那里学到了单一想法 thread 的简洁力量。从 Pieter Levels 那里学到了用收入仪表盘实现极致透明。从 Sahil Bloom 那里学到了 newsletter 优先的品牌策略。",[758,11245,11246,11249],{},[778,11247,11248],{},"Newsletter 搭建。"," GG 研究了各个 newsletter 平台，推荐了 Beehiiv，并帮我一步步搭建了\"Ship with AI\" —— 你现在可能正在阅读的这份 newsletter。然后它还研究了如何将 Beehiiv 和我的 Nuxt 3 博客集成，让订阅者获得完整体验。",[758,11251,11252,11255],{},[778,11253,11254],{},"X/Twitter 策略。"," 这是最重要的一项。GG 深入研究了 X 的算法 —— 回复的权重是点赞的 27 倍、外部链接会被降低约 50% 的曝光、哪些内容格式获得最多触达。然后它把所有内容整合成一套完整的内容策略：每周发布节奏、内容支柱、开头公式、回复互动策略，以及每篇博文的发布时间线。",[758,11257,11258,11261],{},[778,11259,11260],{},"首页设计评审。"," GG 打开了我的网站，分析了布局，并给出了具体的设计建议，每条建议都有理由支撑。",[758,11263,11264],{},"所有这些都在一天之内完成。",[758,11266,11267],{},"但真正重要的是：这些任务不是孤立存在的。博客分析为内容策略提供了信息。创作者研究塑造了定位。Newsletter 搭建由策略驱动。每个任务都建立在前一个任务的上下文之上。",[758,11269,11270],{},"到第二天结束时，GG 已经知道我的定位（\"Ship with AI, not about AI\"）、我的内容支柱（AI 原生执行力、产品领导力、公开构建）、我的目标受众、我的语调风格，以及我正在做的具体项目。它不再是从零开始了。",[758,11272,11273],{},"OpenClaw 做得特别好的就是管理这些上下文。GG 维护着结构化的记忆文件 —— 每日笔记和长期记忆 —— 并把它们关联到它做的每一件事：定时任务、文件访问、shell 命令、网页浏览。上下文不只是\"被记住\"—— 而是在每个任务中被积极使用。当 GG 上午研究顶级创作者、下午制定内容策略时，不需要我来串联这些信息。记忆系统自动完成了这件事。",[758,11275,11276],{},"这就是复利效应。也是这套配置真正强大的原因。",[758,11278,11279],{},[3036,11280],{"alt":11281,"src":3919},"复利效应：每个任务都建立在前一个任务的上下文之上",[753,11283,11284],{"id":11284},"我真正学到了什么",[758,11286,11287],{},"这个转变不是关于 OpenClaw 本身，也不是关于 Mac mini，更不是关于任何单一工具。而是关于从\"把 AI 当搜索引擎\"转变为\"把 AI 当持久伙伴\"。",[758,11289,11290],{},"当你在浏览器标签页里使用 AI，你在做一次性查询。得到一个回复，也许有用，然后继续。没有记忆，没有上下文，没有复利。",[758,11292,11293],{},"当你搭建了一个持久化的智能体 —— 一个活在你网络上、能访问你的文件、记住你的对话、在你睡觉时运行定时任务、并随时间积累上下文的智能体 —— 一些根本性的不同就会发生。智能体不再是工具，而开始成为队友。",[758,11295,11296],{},"48 小时之后，GG 对我的工作有了足够的上下文，不再问显而易见的问题。它不用提醒就知道我的目标。它引用之前对话中的决定。它开始串联起我自己都没有串联的信息。",[758,11298,11299],{},"而现在，在我写这篇博文的时候？GG 参与了它的规划。内容计划、分发策略、X thread 大纲 —— 这些都是协作的成果。这不是\"AI 生成的内容\"。这是 AI 增强的构建。我做决策。GG 做重活。",[758,11301,11302],{},"如果你有一台 Mac mini —— 或者任何能永远在线的机器 —— 尝试这一切的门槛就是一个下午。两小时的配置，你就拥有了一个永不休息、永不丢失上下文、每天都在变好的伙伴。",[758,11304,11305],{},"最难的部分不是配置。而是改变你的心智模型：从\"AI 是我使用的工具\"变成\"AI 是我合作的伙伴\"。",[758,11307,11308],{},[3036,11309],{"alt":11310,"src":3950},"从 AI 作为查询工具到 AI 作为工作伙伴的转变",[758,11312,11313,11314,11316],{},"我正在记录这段旅程的全部 —— 成功、失败、真实数据 —— 在我的 newsletter ",[778,11315,3956],{}," 中。如果你想跟随我用 AI 伙伴公开构建的旅程，请在下方订阅。",{"title":839,"searchDepth":708,"depth":708,"links":11318},[11319,11320,11321,11322],{"id":11135,"depth":708,"text":11136},{"id":11189,"depth":708,"text":11190},{"id":11224,"depth":708,"text":11225},{"id":11284,"depth":708,"text":11284},"我在一台 Mac mini 上安装了 OpenClaw，给 AI 智能体取名 GG，48 小时内它分析了我的博客、研究了顶级创作者的策略、搭建了 newsletter、制定了完整的 X 内容策略。",{"date":3966,"image":3967,"alt":11325,"tags":11326,"category":3758,"published":859},"Mac mini 和 OpenClaw 龙虾吉祥物 —— 永不停歇的 AI 内容中心",[2130,11327,11328,11329,11330],"OpenClaw","Mac Mini","效率","公开构建","/blogs/zh/mac-mini-ai-agent-content-team",{"title":11127,"description":11323},"blogs/zh/16.mac-mini-ai-agent-content-team","ghOhWsgJYry1k3vyqFpMl6rNYOhlZOZxXE8L0Bem3L4",{"id":11336,"title":11337,"body":11338,"description":11617,"extension":2126,"meta":11618,"navigation":859,"ogImage":4276,"path":11626,"seo":11627,"stem":11628,"__hash__":11629},"zh/blogs/zh/17.mckinsey-wealth-management-2035-ai-insider-take.md","麦肯锡说 2035 年。我说他们预测晚了 5 年。",{"type":750,"value":11339,"toc":11609},[11340,11343,11346,11349,11353,11356,11394,11397,11405,11410,11413,11419,11425,11431,11434,11437,11440,11445,11448,11451,11454,11457,11462,11465,11468,11471,11476,11479,11482,11487,11490,11493,11498,11509,11512,11515,11518,11523,11526,11533,11536,11541,11544,11547,11550,11555,11558,11565,11571,11577,11583,11587,11590,11593,11596,11599,11602,11604],[758,11341,11342],{},"麦肯锡刚刚发布了一份 10 页的报告，预测 AI、人口结构变化和信任重塑将如何在 2035 年之前重塑美国财富管理行业。这是一份扎实的报告——研究严谨，措辞谨慎，顾问风格一流。数据是真实的，趋势无可争辩，他们提出的六大主题方向上也没有问题。",[758,11344,11345],{},"但我读这份报告的角度，和大多数人不一样。因为我不是顾问。我是一家金融机构的技术产品负责人（是技术产品，不是金融产品），同时每天还在写代码，业余时间在搭建 AI 原生系统。他们描述的那个未来？一部分已经来了。剩下的部分也不是 10 年后的事——是 5 年后的事。",[758,11347,11348],{},"顾问的时间线和建造者的现实之间的落差，才是这个故事真正的主角。那些按 2035 年规划的公司，会被那些按 2030 年建造的公司颠覆。",[753,11350,11352],{"id":11351},"报告摘要60-秒读懂麦肯锡","报告摘要：60 秒读懂麦肯锡",[758,11354,11355],{},"麦肯锡识别出将定义 2035 年美国财富管理行业的六大主题：",[772,11357,11358,11364,11370,11376,11382,11388],{},[775,11359,11360,11363],{},[778,11361,11362],{},"人口结构变化"," —— 22 万亿美元的财富向 X 世代和千禧一代转移，女性掌控 40% 以上的财富，预计出现 10 万名顾问缺口",[775,11365,11366,11369],{},[778,11367,11368],{},"Agentic AI（自主 AI）"," —— 从任务自动化向能够自主推理、规划和行动的 AI 系统演进",[775,11371,11372,11375],{},[778,11373,11374],{},"资产类别扩张"," —— 私募市场、实物资产以及数字/代币化资产走向主流",[775,11377,11378,11381],{},[778,11379,11380],{},"通过透明度重建信任"," —— 机构声誉让位于实际表现和彻底的开放性",[775,11383,11384,11387],{},[778,11385,11386],{},"每位客户都成为家族办公室"," —— 超越投资，向税务、遗产和生活方式的全面\"人生管理\"拓展",[775,11389,11390,11393],{},[778,11391,11392],{},"顾问从规划师进化为人生教练"," —— 人的角色转向关系和引导，而非信息传递",[758,11395,11396],{},"麦肯锡同时提出了四种将主导市场的竞争原型：凭借规模取胜的超级平台、凭借关系取胜的精品机构、凭借灵活性取胜的独立平台，以及凭借效率取胜的 AI 原生公司。",[758,11398,11399,11400,11404],{},"完整报告值得一读——",[2177,11401,11403],{"href":4044,"rel":11402},[2181],"报告链接","。但以下是我这个行业内部人的解读。",[758,11406,11407],{},[3036,11408],{"alt":11409,"src":4053},"麦肯锡财富管理六大主题——地图是对的，但时间线不对",[753,11411,11412],{"id":11412},"麦肯锡说对了的地方",[758,11414,11415,11418],{},[778,11416,11417],{},"顾问短缺是真实且紧迫的。"," 这不需要麦肯锡的数据来证明——我每天在自己的公司都能看到。有经验的顾问陆续退休，新入职的人需要好几年才能真正上手，而客户需求增长的速度远超人才供给。麦肯锡预测的 10 万顾问缺口可能还是保守估计。不引入 AI 辅助，这道数学题根本算不过来，而大多数公司心里都清楚。",[758,11420,11421,11424],{},[778,11422,11423],{},"信任的转变已经在年轻客户身上发生了。"," 麦肯锡引用数据：76% 的 Z 世代在网上寻求理财建议，而只有 14% 会首先去找专业的财务顾问（婴儿潮一代是 39%）。这与我从自己公司产品数据中看到的完全吻合。年轻客户不想被有资历的人告知该怎么做——他们想要的是实时、透明的\"让我看到\"。\"说给我听不如做给我看\"这种互动方式，已经渗透到现在每一次客户沟通中了。",[758,11426,11427,11430],{},[778,11428,11429],{},"\"人人都有家族办公室\"的趋势正在发生。"," 麦肯锡的数据显示，2023 年有 52% 的投资者寻求综合性建议，2018 年这个数字只有 29%。客户现在明确要求理财顾问提供的不只是投资组合管理——他们想要税务策略、遗产规划、保险，甚至生活方式决策。而在较低财富层级实现这一切的技术正在到来。我们现在就在将它纳入产品路线图。",[758,11432,11433],{},"在这三点上，麦肯锡不只是说对了——他们甚至低估了紧迫程度。",[753,11435,11436],{"id":11436},"麦肯锡太保守的地方",[758,11438,11439],{},"这里是我的异议所在。",[758,11441,11442],{},[778,11443,11444],{},"Agentic AI 不是 2035 年的故事——是 2026 年的故事。",[758,11446,11447],{},"麦肯锡把 Agentic AI 定性为一种正在涌现的未来能力。与此同时，我上周在一台 Mac mini 上部署了一个 AI 代理。它全天候运行，做深度研究、管理内容流水线、每天早上给我发简报，还在我睡觉时自主执行定时任务。昨天我还通过 WhatsApp 和它聊了聊我正在规划的一篇博客文章。",[758,11449,11450],{},"这不是实验室里的原型，是一台积了灰的 Mac mini 加上每月 15 美元的 API 费用，跑着我内容业务的生产工作流。",[758,11452,11453],{},"在我的本职工作中，我们已经在用 AI 来增强产品工作流程、研究和报告。麦肯锡 2024 年数据里那 62% \"打算使用 AI\"的顾问——其中真正在行动的那部分人，已经不是\"打算\"了。他们在交付了。",[758,11455,11456],{},"以前需要研究团队好几天才能完成的工作，现在我的 AI 代理几个小时搞定。这不是 2035 年的事。这是今天正在发生的事。",[758,11458,11459],{},[778,11460,11461],{},"四种竞争原型低估了 AI 原生的破坏力。",[758,11463,11464],{},"麦肯锡的四种竞争分类（超级平台、精品机构、独立平台、AI 原生公司）把\"AI 原生\"当成一个独立的细分市场。我认为这个框架是错的。",[758,11466,11467],{},"麦肯锡其实已经暗示了正确答案——他们提出了\"独立顾问公司\"（firm of one）的概念：一位顾问借助数百个 AI 代理，服务原本需要一个十人团队才能承接的客户量。但他们没有把这条线索追到底：当一个人能做到十个人的工作，每一种模式的经济逻辑都会被打破。不只是 AI 原生这一类。",[758,11469,11470],{},"AI 原生不是一个细分市场，而是一种能力——它将从内部蚕食所有类别。引入 AI 原生工具的精品机构，会变成截然不同的精品机构。搞定大规模 Agentic 客户服务的超级平台，面目会和现在完全不同。每种竞争原型都会被颠覆，不只是右下角那一个。",[758,11472,11473],{},[778,11474,11475],{},"投资组合转型已经在发布产品了。",[758,11477,11478],{},"实物资产代币化？直接指数化？数字资产机构化？麦肯锡把这些描述为 2030-2035 年的概念。但现货加密 ETF/ETP 的管理规模已经超过 1500 亿美元。直接指数化平台今天正在抢客户。\"统一托管家庭账户\"系统已经在领先机构以早期产品形式运行了。",[758,11480,11481],{},"这些不是试点。是产品。问题不是它们什么时候到来——而是落后者什么时候追得上。",[758,11483,11484],{},[3036,11485],{"alt":11486,"src":4137},"时间线落差：麦肯锡的 2035 地图 vs. 建造者的 2026 现实",[753,11488,11489],{"id":11489},"麦肯锡没有说到的地方",[758,11491,11492],{},"除了时间线之外，这份报告有几个点处理得不够到位。",[758,11494,11495],{},[778,11496,11497],{},"建造者 vs. 采购者的分野。",[758,11499,11500,11501,11504,11505,11508],{},"未来五年，财富管理领域最大的竞争差距，不会在不同类型的公司之间产生——而会在",[2156,11502,11503],{},"自建"," AI 的公司和",[2156,11506,11507],{},"采购"," AI 的公司之间产生。",[758,11510,11511],{},"这个模式我们见过。数字化转型浪潮中，把技术策略外包给供应商的公司，几乎无一例外地输给了自建内部能力的公司。供应商的解决方案总是落后一个世代。拥有自己技术栈的公司可以按周迭代；依赖供应商的公司只能等发版周期。",[758,11513,11514],{},"AI 是同样的模式，只是放大了。自建 AI 系统的公司内化了反馈回路——他们学到什么东西真正打动客户，他们针对自己的具体场景调优模型，他们拥有数据优势。采购 AI 解决方案的公司拿到的是针对平均情况构建的通用产品，而不是针对他们自己的客户。",[758,11516,11517],{},"这个区分在麦肯锡的四个原型里看不到。但它是未来五年最重要的差异化因素。",[758,11519,11520],{},[778,11521,11522],{},"真正的人才危机不只是顾问的问题。",[758,11524,11525],{},"麦肯锡把顾问短缺列为核心挑战。没错。但它下面还藏着第二场人才危机：谁来搭建 AI 系统？",[758,11527,11528,11529,11532],{},"你需要的是既懂金融服务",[2156,11530,11531],{},"又","懂 AI 架构的产品负责人和工程师。不是学会用 ChatGPT 提问的顾问，也不是从没碰过金融产品的 AI 工程师。那个真正的交叉点——能设计客户入职的 Agentic 工作流，同时理解受信任义务、合规约束，还有当客户在市场下跌 20% 时那种情绪动态——这种人真的凤毛麟角。",[758,11534,11535],{},"我正好在这个交叉点上。白天，我在一家金融机构领导技术产品和工程团队——是搭建支撑业务运转的系统，不是管理金融产品。业余时间，我在搭建 AI 原生系统。我可以从直接经验告诉你：这种技能组合才是真正的瓶颈。不是顾问，不是资本，不是监管。横跨金融和 AI 的人才，才是那个约束条件。",[758,11537,11538],{},[778,11539,11540],{},"客户体验会跳跃式进化，而不是渐进式演变。",[758,11542,11543],{},"麦肯锡描述了一种从\"点击到对话\"的渐进演变——随着 AI 越来越好，面向客户的界面平滑地改善。我认为这个心理模型误解了相变的工作方式。",[758,11545,11546],{},"当一个客户可以和一个完全了解他们财务状况的 AI 进行真正的对话——每一个账户、每一个目标、每一个已记录的人生事件——并且在周日晚上 11 点就能得到即时的、个性化的、有数据支撑的答案时，人类顾问那场对话的标准会在一夜之间改变。不是渐渐地。是一夜之间。",[758,11548,11549],{},"能够在这个时代脱颖而出的顾问，不会是那些逐渐改善服务模式的人。而是那些当 AI 处理好所有它能处理的事情后，彻底重新定义\"人的附加价值是什么\"的人。这不是演变。这是相变。",[758,11551,11552],{},[3036,11553],{"alt":11554,"src":4211},"「独立顾问公司」——一位顾问，数百个 AI 代理",[753,11556,11557],{"id":11557},"对你意味着什么",[758,11559,11560,11561,11564],{},"如果你是",[778,11562,11563],{},"一位理财顾问","：那 62% \"打算\"使用 AI 的顾问会分成两批。2024 年或 2025 年就开始的人，正在建立一个随时间复利增长的能力护城河。2027 年才开始的人，要追赶的是已经积累了 2-3 年 AI 辅助实践经验的同行。现在就开始。不要等公司全面推行。",[758,11566,11560,11567,11570],{},[778,11568,11569],{},"金融机构的产品负责人","：推动自建，而不是采购。每多等一年让供应商来交付解决方案，就是多给竞争对手一年时间，让他们的内部团队比你的供应商产品团队学得更快。AI 浪潮以软件速度移动，而不是供应商合同速度。",[758,11572,11560,11573,11576],{},[778,11574,11575],{},"建造者或工程师","：金融服务 AI 严重供给不足。大多数金融科技关注的是消费者支付或借贷。财富管理——也就是数万亿美元资产的实际配置和管理——还在运行 20 年来没有根本改变的顾问工作流。机会巨大，而金融 + AI + 产品的交叉点，正是最大的空白所在。",[758,11578,11560,11579,11582],{},[778,11580,11581],{},"一位客户","：去问问你的财富顾问他们是怎么使用 AI 的。不要问\"你们在研究吗\"——那是任何人都能回答的是非题。问他们过去 12 个月里构建了什么。问他们的初级顾问现在能做什么，是两年前做不到的。\"我们还在探索\"在 2026 年是一个红旗信号。",[753,11584,11586],{"id":11585},"地图是对的时间线不是","地图是对的。时间线不是。",[758,11588,11589],{},"麦肯锡的报告是一张扎实的地图。六大主题是真实的，人口压力是真实的，技术轨迹是真实的。作为一份研究成果，它非常出色。",[758,11591,11592],{},"但由顾问绘制的地图，往往低估了建造者的速度。那些正在真正交付 AI 系统的公司内部人，没有在等 2035 年。他们没有在等 2030 年。他们现在就在迭代，就在这周，而且速度还在加快。",[758,11594,11595],{},"我同时生活在两个世界——白天是金融机构的技术产品负责人，晚上是 AI 原生的建造者。麦肯锡描述的那个未来，不是 10 年后的事。对于移动最快的公司来说，它已经在发生了。对于还在规划的公司来说，它大约还有 5 年——而不是 10 年。",[758,11597,11598],{},"问题不是财富管理会不会被 AI 改变。这个问题已经有了答案。问题是，你的公司会成为这场转变的引领者，还是成为\"为什么落后者输了\"这个案例分析里的主角。",[758,11600,11601],{},"按 2035 年规划，会被 2030 年颠覆。按 2030 年建造，才能引领 2035 年。",[2163,11603],{},[758,11605,11606],{},[2156,11607,11608],{},"这是我的 AI 原生旅程系列的一部分，我在这里分享从金融机构内部看到的 AI 转型真实面貌。如果你也在金融与 AI 的交叉点上构建，欢迎联系。",{"title":839,"searchDepth":708,"depth":708,"links":11610},[11611,11612,11613,11614,11615,11616],{"id":11351,"depth":708,"text":11352},{"id":11412,"depth":708,"text":11412},{"id":11436,"depth":708,"text":11436},{"id":11489,"depth":708,"text":11489},{"id":11557,"depth":708,"text":11557},{"id":11585,"depth":708,"text":11586},"麦肯锡关于财富管理的新报告方向正确，但在时间预测上过于保守。作为一家金融机构内部的技术产品负责人，同时也在搭建 AI 系统，我分析了他们说对的、说晚了 5 年的，以及完全没提到的部分。",{"date":4275,"image":4276,"alt":11619,"tags":11620,"category":3417,"published":859},"麦肯锡 2035 年顾问报告与建造者现实 2030 年——对比图示",[11621,11622,11623,11624,11625],"财富管理","自主AI","金融科技","AI原生","产品领导力","/blogs/zh/mckinsey-wealth-management-2035-ai-insider-take",{"title":11337,"description":11617},"blogs/zh/17.mckinsey-wealth-management-2035-ai-insider-take","dC9YkNjFyQljGIq5_uj56mkKswSBoTmD7IPBHlSZ61g",{"id":11631,"title":11632,"body":11633,"description":11837,"extension":2126,"meta":11838,"navigation":859,"ogImage":4500,"path":11844,"seo":11845,"stem":11846,"__hash__":11847},"zh/blogs/zh/18.ai-wont-make-everyone-a-creator.md","相机没有让每个人都成为斯皮尔伯格，AI 也不会。",{"type":750,"value":11634,"toc":11829},[11635,11638,11641,11644,11647,11654,11657,11660,11665,11669,11672,11675,11678,11689,11696,11699,11702,11707,11711,11714,11717,11720,11726,11729,11734,11737,11740,11743,11746,11753,11756,11759,11762,11767,11771,11777,11780,11783,11786,11789,11792,11798,11803,11806,11809,11812,11815,11818,11821,11823],[758,11636,11637],{},"每隔几十年，就会有一项新技术承诺要让创作平民化。相机、合成器、Photoshop、Auto-Tune。每一次都是同样的预言：现在人人都能成为创作者。每一次都是同样的结果：工具变简单了，但做出优秀的作品并没有。AI 是这个故事的最新一章 —— 而我正身处其中。",[753,11639,11640],{"id":11640},"反复出现的规律",[758,11642,11643],{},"仔细想想。当相机变得人人买得起的时候，人们预言专业摄影将走向终结。任何人都能拍照了。确实如此 —— 任何人都能拍。但安塞尔·亚当斯之所以是安塞尔·亚当斯，不是因为他有一台相机，而是因为他有一双眼睛。",[758,11645,11646],{},"当合成器面世的时候，人们预言任何人都能做音乐了，不需要乐队。确实如此 —— 制作一首曲子的门槛降到了几乎为零。但拥有一台合成器并不能给你品味。排行榜上不会突然充满杰作。它充满的是噪音，而那些伟大的音乐人依然脱颖而出，因为他们有话想说。",[758,11648,11649,11650,11653],{},"Photoshop、桌面出版、Auto-Tune、GarageBand、TikTok —— 故事在不断重复。工具让 ",[2156,11651,11652],{},"制作"," 变得容易。但制作从来不是难点。难的是知道该做什么、为什么有意义、什么时候才算真正做好了。",[758,11655,11656],{},"现在轮到 AI 了。预言比以往任何时候都响亮：AI 将让每个人都能创作专业的内容、视频、音乐、艺术。人人都能成为创作者。",[758,11658,11659],{},"我每天都在用 AI，我热爱这个工具。但我想告诉你 —— 那个预言只讲了故事的一半。",[758,11661,11662],{},[3036,11663],{"alt":11664,"src":4323},"同样的规律反复出现：每一种新工具降低了门槛，但从未提高天花板",[753,11666,11668],{"id":11667},"我用-ai-创作后学到了什么","我用 AI 创作后学到了什么",[758,11670,11671],{},"过去几周，我一直在用 AI 生成博客文章、视频和图片。不是尝鲜 —— 是真正的工作流程。AI 就是我生产内容的方式。",[758,11673,11674],{},"AI 做得极好的地方：初稿、变体、视觉生成。它把我脑海中的想法快速变成实体，比我独自工作快得多。速度是真实的。以前花一整天的事，现在几小时就能搞定。这部分承诺是兑现了的。",[758,11676,11677],{},"AI 做不到的事：决定什么值得做。",[758,11679,11680,11681,11684,11685,11688],{},"它不知道一张生成的图片是 ",[2156,11682,11683],{},"差一点就好"," 还是 ",[2156,11686,11687],{},"真的好","。它不带个人经验和信念。它没法告诉你一篇博客到底有没有观点，还是只是一堆看起来合理的文字。它没有品味。",[758,11690,11691,11692,11695],{},"当我开始使用 AI 后，工作量并没有消失，而是转移了。我花在执行上的时间少了 —— 写初稿、生成图片、排版内容。我花在甄选、打磨和创意决策上的时间 ",[2156,11693,11694],{},"多了","。更多时间在决定留什么、扔什么。更多时间在问自己：这东西到底有没有话说？",[758,11697,11698],{},"这个转变是没人谈论的那部分。那些给你展示\"我用 AI 5 分钟就做出了这个\"的人，没有给你看他们花了 2 小时弄清楚该做什么，扔掉了 30 个版本，以及最后花时间重写了那些听起来像机器写的部分。",[758,11700,11701],{},"\"AI 生成了这个\"和\"这个真的好\"之间的差距，填充它的依然是人。",[758,11703,11704],{},[3036,11705],{"alt":11706,"src":4362},"工作量转移了：执行减少，判断和甄选增多",[753,11708,11710],{"id":11709},"ai-垃圾问题","AI 垃圾问题",[758,11712,11713],{},"如果 AI 能让质量毫不费力，我们现在应该生活在一个创意黄金时代。然而实际情况是，我们迎来了\"AI 垃圾\"（AI slop）。",[758,11715,11716],{},"\"AI slop\"这个词被韦氏词典和澳大利亚国家词典同时评为 2025 年度词汇。它描述的是充斥互联网的低质量 AI 生成内容。数据触目惊心：Kapwing 的研究估计 YouTube 上 21% 到 33% 的内容可能是 AI 生成的垃圾，每年带来约 1.17 亿美元的广告收入。",[758,11718,11719],{},"2026 年 1 月，Bandcamp 全面禁止 AI 生成音乐。调查发现 52% 的消费者在怀疑内容是 AI 生成的时候会减少互动。受众已经对低质量 AI 内容产生了\"第六感\" —— 他们能发现那些过于热情的形容词、重复的结构、以及缺乏具体细节的表达 —— 这些都是有人按了\"生成\"键然后就完事了的标志。",[758,11721,11722,11723,11725],{},"问题不在 AI。问题在于有人认为工具 ",[2156,11724,10790],{}," 产品。",[758,11727,11728],{},"相机不能让一张照片变得伟大。Photoshop 不能让一个设计变得伟大。AI 也不能让内容变得伟大。工具就是工具。让作品伟大的，是背后的人 —— 他们的判断力、品味，以及愿意扔掉前十个输出、继续打磨直到满意的决心。",[758,11730,11731],{},[3036,11732],{"alt":11733,"src":4390},"AI 垃圾：当工具变成了产品，质量就消失了",[753,11735,11736],{"id":11736},"炒作机器",[758,11738,11739],{},"那为什么\"AI 能做一切\"的叙事还在持续？",[758,11741,11742],{},"因为这是一个好讲的故事。很多人在重复这个说法 —— 不是因为它符合他们的经验，而是因为这是趋势。",[758,11744,11745],{},"TechCrunch 宣称 2026 年是 AI 从\"炒作走向务实\"的一年。市场正在清醒。但炒作机器还在运转，推动它的是那些卖 AI 工具、AI 课程和 AI 梦想的人。",[758,11747,11748,11749,11752],{},"有一个数据能说明真实情况。研究人员用 10 万人对 AI 做了创造力测试。标题写的是：\"AI 击败人类。\"实际数据呢？AI 击败的是 ",[2156,11750,11751],{},"平均水平"," 的人类。但最优秀的 10% 远远超越了每一个 AI 系统。最好的人类不是接近 —— 是碾压。",[758,11754,11755],{},"标题说\"AI 击败人类\"。数据说的是\"AI 击败了平庸的人类\"。这是两个完全不同的故事。",[758,11757,11758],{},"一个叫做\"人性护城河\"（Humanity Moat）的概念正在兴起。在一个人人都能用同样 AI 工具的世界里，竞争优势不在工具 —— 在于信任、真实和品味。这些是 AI 生成不了的。",[758,11760,11761],{},"那些告诉你\"AI 让一切变得毫不费力\"的人，通常在卖东西。真正每天用 AI 在做事的人知道真相：AI 非常强大，但它依然需要你带着值得放大的东西来。",[758,11763,11764],{},[3036,11765],{"alt":11766,"src":4425},"标题说\"AI 击败人类\" —— 数据讲的是另一个故事",[753,11768,11770],{"id":11769},"ai-是电动工具不是魔法棒","AI 是电动工具，不是魔法棒",[758,11772,11773,11774],{},"经过几周的 AI 创作实践，我找到了一个心智模型：",[778,11775,11776],{},"AI 是有史以来最好的放大器。",[758,11778,11779],{},"如果你有品味、有远见、有话想说 —— AI 让你快 10 倍。它处理繁重的工作，让你专注于创意决策。它生成变体，让你选出最好的。它在几分钟内把粗糙的想法变成成品，而不是几小时。",[758,11781,11782],{},"如果你没有品味、没有远见、没有话想说 —— AI 帮你更快地生产平庸内容。仅此而已。更快的平庸。",[758,11784,11785],{},"一项针对 1,100 多名音乐制作人的调查发现，他们对 AI 的期望是\"在节省时间的同时不抹杀创造力\"。他们要的是放大，不是替代。他们希望工具强大，但音乐人依然做主。",[758,11787,11788],{},"这才是正确的类比。AI 给了每个人一台相机，但没有给每个人一双眼睛。AI 给了每个人一间录音棚，但没有给每个人一首值得录制的歌。",[758,11790,11791],{},"我想说清楚：这不是反对 AI 的论点。我每天都在用它，它真正改变了我的工作方式。我产出更多、迭代更快、探索了以前不会尝试的想法。",[758,11793,11794,11795,11797],{},"但我没有搞混。AI 在 ",[2156,11796,10921],{}," 层面做着重活。品味、判断、视角 —— 那依然是我。而那才是让作品真正好的东西。",[758,11799,11800],{},[3036,11801],{"alt":11802,"src":4462},"AI 是放大器：你带来什么决定了产出什么",[753,11804,11805],{"id":11805},"那个不变的常量",[758,11807,11808],{},"技术在变。根本等式不变。",[758,11810,11811],{},"在印刷术时代，瓶颈不是印刷 —— 而是有没有值得印刷的东西。在 YouTube 时代，瓶颈不是上传 —— 而是有没有值得观看的东西。在 AI 时代，瓶颈不是生成 —— 而是有没有值得生成的东西。",[758,11813,11814],{},"在 AI 时代能茁壮成长的创作者，不是最依赖 AI 的人，而是那些知道自己想说什么 —— 并用 AI 把它说得更好的人。",[758,11816,11817],{},"AI 是我们这一代最好的创作工具。用它，拥抱它。但不要把工具当成才华。",[758,11819,11820],{},"技术从来不是瓶颈。你的品味、你的判断、你的视角 —— 那才一直是难点。现在依然是。",[2163,11822],{},[758,11824,11825,11826,11828],{},"我每周分享用 AI 构建的真实经验 —— 不是炒作，只有真实的体会。如果你想要关于 AI 创作的诚实看法，欢迎订阅 ",[778,11827,3956],{}," 通讯。",{"title":839,"searchDepth":708,"depth":708,"links":11830},[11831,11832,11833,11834,11835,11836],{"id":11640,"depth":708,"text":11640},{"id":11667,"depth":708,"text":11668},{"id":11709,"depth":708,"text":11710},{"id":11736,"depth":708,"text":11736},{"id":11769,"depth":708,"text":11770},{"id":11805,"depth":708,"text":11805},"AI 是有史以来最强大的创作工具，但技术从来不是优秀作品的瓶颈。品味、远见和判断力才是——现在依然是。",{"date":4499,"image":4500,"alt":11839,"tags":11840,"category":3758,"youtube":4506,"published":859},"创作工具的进化——从相机到 AI——不变的是人的品味",[2130,11841,11842,11843],"创造力","内容创作","观点","/blogs/zh/ai-wont-make-everyone-a-creator",{"title":11632,"description":11837},"blogs/zh/18.ai-wont-make-everyone-a-creator","fU03Yb_mni7-JNxELZ0Nnp0UaDeHndL8diuglBC6UDw",{"id":11849,"title":11850,"body":11851,"description":12111,"extension":2126,"meta":12112,"navigation":859,"ogImage":4788,"path":12120,"seo":12121,"stem":12122,"__hash__":12123},"zh/blogs/zh/19.ai-two-work-modes.md","AI 让我效率提升了 10 倍，然后我差点累垮了",{"type":750,"value":11852,"toc":12103},[11853,11856,11859,11863,11866,11869,11872,11875,11880,11883,11886,11889,11892,11895,11898,11903,11907,11910,11913,11920,11923,11926,11929,11934,11938,11949,11960,11965,11982,11989,11994,12011,12014,12021,12026,12029,12032,12038,12044,12050,12056,12061,12064,12067,12078,12081,12084,12087,12089,12092,12098],[758,11854,11855],{},"我同时跑着 5 个 AI Agent，一边写内容，一边生成视频，一边自动化各种工作流。我的产出比以往任何时候都高。然后我撞了墙 — 不是因为 AI 出了问题，而是因为我在错误的任务上用了错误的模式。",[758,11857,11858],{},"这是没人提醒你的部分。",[753,11860,11862],{"id":11861},"_10-倍陷阱","10 倍陷阱",[758,11864,11865],{},"几周前，我搭建了一套 AI 内容管线：从一个粗略的想法开始，经过 AI 辅助，最终产出一篇博客、配图、视频以及 X 和 Newsletter 的分发内容。我同时运行着多个 Agent：一个生成图片，一个起草文案，另一个组装视频。感觉像是拥有了超能力。",[758,11867,11868],{},"数据也证明了这一点。我的产出前所未有地高。过去一周才能完成的工作，现在一天半就搞定了。我成了一个人的内容工作室。",[758,11870,11871],{},"但有些地方不对劲。我发货更快了，却更累了。我的决策质量在下降。回头检查自己的产出时，我会发现一些平时不会犯的错误。我在五个并行任务之间来回切换，没有一个得到了应有的专注。",[758,11873,11874],{},"更多的产出，更少的清晰度。不只是我一个人有这种感觉。",[758,11876,11877],{},[3036,11878],{"alt":11879,"src":4543},"10 倍陷阱：更多产出，更少清晰度，不断累积的疲劳",[753,11881,11882],{"id":11882},"研究证实了这一点",[758,11884,11885],{},"2026 年 2 月，UC Berkeley 的研究团队发表了一项引爆全网的研究。标题：AI 不会减少工作 — 它会加剧工作强度。采用 AI 工具的员工并没有减少工作量，反而承担了更广泛的任务、以更快的节奏工作，并把工作时间延伸到了更多时段。正如一位受访者所说：\"你本来以为可以少工作一些，但实际上你没有。你工作的量一样多，甚至更多。\"",[758,11887,11888],{},"Harvard Business Review 识别了 AI 工作强化的三种形式：任务膨胀（你开始承担以前会委托出去的工作）、边界模糊（你在午餐、通勤、晚间都在跟 AI 对话）、以及多任务增加（你在日常工作之上还要管理并行的 AI 工作流）。",[758,11890,11891],{},"最令人震惊的发现来自 TechCrunch：最积极拥抱 AI 的人，反而最先出现倦怠。不是那些持怀疑态度的人在倒下，而是最热情的信徒。",[758,11893,11894],{},"与此同时，2026 年初对数千名 CEO 的调查显示，AI 对其公司的生产力没有可衡量的影响。Solow 悖论 — \"你到处都能看到计算机时代，唯独在生产力统计中看不到\" — 以 AI 的面貌重新出现了。",[758,11896,11897],{},"问题不在工具，而在于我们没有一个使用它的框架。",[758,11899,11900],{},[3036,11901],{"alt":11902,"src":4568},"研究数据：AI 加剧了工作强度，而非减轻了它",[753,11904,11906],{"id":11905},"kahneman-在-15-年前就给出了答案","Kahneman 在 15 年前就给出了答案",[758,11908,11909],{},"Daniel Kahneman 的《思考，快与慢》提出了一个能精确解释当前状况的框架。他描述了两种思维系统：系统 1 — 快速、直觉、自动化；系统 2 — 缓慢、刻意、费力。",[758,11911,11912],{},"系统 1 是你认出一张面孔、对突然的声响做出反应、快速浏览邮件的方式。它快速且低成本。系统 2 是你解一道复杂数学题、做一个艰难的职业决定、调试一个微妙的软件 Bug 的方式。它缓慢且消耗精力。",[758,11914,11915,11916,11919],{},"关键在于：AI 是有史以来最强大的系统 1 加速器。它让快速的工作变得",[2156,11917,11918],{},"更快","。起草、生成、研究、搭框架、探索 — 所有需要速度和广度的并行生成性任务 — AI 都能大幅加速。",[758,11921,11922],{},"但系统 2 的工作 — 深度推理、判断、解决矛盾、在答案不明显时做出决策 — 无法并行化。你不可能同时进行五个深度思考，就像你不可能同时进行五场严肃的对话一样。",[758,11924,11925],{},"我们早就知道这种张力的存在。有些人擅长发散性、探索性思维 — 头脑风暴、构思、在不同线程之间切换。有些人则在收敛性、专注性思维中表现最佳 — 深入钻研一个问题直到突破。左脑/右脑的区分是个神话，但其背后的直觉是真实的：不同的任务需要根本不同的认知模式。",[758,11927,11928],{},"AI 没有创造这种张力，它放大了这种张力。而我们大多数人还没有适应过来。",[758,11930,11931],{},[3036,11932],{"alt":11933,"src":4604},"Kahneman 的系统 1 和系统 2，对应到 AI 工作模式",[753,11935,11937],{"id":11936},"两种模式scatter-和-laser","两种模式：Scatter 和 Laser",[758,11939,11940,11941,11944,11945,11948],{},"在撞墙之后，我开始留意什么时候 AI 让我变得更好，什么时候让我变得更差。规律很清晰：关键不在于用更多还是更少的 AI，而在于把正确的",[2156,11942,11943],{},"模式","匹配到正确的",[2156,11946,11947],{},"任务","上。",[758,11950,11951,11952,11955,11956,11959],{},"我把它们叫做 ",[778,11953,11954],{},"Scatter Mode（发散模式）"," 和 ",[778,11957,11958],{},"Laser Mode（聚焦模式）","。",[758,11961,11962,11964],{},[778,11963,4624],{}," 是并行的、探索性的、AI 赋能的。它是加了涡轮的系统 1。适合追求广度的时候：",[804,11966,11967,11970,11973,11976,11979],{},[775,11968,11969],{},"同时在多个方向上研究一个课题",[775,11971,11972],{},"生成草稿、变体和选项",[775,11974,11975],{},"让多个 AI Agent 并行处理独立任务",[775,11977,11978],{},"探索新想法、快速原型、代码脚手架",[775,11980,11981],{},"搭建内容管线：博客、视频、配图、分发",[758,11983,11984,11985,11988],{},"Scatter Mode 是 AI 最闪耀的地方。你可以跑五个 Agent、探索十个方向、生成二十个版本 — 然后再收敛。关键词是",[2156,11986,11987],{},"独立","。如果任务之间互不依赖，就分散去做。",[758,11990,11991,11993],{},[778,11992,4628],{}," 是串行的、专注的、以人为主导的。它是系统 2，被保护免受干扰。适合需要深度的时候：",[804,11995,11996,11999,12002,12005,12008],{},[775,11997,11998],{},"调试一个微妙且层层嵌套的问题",[775,12000,12001],{},"做一个有多方权衡的产品架构决策",[775,12003,12004],{},"打磨一篇文章直到它准确表达你的意思",[775,12006,12007],{},"设计一个每一步都依赖前一步的复杂工作流",[775,12009,12010],{},"和队友进行一次困难的对话",[758,12012,12013],{},"Laser Mode 是人类不可替代的领域。你可以在这里使用 AI — 一次专注的对话 — 但你无法并行化。这种工作需要你把整个问题装在脑子里，追踪依赖链条，行使只有深度投入才能产生的判断力。",[758,12015,12016,12017,12020],{},"改变我工作方式的洞察：",[778,12018,12019],{},"模式比工具重要。"," Scatter Mode 下的 AI（五个 Agent 生成选项）和 Laser Mode 下的 AI（一次专注的对话解决难题）完全是两回事。同一个工具，不同的认知语境，截然不同的结果。",[758,12022,12023],{},[3036,12024],{"alt":12025,"src":4695},"两种模式：Scatter（并行、广度、AI 赋能）vs Laser（串行、深度、人类主导）",[753,12027,12028],{"id":12028},"我的实际做法",[758,12030,12031],{},"这是我日常工作中的具体实践。",[758,12033,12034,12037],{},[778,12035,12036],{},"上午是 Scatter Mode。"," 我用探索性、生成性的工作开启一天。多个方向的研究，启动并行的 AI 任务 — 图片生成、草稿写作、代码脚手架。把能在后台运行的任务排好队，同时在不同线程之间穿梭。这时候我最像一个指挥家 — 编排多条流水线，而不是深入执行任何一条。",[758,12039,12040,12043],{},[778,12041,12042],{},"下午是 Laser Mode。"," 我关闭并行的流水线。一个问题，一个屏幕。有时候用一次 AI 对话，但通常就是我和代码或文档面对面。这是我做艰难决定的时候：产品决策、工作流设计、调试那些不会轻易就范的问题。",[758,12045,12046,12049],{},[778,12047,12048],{},"我的经验法则："," 如果任务有依赖关系或需要在多个权衡之间做判断 — 用 Laser。如果是探索性的、生成性的、或者子任务彼此独立 — 用 Scatter。",[758,12051,12052,12055],{},[778,12053,12054],{},"切换仪式很重要。"," 从 Scatter 切换到 Laser 不是瞬间完成的。我会关闭标签页，停掉后台 Agent，给自己五分钟缓冲来调整认知模式。从\"管理五条并行流\"到\"解决一个难题\"需要一次精神重置。试图直接从 Scatter 跳到 Laser，就是犯错的开始。",[758,12057,12058],{},[3036,12059],{"alt":12060,"src":4732},"每日实践：上午 Scatter，下午 Laser，中间有切换仪式",[753,12062,12063],{"id":12063},"不变的是你",[758,12065,12066],{},"在上一篇文章中，我写道 AI 是有史以来最好的放大器。我仍然相信这一点。但现在我看到了更细微的图景。",[758,12068,12069,12070,12073,12074,12077],{},"在 Scatter Mode 下，AI 放大你的",[2156,12071,12072],{},"广度"," — 你能探索多少方向，生成多少选项，并行运行多少任务。在 Laser Mode 下，AI 放大你的",[2156,12075,12076],{},"深度"," — 你推理的质量，输出的精确度，达到深思熟虑决定的速度。",[758,12079,12080],{},"但在两种模式下，做选择的都是人。你决定什么时候发散，什么时候聚焦。你决定广度什么时候服务于你，什么时候只是噪音。你决定什么时候让五个 Agent 运行，什么时候关掉一切，安静地思考。",[758,12082,12083],{},"AI 时代能脱颖而出的人，不是最快的多任务处理者，也不是最深度的专注者。而是那些能在两者之间自如切换的人 — 他们知道元技能不是使用 AI，而是知道自己该处于哪种模式。",[758,12085,12086],{},"Kahneman 是对的：知道什么时候快速思考、什么时候慢速思考，一直是人类的核心优势。AI 只是提高了赌注。",[2163,12088],{},[758,12090,12091],{},"你的体验是什么？你是否发现自己总是卡在一种模式里 — 总在发散，或者总在聚焦？我真心好奇，欢迎在评论区告诉我。",[758,12093,12094,12095,12097],{},"我每周都会分享用 AI 构建产品的真实经验 — 不是炒作，只是实践中学到的东西。订阅 ",[778,12096,3956],{},"，获取真实故事。",[758,12099,12100],{},[3036,12101],{"alt":12102,"src":4776},"不变的是你：你选择模式",{"title":839,"searchDepth":708,"depth":708,"links":12104},[12105,12106,12107,12108,12109,12110],{"id":11861,"depth":708,"text":11862},{"id":11882,"depth":708,"text":11882},{"id":11905,"depth":708,"text":11906},{"id":11936,"depth":708,"text":11937},{"id":12028,"depth":708,"text":12028},{"id":12063,"depth":708,"text":12063},"AI 不仅改变了我们能做什么，还改变了我们应该如何思考工作本身。一个帮你判断何时发散（并行、探索、AI 赋能）、何时聚焦（深度、专注、人类主导）的框架。",{"date":4787,"image":4788,"alt":12113,"tags":12114,"category":3016,"youtube":4796,"published":859},"两种工作模式 — Scatter 和 Laser — 以及在两者之间做选择的人",[12115,12116,12117,12118,12119],"AI生产力","工作模式","深度工作","系统思维","职业倦怠","/blogs/zh/ai-two-work-modes",{"title":11850,"description":12111},"blogs/zh/19.ai-two-work-modes","Eb8M8q2dwYmHF61JxGqoqBrJCZMnUEGKAmFzVipqB6U",{"id":12125,"title":12126,"body":12127,"description":12348,"extension":2126,"meta":12349,"navigation":859,"ogImage":5040,"path":12357,"seo":12358,"stem":12359,"__hash__":12360},"zh/blogs/zh/20.chamath-process-over-goals.md","我追了 15 年的目标，Chamath 让我看到了问题所在",{"type":750,"value":12128,"toc":12341},[12129,12132,12135,12138,12141,12147,12150,12153,12156,12159,12161,12168,12171,12176,12180,12183,12189,12195,12201,12207,12213,12223,12228,12231,12238,12241,12244,12250,12253,12256,12263,12266,12271,12274,12277,12280,12286,12292,12298,12304,12311,12314,12319,12322,12325,12328,12331,12333,12336],[758,12130,12131],{},"我这辈子都在设定目标。30 岁前升 Senior Manager。Q3 前发布产品。年底达成营收指标。每一个都完成了。然后我看了 Chamath Palihapitiya —— 那个打造了 Facebook 增长引擎和 Social Capital 的人 —— 解释为什么目标导向从根本上就是错误的框架。",[758,12133,12134],{},"他的观点是：真正赢的人不是在追逐终点，而是承诺于一个持续学习、永不停止的过程。我用这个观点审视了自己的职业生涯 —— 他说得令人不安地正确。",[753,12136,12137],{"id":12137},"目标陷阱",[758,12139,12140],{},"十五年来，我围绕里程碑优化自己的职业。30 岁前升 Senior Manager，35 岁前做到 Partner。发布产品，完成指标。每个目标都清晰、可衡量、有驱动力 —— 完全符合管理学教科书的标准。",[758,12142,12143,12144],{},"我确实达到了，大部分都达到了。但没人提前告诉你的是：到达之后的那一刻。有一阵短暂的兴奋 —— 一周，也许两周 —— 然后是一种安静的空虚。你的大脑唯一给出的回应是：",[2156,12145,12146],{},"然后呢？",[758,12148,12149],{},"于是你设定新目标。再一个。目标 → 达成 → 空虚 → 新目标 → 循环往复。",[758,12151,12152],{},"心理学家 Tal Ben-Shahar 把这叫做\"到达谬误\"（arrival fallacy）—— 一种认为到达目的地就会获得持久幸福的幻觉。他通过自己作为精英壁球运动员的经历发现了这一点：每场胜利带来的兴奋几乎立刻消退。目的地从来不是重点。但他花了很多年才意识到这一点。",[758,12154,12155],{},"我也是。",[758,12157,12158],{},"然后我看了 Chamath 的视频 ——\"30 Years of Business Advice\" —— 一下子就通了。",[4838,12160],{"title":4840,"videoId":4841},[758,12162,12163,12164,12167],{},"他的第一条规则是\"永远不要停下来\"。但他说的不是内卷文化，而是承诺于",[2156,12165,12166],{},"学习和冒险的过程","，而不是终点。Buffett 95 岁了还在做投资，不是因为他在追目标，而是因为他在运行一个过程。Munger 也一样。那些能在几十年间保持卓越的人，不是目标设得最好的人，而是从未停止学习的人。",[758,12169,12170],{},"Chamath 把 30 年的经验浓缩成 6 条规则。当你用\"过程大于目标\"的视角去看，它们不再是一张建议清单，而是一个连贯的操作系统。",[758,12172,12173],{},[3036,12174],{"alt":12175,"src":4857},"目标陷阱：达成、空虚、循环",[753,12177,12179],{"id":12178},"_6-条规则构成的系统","6 条规则构成的系统",[758,12181,12182],{},"大多数关于 Chamath 视频的总结都忽略了一点：这不是 6 条独立的建议，而是相互关联的原则，它们都在保护同一件事 —— 你持续学习的能力。去掉任何一条，系统就会崩溃。",[758,12184,12185,12188],{},[778,12186,12187],{},"规则 1：永远不要停下来。"," 这是基础。不要把人生当作一系列目标，达到了就宣告胜利、停止前进。承诺于持续学习和持续冒险。不是\"每周工作 80 小时\"，而是永远不允许自己停止成长。我职业生涯中最停滞的几年，恰恰是我在\"按计划执行\"而不是在学习的时候。我有路线图、时间表、清晰的交付物 —— 但在智力上处于惯性滑行。最有生产力的时期？最近六个月。没有计划，只是每天用 AI 去构建东西来学习。",[758,12190,12191,12194],{},[778,12192,12193],{},"规则 2：零负债。"," 债务 —— 财务的和技术的 —— 会迫使你转向目标导向。当你欠债时，你会优化短期回报，而不是长期学习。债务压缩了选择空间。Chamath 指出，社交媒体上精心策划的生活方式把年轻人推向物质消费 → 负债 → 被困在讨厌的工作里。我在产品工作中看到同样的模式：当代码库被技术债淹没时，每个 sprint 都在打补丁，而不是在学习。个人层面也一样：财务自由不是为了奢侈，而是为了保留学习的自由。",[758,12196,12197,12200],{},[778,12198,12199],{},"规则 3：以谦逊管理。"," 对你现在所处的位置保持绝对诚实。不是你希望自己在哪里，不是你的目标说你应该在哪里。是对现状的真实评估。我 AI 转型中最困难的部分，是承认自己是从零开始的。不是\"我有 15 年经验，我会很快搞定\"。真实的评估是：我是 AI 原生构建的初学者。我不知道自己不知道什么。那种谦逊让人不舒服 —— 但它解锁了比任何自信都更快的学习。",[758,12202,12203,12206],{},[778,12204,12205],{},"规则 4：和年轻人在一起。"," Chamath 说你的认知会在某个时间点固化。年轻人是\"未来的预警系统\" —— 他们帮你看到盲区。这不是指导（mentorship），那是一个目标导向的框架（\"我来教他们\"）。这是保持与新兴模式的连接，那是一个过程。我团队里的初级开发者比任何人都更快地采用了 AI 工具。他们没有 15 年的\"事情就该这么做\"需要忘掉。他们是我意识到变革是真实的信号 —— 以及我需要行动的信号。",[758,12208,12209,12212],{},[778,12210,12211],{},"规则 5：保留选择权。"," 在商业、谈判、人生中都追求双赢。不要锁死在封闭选择的路径上。这从根本上是反目标的：目标把你的注意力收窄到一个结果，而选择权保持它的开放。我一直在保持日常工作的同时做副项目。不是\"我要辞职 all-in\" —— 那是目标导向的思维。而是\"我在积累能力，保持大门敞开\" —— 过程导向的思维保护了关系、保留了选项、减少了被锁定在单一路径上的自毁行为。",[758,12214,12215,12218,12219,12222],{},[778,12216,12217],{},"规则 6：地位是被制造的。"," 所有外部认可 —— 排行榜、俱乐部、邀请 —— 都是钩子，让别人对你有裁判权。追逐地位是终极的目标导向陷阱，它优化的是",[2156,12220,12221],{},"别人的","记分卡，而不是你自己的学习。Chamath 说脱离地位游戏是一种超能力。我亲身体会到了这一点：头衔和晋升让人兴奋一周。真正的技能积累 —— 学习编写 AI agent、每天发布内容、从零搭建内容工作室 —— 产生的复利效应是任何头衔都无法比拟的。没人为此颁发证书。这恰恰是重点。",[758,12224,12225],{},[3036,12226],{"alt":12227,"src":4910},"6 条规则构成一个互联系统，都在保护持续学习",[753,12229,12230],{"id":12230},"为什么现在比以往任何时候都重要",[758,12232,12233,12234,12237],{},"你可以说\"过程大于目标\"一直都是好建议。你说得对 —— Kahneman、Seligman 和一个世纪的心理学研究都支持这一点。但有一个原因让这个框架",[2156,12235,12236],{},"此刻","特别关键。",[758,12239,12240],{},"AI 正在加速世界变化的速度。当格局每六个月就会改变一次，基于今天假设设定的目标在你到达之前就已经过时了。2025 年 1 月设定\"精通 prompt engineering\"目标的人，到 2025 年 12 月看着 AI agent 让这项技能的重要性大幅降低。而那个承诺\"理解 AI 系统如何运作\"的人无缝适应了 —— 因为他的北极星是过程，不是目的地。",[758,12242,12243],{},"过程导向的人自动适应。他们不需要撕掉计划重新来过。他们一开始就没有执着于某个具体的计划。",[758,12245,12246,12247,11959],{},"Chamath 引用了 Curt Richter 1957 年的一个实验，让这个道理变得直观。Richter 把野生老鼠放进水罐里，测量它们在溺水前能游多久。大多数在 15 分钟内放弃。但当他短暂地救出它们 —— 把它们拉出来，让它们休息，然后放回去 —— 被救过一次的老鼠能游长达 60 小时。不是 60 分钟。60 ",[2156,12248,12249],{},"小时",[758,12251,12252],{},"区别不在于体力。而在于信念。被救过的老鼠相信情况不是绝望的。它们有证据表明坚持可能带来救援。",[758,12254,12255],{},"对于在 AI 不确定性中航行的创造者来说，等价物不是相信某个具体结果会发生。而是相信你自己的过程 —— 如果你持续学习、持续构建、持续适应，你会找到出路。这种信念是在一切都在变化时支撑你的东西。",[758,12257,12258,12259,12262],{},"Chamath 给年轻人的实操建议强化了这一点：\"去主舞台。\"做科技就去硅谷，做金融就去纽约。这听起来像个目标 —— 搬到某个城市 —— 但实际上是过程思维。你在优化",[2156,12260,12261],{},"学习环境和机会密度","，而不是某个具体的回报。",[758,12264,12265],{},"他最有争议的观点：\"工作与生活的平衡是个伪命题。\"先别急着否定。当你找到一个让你进入心流状态的过程时，平衡的问题自然消解了。不是因为你 24 小时工作，而是因为这份工作不像牺牲。我做 AI 构建的体验就是这样：有时候我晚上 11 点还在做 —— 不是因为我必须，而是因为我在心流中。那不是失衡，那是对齐。",[758,12267,12268],{},[3036,12269],{"alt":12270,"src":4955},"为什么过程导向在 AI 时代比以往更重要：格局变化太快，固定目标来不及追",[753,12272,12273],{"id":12273},"反目标操作系统",[758,12275,12276],{},"吸收了 Chamath 的框架之后，我审视了自己 AI 转型期间的运作方式 —— 这段时间是我生产力最高、也最有满足感的时期 —— 然后意识到我其实已经在运行这个系统的某个版本了。只是之前没有明确说出来。",[758,12278,12279],{},"这是它在实践中的样子：",[758,12281,12282,12285],{},[778,12283,12284],{},"没有 5 年计划。"," 取而代之的是：每日学习仪式。每天用 AI 构建一些东西。有时候是一个新的工作流，有时候是一篇博客，有时候是一个失败的实验 —— 但它教会我的东西是其他方式学不到的。",[758,12287,12288,12291],{},[778,12289,12290],{},"副项目没有营收目标。"," 取而代之的是：发布和迭代。让市场教我。从发布和构建中获得的反馈，比我在电子表格里能写的任何预测都更有价值。",[758,12293,12294,12297],{},[778,12295,12296],{},"Newsletter 没有粉丝目标。"," 取而代之的是：持续发布，每篇比上一篇更好。增长是过程的副产品，而不是目标。",[758,12299,12300,12303],{},[778,12301,12302],{},"没有职业目的地。"," 取而代之的是：积累能复利的能力。我构建的每一项技能都会反哺下一项。写作提升了我的思维，构建 AI 工作流让我看到了什么是可能的，发布内容让我知道什么能引起共鸣。",[758,12305,12306,12307,12310],{},"悖论是真实存在的：过程导向往往比目标导向更快地达成\"目标\"。当你不再把精力浪费在焦虑差距 —— 你现在在哪里和目标说你",[2156,12308,12309],{},"应该","在哪里之间的距离 —— 你就把那份精力重新导向了实际的工作。James Clear 完美地表达了这一点：\"你不会升到你的目标的高度。你会降到你的系统的水平。\"Chamath 从不同的角度在说同一件事。",[758,12312,12313],{},"复利效应是关键。当你的北极星是\"持续学习\"，每天的产出 —— 即使是失败 —— 都会反哺明天的能力。当你的北极星是\"在 Y 日期前达到 X\"，那些没有推进指标的日子就感觉被浪费了。一个框架产生复利。另一个制造焦虑。",[758,12315,12316],{},[3036,12317],{"alt":12318,"src":5005},"反目标操作系统：每日学习、不设目标、能力复利",[753,12320,12321],{"id":12321},"过程就是意义",[758,12323,12324],{},"Chamath 的 6 条规则不是 6 个独立的教训。它们是同一个教训的 6 个面向：承诺于学习的过程，移除那些阻止你的东西 —— 债务、地位追逐、固定目标、自我 —— 并对你真正所处的位置保持绝对诚实。",[758,12326,12327],{},"我近年来收到的最好的职业建议，不是\"设更大的目标\"或\"更努力地拼\"。而是：真正赢的人是那些从不停止学习的人。不是因为他们有一个目标要达成。而是因为过程本身就是意义所在。",[758,12329,12330],{},"我的 AI 转型就是活生生的证明。没有商业计划。没有时间表。没有营收目标。只是承诺每一天都用 AI 去构建。而不知怎的，它产出的成果比我写过的任何目标导向的计划都多。",[2163,12332],{},[758,12334,12335],{},"我正在记录这个过程 —— 用 AI 构建、公开学习、每周发布。如果这些引起了你的共鸣，欢迎关注。",[758,12337,12338],{},[3036,12339],{"alt":12340,"src":5029},"过程就是意义：承诺于学习，让结果自然复利",{"title":839,"searchDepth":708,"depth":708,"links":12342},[12343,12344,12345,12346,12347],{"id":12137,"depth":708,"text":12137},{"id":12178,"depth":708,"text":12179},{"id":12230,"depth":708,"text":12230},{"id":12273,"depth":708,"text":12273},{"id":12321,"depth":708,"text":12321},"Chamath Palihapitiya 的 6 条规则不是独立的建议，它们构成了一个保护持续学习的操作系统。我将每条规则映射到自己 15 年的职业生涯和 AI 转型经历，展示为什么过程导向在当下比以往任何时候都重要。",{"date":5039,"image":5040,"alt":12350,"tags":12351,"category":3016,"youtube":5048,"published":859},"过程大于目标 — Chamath 30 年商业智慧背后的操作系统",[12352,12353,12354,12355,11625,12356],"Chamath","过程大于目标","职业策略","持续学习","AI职业","/blogs/zh/chamath-process-over-goals",{"title":12126,"description":12348},"blogs/zh/20.chamath-process-over-goals","NxX5GFQwadjGGLxkaJ2KjMd3hXV1zQ-FxraTX9SSnhY",{"id":12362,"title":12363,"body":12364,"description":12795,"extension":2126,"meta":12796,"navigation":859,"ogImage":5526,"path":12803,"seo":12804,"stem":12805,"__hash__":12806},"zh/blogs/zh/21.anthropic-finance-plugins-insider-take.md","Anthropic把一套完整的AI技术栈交给了金融团队。我用真实工作跑了一遍——以下是我的发现。",{"type":750,"value":12365,"toc":12784},[12366,12369,12372,12375,12378,12428,12445,12448,12451,12456,12459,12462,12465,12468,12473,12476,12480,12483,12486,12491,12494,12505,12510,12515,12518,12523,12531,12534,12538,12541,12544,12547,12552,12555,12561,12567,12573,12576,12579,12582,12585,12588,12591,12595,12598,12601,12604,12607,12610,12613,12618,12626,12629,12632,12637,12640,12643,12652,12658,12667,12673,12678,12681,12684,12687,12690,12693,12696,12701,12705,12708,12770,12772,12777],[758,12367,12368],{},"Anthropic本周发布了financial-services-plugins——五个Claude插件，覆盖DCF估值模型、LBO分析、股票研究报告、IC备忘录和财富管理工作流，并直接连接包括FactSet、标普全球、晨星、穆迪在内的11个机构级数据源。",[758,12370,12371],{},"我在一家金融公司做软件开发。这周我把整个repo读了一遍——然后真正跑了一遍：生成了一个DCF模型和一份客户报告。以下是我实际发现的东西。",[753,12373,12374],{"id":12374},"里面有什么",[758,12376,12377],{},"五个插件。一个核心基础层需要先装（Financial Analysis），然后根据团队职能叠加四个专业模块：",[5072,12379,12380,12390],{},[5075,12381,12382],{},[5078,12383,12384,12387],{},[5081,12385,12386],{},"插件",[5081,12388,12389],{},"做什么",[5088,12391,12392,12401,12410,12419],{},[5078,12393,12394,12398],{},[5093,12395,12396],{},[778,12397,5097],{},[5093,12399,12400],{},"起草CIM、构建买家列表、运行并购模型、追踪交易进展",[5078,12402,12403,12407],{},[5093,12404,12405],{},[778,12406,5107],{},[5093,12408,12409],{},"撰写财报报告、构建投资论点、覆盖行业研究",[5078,12411,12412,12416],{},[5093,12413,12414],{},[778,12415,5117],{},[5093,12417,12418],{},"寻源交易、开展尽调、起草IC备忘录、监控组合KPI",[5078,12420,12421,12425],{},[5093,12422,12423],{},[778,12424,5127],{},[5093,12426,12427],{},"准备客户评审、再平衡投资组合、构建财务规划",[758,12429,12430,12431,12434,12435,12434,12438,12434,12441,12444],{},"每个插件都新增斜杠命令，直接在Claude里运行——",[841,12432,12433],{},"/comps [公司]","、",[841,12436,12437],{},"/dcf [公司]",[841,12439,12440],{},"/earnings [公司] [季度]",[841,12442,12443],{},"/ic-memo [项目]","。输出到Excel（带实时公式）、PowerPoint（你的品牌模板）或Word文档。零代码，安装即用。",[758,12446,12447],{},"大多数报道略过了最值得关注的部分：技能文件。它们不是套着金融术语的通用提示词，而是编码了资深分析师处理可比公司分析或IC备忘录时实际运用的那种推理框架。读几个就能感受到设计用心。",[758,12449,12450],{},"但命令和技能文件都不是这次发布最有意思的部分。",[758,12452,12453],{},[3036,12454],{"alt":12455,"src":5159},"Repo架构：financial-analysis核心插件、四个专业插件、11个MCP数据连接器",[753,12457,12458],{"id":12458},"数据巨头们刚刚宣告了一个标准",[758,12460,12461],{},"FactSet。标普全球。晨星。穆迪。LSEG。PitchBook——11家机构级数据供应商，全部在同一个发布窗口内构建了原生MCP连接器。不是初创公司，是深度嵌入每家主要金融机构工作流、签着多年期企业合同的数据基础设施供应商。",[758,12463,12464],{},"当这么多行业巨头同时朝同一个方向移动，这是一个标准在宣告自己的存在。历史对照：当Bloomberg和FactSet在2010年代开放REST API，整整一代金融科技公司在其上构建。MCP是2026年版的那个时刻。",[758,12466,12467],{},"数据提供商是管道，Claude是水管工。",[758,12469,12470],{},[3036,12471],{"alt":12472,"src":5178},"MCP协议汇聚：11家机构级数据供应商全部接入MCP协议，经由Claude输出结果",[758,12474,12475],{},"在我的公司，问题不再是\"要不要用AI\"——那个已经有了答案。现在问的是：对于已经付费的数据供应商，开通MCP连接器要花多少钱？这是一个具体得多、可操作得多的对话。",[753,12477,12479],{"id":12478},"我真的跑了一遍dcf和客户报告","我真的跑了一遍——DCF和客户报告",[758,12481,12482],{},"读repo是一回事。我想知道在真实工作上它实际产出什么。",[758,12484,12485],{},"我跑了两个测试：一个DCF模型，一份客户报告。下面的文件都是真实的——Claude直接生成的产出，没有经过任何编辑，原样分享出来，方便你看插件实际能产出什么。",[758,12487,12488],{},[778,12489,12490],{},"DCF的产出：",[758,12492,12493],{},"回来的是一个真正的Excel工作簿——公式引用、关联单元格、内置敏感性分析表。结构遵循投行惯例：收入构建、利润率假设、终值、WACC计算、隐含估值区间。不是我会直接发出去的东西——有些假设我会调整，有几处会覆盖默认值——但脚手架是对的，从这个起点编辑比从零开始构建要快得多。",[758,12495,5202,12496,12501,12502,12504],{},[778,12497,12498],{},[2177,12499,12500],{"href":5207},"下载：AMZN DCF模型（Excel）"," — ",[841,12503,5212],{}," 的实际产出，生成后未作任何修改",[758,12506,12507],{},[3036,12508],{"alt":12509,"src":5219},"AMZN DCF模型 Excel 内页——情景假设、收入构建、WACC与敏感性分析表",[758,12511,12512],{},[778,12513,12514],{},"客户报告的产出：",[758,12516,12517],{},"报告输出（使用样本数据生成，非真实客户信息）同样结构清晰——执行摘要、关键指标、组合评述，格式针对客户场景。语言专业，结构标准。同样不是直接发出去的状态，但是扎实的初稿，把\"空白页问题\"完全消除了。",[758,12519,12520],{},[3036,12521],{"alt":12522,"src":5233},"客户报告产出——面向客户的格式化报告",[758,12524,5202,12525,12530],{},[778,12526,12527],{},[2177,12528,12529],{"href":5240},"下载：客户绩效报告（Excel）"," — 客户报告命令的实际产出，真实测试运行",[758,12532,12533],{},"我的核心感受：价值不在于最终输出是完美的，而在于起点是正确的。一个空白Excel和一个正确结构的工作簿是两件完全不同的事。一个空白文档和一份格式化的初稿是两件完全不同的事。前者是一项任务，后者是一个编辑工作。从空白页到结构化起点之间的那个间隙——那是大多数时间实际花在哪里的地方。",[753,12535,12537],{"id":12536},"skills正在成为产品本身","Skills正在成为产品本身",[758,12539,12540],{},"用过之后，我开始思考一个比这些具体插件更宏观的问题。",[758,12542,12543],{},"这里的架构——领域知识编码在Markdown文件里，组织为技能和命令，通过JSON连接数据源——是一个将远远超出金融服务范围而泛化的模式。软件不是Python代码，不是API调用。软件是Markdown。",[758,12545,12546],{},"这是对\"构建\"意味着什么的有意义的转变。面向技能的项目设计——核心IP存在于结构化的、人类可读的领域知识文件中，而不是在代码里——将变得越来越普遍。对于很多知识工作自动化来说，未来的代码库可能主要就是一套结构良好的Markdown文件。",[758,12548,12549],{},[3036,12550],{"alt":12551,"src":5264},"面向技能的编程——专业知识编码在结构化文件中，而非代码里",[758,12553,12554],{},"由此衍生出几点：",[758,12556,12557,12560],{},[778,12558,12559],{},"技能文件本身就是护城河。"," 任何人都可以安装这个repo。差异化来自叠加在上面的公司专属技能——专有的交易框架、编码为约束条件的内部观点、被正式化为技能文件的机构沟通风格。那些被编码和版本控制的机构知识，比任何SaaS订阅都更难复制。",[758,12562,12563,12566],{},[778,12564,12565],{},"围绕技能配置会有真实的咨询市场。"," 把一个组织的实际流程翻译成结构良好的领域知识文件是一门技艺。一个认真配置这些文件的机构——把真实的工作流转化为精心设计的技能——得到的东西和安装默认配置就放在那里的机构是质的不同。随着这个模式扩散，知道如何做好那个翻译工作的人会有需求。",[758,12568,12569,12572],{},[778,12570,12571],{},"\"只是Markdown\"是对本周怀疑声音的错误解读。"," 流传的一个批评是：这些技能\"只是提示词\"，用更便宜的模型可以更低成本复制。技术上正确，对企业买家实际上无关紧要。机构不会去雇人为41个金融工作流编写并维护自定义系统提示。这个repo里的文件是预置的、有版本控制的、经过专业设计的领域知识。\"只是Markdown\"是特性，不是缺陷——任何理解这个工作流的资深人员都可以读懂、验证、扩展，无需碰一行代码。",[753,12574,12575],{"id":12575},"我内部实际上在考虑的事",[758,12577,12578],{},"基于这次使用体验，我在认真考虑把Claude Code加上技能集分别推给我们的销售团队和研发团队。",[758,12580,12581],{},"技能格式让这对非技术团队是可行的。我不需要构建自定义工具，不需要培训新界面，不需要维护集成。我写技能文件来编码我们的具体工作流，配置我们已经有合同的数据连接器，团队就有了一个从第一天就理解我们领域的AI协作者。",[758,12583,12584],{},"销售团队：编码我们产品、典型客户画像、提案格式、常见异议的技能集。",[758,12586,12587],{},"研发团队：编码我们研究框架、数据源、文档标准的技能集。",[758,12589,12590],{},"当前活在人脑中的机构知识——随着员工离开而流失的那些——被编码并变得可查询。这是和通用AI助手不同量级的价值。",[753,12592,12594],{"id":12593},"一个真实的例子从搭建-agentic-工作流到写一个-skill-文件","一个真实的例子：从搭建 Agentic 工作流到写一个 Skill 文件",[758,12596,12597],{},"这个转变对我来说通过一次具体的对话变得非常清晰——就在这次发布前几周，我和客户团队的同事在讨论一个问题。",[758,12599,12600],{},"我们想解决的问题：自动化个性化投资建议。一个客户带着特定的投资组合、特定的目标、特定的风险偏好来找我们。以前，产出一份符合公司标准的、有针对性的投资建议需要大量手工工作——拉数据、梳理配置逻辑、按我们的格式起草文档。",[758,12602,12603],{},"最初的方案是从头搭建一个 Agentic 工作流。设计数据管道、写编排逻辑、构建输出模板、接入我们的系统、测试、维护。真正的工程工作——几周的开发周期，持续的维护成本，需要专人负责。",[758,12605,12606],{},"读完这些插件之后，我意识到用一个 Skill 文件就能实现大部分需求。",[758,12608,12609],{},"一个技能文件，编码我们的投资建议逻辑——我们的配置框架、客户沟通标准、必要的合规披露、输出格式。一个命令，拉取客户的投资组合数据，通过这套推理逻辑处理。输出：一份结构化的投资建议文档，按我们的标准格式化，准备好供顾问审阅。",[758,12611,12612],{},"我利用财富管理技能，用一个样本客户档案跑了一遍——产出不言自明。",[758,12614,12615],{},[3036,12616],{"alt":12617,"src":5332},"产出：基于技能文件生成的个性化投资建议，按公司标准格式化",[758,12619,5202,12620,12625],{},[778,12621,12622],{},[2177,12623,12624],{"href":5339},"下载：投资建议（Markdown）"," — 针对样本客户档案的技能实际产出，未作修改",[758,12627,12628],{},"让我意外的不是输出的质量，而是构建的速度。我原本估计是几周的工程项目，变成了几个小时写一个结构良好的 Markdown 文件。技能文件本身对团队里任何资深人员都是可读的——可以被审阅、质疑、迭代，不需要碰一行代码。",[758,12630,12631],{},"从\"我们需要搭建一个 Agentic 系统\"到\"我们需要写一个技能文件\"——这个间隙就是范式的转变。不只是针对投资建议，任何可以被编码为结构化推理的工作流，现在都属于技能文件，而不是定制化的 Agentic 管道。",[758,12633,12634],{},[3036,12635],{"alt":12636,"src":5353},"范式转变：搭建 Agentic 工作流（数周工程量）vs 写一个技能文件（数小时）",[753,12638,12639],{"id":12639},"什么会真正被采用",[758,12641,12642],{},"几点从技术和实际使用角度的判断：",[758,12644,12645,12648,12649,12651],{},[778,12646,12647],{},"会快速落地的："," 已知格式的交付物，且数据订阅已经到位。财报摘要、可比公司分析表、研究报告初稿、客户报告。增量工作量：安装插件，配置",[841,12650,5369],{},"指向已付费的供应商，完成。几个月内，而不是几年。",[758,12653,12654,12657],{},[778,12655,12656],{},"会慢的："," 完整模型直接进入客户材料。合规和风险团队需要在模型架构和审核流程上签字后，AI生成的DCF才能进路演PPT。这不是Claude的问题——这是Excel模板上就已经存在的管控问题，当AI介入时更加正式。",[758,12659,12660,12663,12664,12666],{},[778,12661,12662],{},"真正的阻力是数据，不是工具。"," 如果你的公司没有FactSet或Capital IQ的合同，MCP连接器无法充分释放价值。但开源性质改变了中型机构的可行性计算：fork这个repo，配置",[841,12665,5369],{},"指向你有合同的那一家供应商，加入你公司自己的模型和术语。以前需要软件项目的事，现在需要一个会编辑Markdown的人。",[758,12668,12669,12672],{},[778,12670,12671],{},"在你准备好部署之前就开始合规对话。"," 技术工作量比预期的轻。组织层面的工作——合规审批、IT安全、数据治理——才是时间线真正的所在。等到技术准备好再开始，已经损失了两三个月。",[758,12674,12675],{},[3036,12676],{"alt":12677,"src":5397},"快速落地 vs 需要更长时间——中型金融机构的采用现实框架",[753,12679,12680],{"id":12680},"蓝图",[758,12682,12683],{},"Anthropic本周发布的不只是插件，而是一份设计规范——AI原生金融机构的工具层应该是什么样子的。",[758,12685,12686],{},"架构：带共享建模基础设施的核心插件、各职能专用的附加模块、机构级数据的MCP连接器、直接输出到Excel和PowerPoint。模块化，数据连通，工作流专用，Markdown可定制。",[758,12688,12689],{},"我越来越觉得技能文件模式才是这里持久的贡献，而不是任何具体的命令或产出。领域知识编码在结构化的、有版本控制的、人类可读的文件中，该领域任何足够资深的人都可以扩展。这是一个真正新的软件基元。",[758,12691,12692],{},"现在就把机构知识编码进技能文件的机构——他们的交易框架、研究方法论、客户沟通标准——在构建一种复利增长的东西。不只是生产力工具，而是一个随时间变得更有用的组织知识系统。",[758,12694,12695],{},"等待供应商把这些打包成授权产品的机构会付出三倍代价，拿到一个已经落后一代的解决方案。这个模式在金融技术领域发生过，正在以更快的速度再次发生。",[758,12697,12698],{},[3036,12699],{"alt":12700,"src":5422},"蓝图——AI原生金融机构架构：顶层为技能文件中的机构知识，中层为核心基础设施与专业模块，底层为数据连接器，产出至Excel/PowerPoint/Word",[753,12702,12704],{"id":12703},"下载-本文引用的真实产出文件","下载 — 本文引用的真实产出文件",[758,12706,12707],{},"以下文件均为本文测试过程中Claude生成的原始产出，未作任何修改，供参考。",[5072,12709,12710,12723],{},[5075,12711,12712],{},[5078,12713,12714,12717,12720],{},[5081,12715,12716],{},"文件",[5081,12718,12719],{},"格式",[5081,12721,12722],{},"对应章节",[5088,12724,12725,12737,12748,12760],{},[5078,12726,12727,12732,12734],{},[5093,12728,12729],{},[2177,12730,12731],{"href":5207},"AMZN DCF模型",[5093,12733,5456],{},[5093,12735,12736],{},"我实际跑了一遍",[5078,12738,12739,12744,12746],{},[5093,12740,12741],{},[2177,12742,12743],{"href":5240},"客户绩效报告",[5093,12745,5456],{},[5093,12747,12736],{},[5078,12749,12750,12755,12757],{},[5093,12751,12752],{},[2177,12753,12754],{"href":5477},"投资建议",[5093,12756,5456],{},[5093,12758,12759],{},"从Agentic工作流到技能文件",[5078,12761,12762,12766,12768],{},[5093,12763,12764],{},[2177,12765,12754],{"href":5339},[5093,12767,5492],{},[5093,12769,12759],{},[2163,12771],{},[758,12773,12774],{},[2156,12775,12776],{},"我写的是在金融服务和技术产品交叉地带构建AI原生——从一家真实机构内部看是什么样的。Newsletter在这里：",[758,12778,12779],{},[2156,12780,5506,12781],{},[2177,12782,5511],{"href":5509,"rel":12783},[2181],{"title":839,"searchDepth":708,"depth":708,"links":12785},[12786,12787,12788,12789,12790,12791,12792,12793,12794],{"id":12374,"depth":708,"text":12374},{"id":12458,"depth":708,"text":12458},{"id":12478,"depth":708,"text":12479},{"id":12536,"depth":708,"text":12537},{"id":12575,"depth":708,"text":12575},{"id":12593,"depth":708,"text":12594},{"id":12639,"depth":708,"text":12639},{"id":12680,"depth":708,"text":12680},{"id":12703,"depth":708,"text":12704},"Anthropic本周发布了financial-services-plugins——五个Claude插件连接11个机构级数据源。我是一家金融公司的技术产品负责人，用它跑了真实的DCF模型和客户报告。以下是我实际发现的东西。",{"date":5525,"image":5526,"alt":12797,"tags":12798,"category":3417,"published":859},"Anthropic金融AI技术栈蓝图架构图——顶层为Claude + Skills，中层为四个专业模块，底层为11个机构级数据连接器",[12799,11623,12800,12801,12802],"人工智能","Anthropic","MCP协议","金融分析","/blogs/zh/anthropic-finance-plugins-insider-take",{"title":12363,"description":12795},"blogs/zh/21.anthropic-finance-plugins-insider-take","OTzis86rvEUhZHybsbEkMxoPJ00Urx0p6RgXLv3wFTY",{"id":12808,"title":12809,"body":12810,"description":14077,"extension":2126,"meta":14078,"navigation":859,"ogImage":6861,"path":14085,"seo":14086,"stem":14087,"__hash__":14088},"zh/blogs/zh/22.chatgpt-explained-kitchen-metaphor.md","没有魔法：ChatGPT 到底怎么工作的？一间厨房讲清楚",{"type":750,"value":12811,"toc":14047},[12812,12821,12823,12826,12829,12832,12835,12838,12841,12846,12848,12851,12854,12860,12863,12874,12877,12880,12883,12885,12889,12892,12895,12898,12905,12911,12917,12924,12929,12936,12938,12942,12945,12950,12956,12959,12962,12966,12981,12986,12989,12996,12998,13002,13008,13011,13014,13020,13025,13028,13030,13034,13040,13043,13054,13065,13068,13071,13076,13083,13085,13089,13092,13095,13098,13103,13110,13112,13116,13119,13124,13127,13134,13136,13139,13142,13148,13154,13157,13162,13164,13168,13174,13177,13188,13195,13198,13203,13209,13216,13218,13222,13225,13228,13234,13237,13240,13243,13250,13255,13259,13265,13278,13281,13284,13287,13367,13370,13375,13378,13381,13398,13401,13409,13415,13422,13429,13431,13435,13438,13458,13461,13464,13468,13471,13477,13483,13489,13492,13549,13555,13558,13563,13570,13572,13576,13579,13596,13599,13603,13609,13626,13631,13634,13638,13641,13644,13647,13654,13656,13660,13663,13750,13755,13758,13761,13764,13770,13776,13779,13785,13792,13794,13798,13801,13804,13807,13810,13813,13819,13822,13825,13827,13836,13838,13841],[758,12813,12814],{},[2156,12815,12816,12817,12820],{},"灵感来自 Andrej Karpathy 的 ",[2177,12818,5551],{"href":5549,"rel":12819},[2181]," —— 200 行代码，证明这一切不过是一间厨房的事。",[2163,12822],{},[758,12824,12825],{},"你第一次走进一家 Omakase 料理店。",[758,12827,12828],{},"服务员引你坐到吧台前。没有菜单。主厨点点头，开始了。",[758,12830,12831],{},"一道道菜依次呈上 —— 先是刺身，然后是热的，然后是带酸的，然后浓郁，然后清淡。每一道菜来得都像是理所当然。等甜品上桌的时候，你突然发现：你从头到尾没有一次想过「这道菜放在这里不对」。整顿饭像是专门为你设计的，出自某个懂你却从未见过你的人。",[758,12833,12834],{},"上周读完 Andrej Karpathy（OpenAI 联合创始人、特斯拉前 AI 负责人）的文章之后，这个问题就一直在我脑子里转：主厨是怎么「知道」的？",[758,12836,12837],{},"不是那位有二十年经验的老师傅。我是说 —— 任何一个厨师，从零开始，是怎么学会给一个素未谋面的陌生人设计出「感觉对」的上菜顺序的？没有人给他一本规则手册。他也没有把所有可能的套餐顺序都背下来。",[758,12839,12840],{},"答案是：小明。",[758,12842,12843],{},[3036,12844],{"alt":12845,"src":5586},"小明站在怀石料理台前，背后是整面墙的 4,192 个判断旋钮",[2163,12847],{},[758,12849,12850],{},"认识一下小明：一个学徒厨师，一面墙上的 4,192 个旋钮，和 32,000 份怀石料理记录。（4,192 不是为了好记凑出来的数字 —— 这是 Karpathy 那 200 行代码里真实的参数计数。）",[758,12852,12853],{},"那 32,000 份记录是前辈大师的全部作品集 —— 每一套怀石套餐，每一道菜，按顺序写下来。小明研究了全部。",[758,12855,12856,12857],{},"那 4,192 个旋钮是另一回事。它们在小明的脑子里。每个旋钮编码着他后天习得的一小片判断力：",[2156,12858,12859],{},"「第三道菜如果很浓郁，对第四道菜的选择影响有多大？开头几道以海鲜为主，中间是否应该换换口味？一整套下来，口感的变化节奏有多重要？」",[758,12861,12862],{},"4,192 个微小的直觉。合在一起，构成他全部的厨艺判断。",[758,12864,12865,12866,12869,12870,12873],{},"在这间厨房里，小明是唯一的决策者 —— 所有关于「下一道菜是什么」的判断，最终都由他做出。你在后面会遇到「副厨」和「师傅」，先做个简单介绍：",[778,12867,12868],{},"副厨","协助小明思考，出意见但不做决定；",[778,12871,12872],{},"师傅","是训练阶段专门帮小明调节旋钮的老手，练习结束后就离场了。正式上菜时，只有小明。",[758,12875,12876],{},"他从什么都不懂开始。每个旋钮都设在随机的位置。训练结束时，这些旋钮将会编码出近乎一位大师的品味。",[758,12878,12879],{},"这就是 ChatGPT 的工作原理。不是比喻。是字面意思。",[758,12881,12882],{},"Karpathy 用 200 行代码证明了这一点。今天，我带你走进这间厨房。",[2163,12884],{},[753,12886,12888],{"id":12887},"第一章食材-32000-套怀石记录","第一章：食材 —— 32,000 套怀石记录",[758,12890,12891],{},"小明有前辈大师的作品集：32,000 套完整的怀石套餐，每道菜按顺序记录在案。",[758,12893,12894],{},"「emma」是一套 4 道菜的套餐：e → m → m → a。「sophia」是 6 道菜：s → o → p → h → i → a。小明只能看到上了什么，以及上菜的顺序。没有食谱，没有解释，只有序列本身。",[758,12896,12897],{},"他的目标：研究足够多的套餐之后，设计出让食客说「感觉对」的新套餐。",[758,12899,12900,12901,12904],{},"有一点值得先说清楚：在这个简化的厨房里，菜单只有 27 道菜，每道菜用一个字母代号标注。「e」不是「鳗鱼寿司」—— 它只是 27 道菜里第 5 道菜的代号。这些代号本身没有任何具体含义，重要的只是",[2156,12902,12903],{},"顺序","：哪个代号跟在哪个后面，以及在数万套套餐里涌现出什么样的规律。",[758,12906,12907,12910],{},[778,12908,12909],{},"为什么顺序那么重要？"," 因为同样四道菜 —— 「刺身→烤物→汤→甜品」和「甜品→汤→烤物→刺身」，是完全不同的用餐体验。「emma」和「amme」包含相同的食材，却是完全不同的体验。就像「狗咬人」和「人咬狗」，同样四个字，意思天壤之别。",[758,12912,12913,12916],{},[778,12914,12915],{},"放大到 ChatGPT："," 它的作品集不是 32,000 个名字，而是整个互联网 —— 所有写过的书、文章、对话、帖子。万亿套「怀石记录」，全部按顺序消化。",[758,12918,12919,12920,12923],{},"关键感受：小明没有背下每一套套餐的序列。他形成了",[2156,12921,12922],{},"直觉"," —— 对「什么应该跟在什么后面」的感知，对「这个组合感觉对」的判断。没有规则手册，只有模式，反复吸收，直到变成本能。",[758,12925,12926],{},[3036,12927],{"alt":12928,"src":5674},"32,000 套怀石记录叠成一摞卷轴，其中一卷展开显示序列 e→m→m→a，旁边是 ChatGPT 的万亿条序列对比气泡",[3590,12930,12931],{},[758,12932,12933],{},[2156,12934,12935],{},"小明不是在记忆，他在培养品味。",[2163,12937],{},[753,12939,12941],{"id":12940},"第二章拆解-一道菜一道菜地学","第二章：拆解 —— 一道菜一道菜地学",[758,12943,12944],{},"一整套怀石太复杂，没法整套学。第一步是把它拆成一道道单独的菜。",[758,12946,12947,12948],{},"「emma」变成：",[844,12949,5696],{},[758,12951,12952,12953,12955],{},"那个服务铃 🛎️ 标记着每套套餐的开始和结束。每道单独的菜是一个 ",[778,12954,5702],{}," —— 厨房处理的最小意义单位。",[758,12957,12958],{},"但小明的厨房不认识菜名，它只认识数字。所以每道菜都要编号：",[758,12960,12961],{},"a = 0，b = 1，c = 2……z = 25，🛎️ = 26",[758,12963,12947,12964],{},[844,12965,5715],{},[758,12967,12968,12969,12972,12973,12976,12977,12980],{},"光有数字还不够。数字只是一个标签 —— 它告诉你",[2156,12970,12971],{},"是哪道","菜，却什么都没说明这道菜的",[2156,12974,12975],{},"性格","。所以每个编号还附带一张",[778,12978,12979],{},"风味档案卡","：16 个数字，描述这道菜的属性，比如浓郁度、温度、口感、酸度。这些档案一开始完全随机，在训练中逐渐变得有意义。",[758,12982,12983],{},[3036,12984],{"alt":12985,"src":5736},"分词流水线：「emma」→ 带服务铃的 token → 数字 26,4,12,12,0,26 → 附带 16 个格子的风味档案卡",[758,12987,12988],{},"ChatGPT 的编号系统更聪明 —— 常见的词语组合会拿到一个单独的编号，不用一个字母一个字母地拼。比如「hello」在 ChatGPT 的词汇表里是一道独立的菜，不需要拆成 h、e、l、l、o 五道分开处理。更高效，更有表达力。词汇量大约 10 万个「菜品」，而不是 27 个。",[3590,12990,12991],{},[758,12992,12993],{},[2156,12994,12995],{},"你打出的每一个字都变成数字，每个数字都携带一张风味档案。这就是机器开始「读懂」语言的方式。",[2163,12997],{},[753,12999,13001],{"id":13000},"第三章出菜流水线-下一道应该是什么","第三章：出菜流水线 —— 下一道应该是什么？",[758,13003,13004,13005],{},"这是一切汇聚的地方。小明在每个时刻只需要回答一个问题：",[778,13006,13007],{},"下一道菜应该是什么？",[758,13009,13010],{},"假设「emma」这套套餐已经上了前两道菜（e, m），现在第三道「m」刚刚到达厨房。小明的任务：决定第四道是什么。",[758,13012,13013],{},"这道菜从厨房入口出发，经过三个关键站点，另一端出来的是答案：",[834,13015,13018],{"className":13016,"code":13017,"language":3579},[3577],"第三道菜「m」到达\n        ↓\n[1] 回头看 —— 副厨圆桌  ← 唯一回头的地方\n        ↓\n[2] 想清楚 —— 后厨加工\n        ↓\n[3] 下注 —— 最终投票\n        ↓\n27 道菜的概率排名\n",[841,13019,13017],{"__ignoreMap":839},[758,13021,13022],{},[3036,13023],{"alt":13024,"src":5776},"三站式出菜流水线：副厨圆桌 → 后厨加工 → 最终投票，输出为 27 道菜的概率排名",[758,13026,13027],{},"整件事分三步。每一步之间，有两个幕后动作悄悄发生。我们先把三步说清楚，再说那两个配角。",[2163,13029],{},[826,13031,13033],{"id":13032},"第一步回头看-副厨圆桌","第一步：回头看 —— 副厨圆桌",[758,13035,13036,13037,11959],{},"这是整条流水线上",[778,13038,13039],{},"唯一一个会看之前菜品的地方",[758,13041,13042],{},"第三道菜「m」被放到桌子中央。四位副厨围坐在四周，每人负责追踪这顿饭到目前为止的一个不同维度：",[804,13044,13045,13048,13051],{},[775,13046,13047],{},"副厨甲追踪浓郁度，翻开笔记本：「第一道清淡，第二道浓郁……」",[775,13049,13050],{},"副厨乙追踪口感：「连续两道都是软的 —— 该来点有嚼劲的了？」",[775,13052,13053],{},"副厨丙追踪温度，副厨丁追踪酸碱感。",[758,13055,13056,13057,13060,13061,13064],{},"每位副厨提一个问题：",[2156,13058,13059],{},"「按我追踪的维度，前面哪道菜对现在的决定最有参考价值？」"," 之前两道菜逐一回应：",[2156,13062,13063],{},"「这是我的情况。」"," 匹配度高的，把详细信息传递过来；匹配度低的，直接跳过。",[758,13066,13067],{},"四位副厨把各自的发现汇总，传向下一站。",[758,13069,13070],{},"有一点值得一提：副厨不需要重新品尝之前的菜。从第一道菜开始，他们就把每道菜的信息记在了笔记本里。每道新菜只需要新增一页 —— 这是这间厨房的记忆方式，记过的不用重来。",[758,13072,13073],{},[3036,13074],{"alt":13075,"src":5825},"俯视视角的副厨圆桌：四位副厨（甲=浓郁度，乙=口感，丙=温度，丁=酸度）围坐，中央盘子上写着「第三道菜：m」",[3590,13077,13078],{},[758,13079,13080],{},[2156,13081,13082],{},"副厨圆桌做的事只有一件：告诉小明，过去发生的哪些事，和下一道菜的决定最有关。",[2163,13084],{},[826,13086,13088],{"id":13087},"第二步想清楚-后厨加工","第二步：想清楚 —— 后厨加工",[758,13090,13091],{},"副厨圆桌问完了该问的。现在，小明要自己想清楚。",[758,13093,13094],{},"后厨有一个工作习惯：先把问题完全打开 —— 在一个宽阔得多的空间里，同时考虑所有可能的方向。然后，品控把显然行不通的方向全部划掉（凡是出现负面信号的，直接清零）。最后，把结论收拢回来，提炼成清晰的判断。",[758,13096,13097],{},"打开，筛选，收拢。结果是一份精炼的「对第四道菜该是什么的理解」。",[758,13099,13100],{},[3036,13101],{"alt":13102,"src":5853},"后厨工作法三步：全面打开（发散的箭头） → 品控筛选（划掉的路径） → 收拢判断（汇聚成单一箭头）",[3590,13104,13105],{},[758,13106,13107],{},[2156,13108,13109],{},"副厨圆桌是向外看：历史告诉我什么？后厨是向内看：我自己怎么判断？",[2163,13111],{},[826,13113,13115],{"id":13114},"第三步下注-最终投票","第三步：下注 —— 最终投票",[758,13117,13118],{},"后厨加工完成。小明对所有 27 道可能的下一道菜分别打分：",[758,13120,13121],{},[2156,13122,13123],{},"「第四道是 'a' 的可能性 70%，是 'e' 的可能性 15%，是 'i' 的可能性 8%……」",[758,13125,13126],{},"原始分数转换成百分比，加起来等于 100%。这不是一个确定的答案 —— 是一个有把握的押注。",[3590,13128,13129],{},[758,13130,13131],{},[2156,13132,13133],{},"小明不知道答案，他在下注。全部魔法，在于他怎样越押越准。",[2163,13135],{},[826,13137,13138],{"id":13138},"幕后的两个配角",[758,13140,13141],{},"整个流程里还有两个动作，不是主角，但少了它们厨房会出问题。",[758,13143,13144,13147],{},[778,13145,13146],{},"漱口（每个主站点之前）："," 每次进入副厨圆桌或后厨之前，先漱口、重置味觉。这保证每次判断的起点是一致的 —— 刚吃完特别辣的东西马上品精细料理，感知会失真。漱口让每一次判断都从同一个基准出发。",[758,13149,13150,13153],{},[778,13151,13152],{},"混入一勺原味（每个主站点之后）："," 副厨圆桌结束后、后厨加工结束后，各有一个动作：把这道菜最初的原始档案混回去一勺。就像做酱汁浓缩时，永远在旁边留一锅原汤，每次收汁后加回来一点 —— 防止菜品经过一次次加工后，把自己原本的样子完全丢失。",[758,13155,13156],{},"在小明的厨房（MicroGPT）里，这三步走一遍就结束了。ChatGPT 的厨房把同样的三步叠加了几十次：第一轮的输出，是第二轮的输入，同样的副厨和后厨，每一轮理解得更深一层。",[758,13158,13159],{},[3036,13160],{"alt":13161,"src":5914},"两个幕后配角：漱口（每个站点前重置到中性基准）和混入一勺原味（每个站点后把原始档案混回去）",[2163,13163],{},[753,13165,13167],{"id":13166},"第四章食客的打分-你做得有多差","第四章：食客的打分 —— 你做得有多差？",[758,13169,13170,13171],{},"小明设计了一套怀石套餐。每道菜上桌，都有一个评分机制：",[778,13172,13173],{},"小明对正确答案给了多大的把握？",[758,13175,13176],{},"假设下一道菜实际上是「a」：",[804,13178,13179,13182,13185],{},[775,13180,13181],{},"小明给「a」押了 100% 的把握 —— 零惩罚，满分",[775,13183,13184],{},"小明给「a」只押了 10% —— 大惩罚",[775,13186,13187],{},"小明给「a」只押了 0.1% —— 巨大惩罚",[758,13189,13190,13191,13194],{},"这个惩罚就是",[778,13192,13193],{},"损失值（Loss）","。损失越低，厨艺越好。",[758,13196,13197],{},"起始损失：3.3。这是在 27 个选项里纯随机猜测的水平。一个完全的厨房新手，闭着眼睛随机摆盘。",[758,13199,13200],{},[3036,13201],{"alt":13202,"src":5950},"食客打分卡：三种把握程度对应的惩罚，以及损失值从 3.3 降至 2.37 的成长曲线",[758,13204,13205,13206],{},"整个训练过程只有一个使命：",[778,13207,13208],{},"把这个数字降下去。",[3590,13210,13211],{},[758,13212,13213],{},[2156,13214,13215],{},"小明的每一个决定都可以用一个数字来评判。这个数字是北极星。",[2163,13217],{},[753,13219,13221],{"id":13220},"第五章追根溯源-是谁让菜变咸了","第五章：追根溯源 —— 是谁让菜变咸了？",[758,13223,13224],{},"这是整件事的核心 —— 也是让整个系统真正变得聪明的地方。",[758,13226,13227],{},"分数出来了：「太差。」小明面对那面墙上的 4,192 个旋钮，不知道从哪儿下手。一个一个试？要几年。",[758,13229,13230,13231],{},"但他可以做一件更聪明的事：",[778,13232,13233],{},"从盘子开始往回追。",[758,13235,13236],{},"食客吃了一口：「太咸了。」",[826,13238,13239],{"id":13239},"从盘子追到厨房",[758,13241,13242],{},"「这口太咸了 → 最后一步是摆盘（没加盐）→ 再往前是后厨加工（加了酱油）→ 再往前是副厨圆桌（参考了第二道菜的档案）→ 追回到旋钮 #347（酱油浓度调节）。」",[758,13244,13245,13246,13249],{},"这就是",[778,13247,13248],{},"反向传播（Backpropagation）"," —— 从结果出发，沿着因果链条往回追到源头。",[758,13251,13252],{},[3036,13253],{"alt":13254,"src":6004},"反向传播链：「太咸了」← 摆盘 ← 后厨加工 ← 副厨圆桌 ← 旋钮 #347「酱油浓度调节」",[826,13256,13258],{"id":13257},"接力传递链式法则-chain-rule","接力传递（链式法则 Chain Rule）",[758,13260,13261,13262],{},"每个检查站只需要知道一件简单的事：",[2156,13263,13264],{},"「如果我的输入变化了一点点，我的输出会变化多少？」",[804,13266,13267,13270,13273],{},[775,13268,13269],{},"把旋钮 #347 调大 1 → 汤底咸度增加 3",[775,13271,13272],{},"汤底咸度增加 1 → 最终口味变化 2",[775,13274,13275,13276],{},"旋钮 #347 对最终口味的总影响 = 3 × 2 = ",[778,13277,6028],{},[758,13279,13280],{},"沿着流水线走，在每一步相乘。这就是链式法则 —— 不需要微积分，只需要沿路相乘。",[826,13282,13283],{"id":13283},"六种基本烹饪手法",[758,13285,13286],{},"整间厨房只用六种烹饪操作。每一个检查站、每一处计算，最终都归结为这六种之一。更关键的是：每种手法都能精确地追溯自身 —— 这正是「从盘子往回追」能一路追到每个旋钮的原因。",[5072,13288,13289,13299],{},[5075,13290,13291],{},[5078,13292,13293,13296],{},[5081,13294,13295],{},"手法",[5081,13297,13298],{},"在厨房里做什么",[5088,13300,13301,13312,13323,13334,13345,13356],{},[5078,13302,13303,13309],{},[5093,13304,13305,13308],{},[778,13306,13307],{},"合并","（加法）",[5093,13310,13311],{},"把两样东西倒在一起",[5078,13313,13314,13320],{},[5093,13315,13316,13319],{},[778,13317,13318],{},"融合","（乘法）",[5093,13321,13322],{},"两种食材相互放大",[5078,13324,13325,13331],{},[5093,13326,13327,13330],{},[778,13328,13329],{},"浓缩","（幂次）",[5093,13332,13333],{},"把风味集中",[5078,13335,13336,13342],{},[5093,13337,13338,13341],{},[778,13339,13340],{},"提取","（对数）",[5093,13343,13344],{},"一点点就够用",[5078,13346,13347,13353],{},[5093,13348,13349,13352],{},[778,13350,13351],{},"发酵","（指数）",[5093,13354,13355],{},"指数级增长",[5078,13357,13358,13364],{},[5093,13359,13360,13363],{},[778,13361,13362],{},"品控","（ReLU）",[5093,13365,13366],{},"好的通过，差的扔掉",[758,13368,13369],{},"ChatGPT 的整间厨房只用这六种手法。没有第七种。",[758,13371,13372],{},[3036,13373],{"alt":13374,"src":6127},"六种烹饪手法 2×3 图格：合并、融合、浓缩、提取、发酵、品控，每个配手绘烹饪图标",[826,13376,13377],{"id":13377},"一个具体的例子",[758,13379,13380],{},"小明做一道简单的两步菜：",[804,13382,13383,13386,13392],{},[775,13384,13385],{},"材料：2 份盐，3 份糖",[775,13387,13388,13389,13391],{},"第一步：",[778,13390,13318],{}," → 咸甜底味 = 6",[775,13393,13394,13395,13397],{},"第二步：",[778,13396,13307],{}," → 最终口味 = 底味 + 额外一撮盐 = 8",[758,13399,13400],{},"食客说味道不对。往回追：",[804,13402,13403,13406],{},[775,13404,13405],{},"最终口味 = 底味 + 盐 → 合并 → 底味的影响 = 1，盐的影响 = 1",[775,13407,13408],{},"底味 = 盐 × 糖 → 融合 → 盐在这里的影响 = 糖的量 = 3",[758,13410,13411,13412,13414],{},"盐出现了两次（融合里一次，合并里一次），总影响 = 3 + 1 = ",[778,13413,6168],{},"。糖只出现一次：影响 = 2。",[758,13416,13417,13418,13421],{},"现在小明知道了：盐对结果的影响是糖的两倍。这就是",[778,13419,13420],{},"梯度（Gradient）= 4"," 的意思。调盐要小心，调糖可以大胆一点。",[3590,13423,13424],{},[758,13425,13426],{},[2156,13427,13428],{},"小明不是在猜该转哪个旋钮，他在精确计算每个旋钮有多重要。",[2163,13430],{},[753,13432,13434],{"id":13433},"第六章师傅的训练方法-一千道菜","第六章：师傅的训练方法 —— 一千道菜",[758,13436,13437],{},"小明开始苦练。每道菜，同样三个步骤：",[772,13439,13440,13446,13452],{},[775,13441,13442,13445],{},[778,13443,13444],{},"出菜（前向传播）"," —— 用当前旋钮位置，把当前菜品过一遍流水线",[775,13447,13448,13451],{},[778,13449,13450],{},"追溯（反向传播）"," —— 食客打分后，从盘子开始往回追，找出每个旋钮的影响数值",[775,13453,13454,13457],{},[778,13455,13456],{},"调整（参数更新）"," —— 根据每个旋钮的影响数值，往「减少惩罚」的方向拨一点",[758,13459,13460],{},"前两步有明确的规则可循。第三步才是真正的难题：4,192 个旋钮，该怎么调？每个都拨一样的幅度吗？哪些要轻，哪些要重？激进地大幅调整，还是保守地小步试探？",[758,13462,13463],{},"这件事，小明一个人搞不定。所以训练时，旁边站着一位专门负责调节的老手，他叫师傅。",[826,13465,13467],{"id":13466},"调节大师-师傅","调节大师 师傅",[758,13469,13470],{},"不是每个旋钮都接受同样幅度的调整。师傅比「一刀切」聪明得多。",[758,13472,13473,13476],{},[778,13474,13475],{},"记忆（动量 Momentum）："," 师傅记得最近几轮的调整方向。如果过去五轮都说「把旋钮 #347 调小」，第六轮就更大胆地调。就像一个球滚下坡 —— 它积累势头，而不是来回弹跳。",[758,13478,13479,13482],{},[778,13480,13481],{},"因材施教（自适应学习率）："," 有些旋钮极度敏感 —— 轻轻一拨就引发巨大变化。师傅对这些旋钮轻手轻脚。有些旋钮很迟钝 —— 大力拨动才有小小变化。师傅使劲拧。",[758,13484,13485,13488],{},[778,13486,13487],{},"越来越轻（学习率衰减）："," 训练初期调整大胆 —— 你离好还很远，激进一点没关系。越接近优秀，调整越精细。越接近完美，越需要如履薄冰。",[826,13490,13491],{"id":13491},"成长曲线",[5072,13493,13494,13507],{},[5075,13495,13496],{},[5078,13497,13498,13501,13504],{},[5081,13499,13500],{},"第几道菜",[5081,13502,13503],{},"损失值",[5081,13505,13506],{},"厨房里发生了什么",[5088,13508,13509,13519,13529,13539],{},[5078,13510,13511,13514,13516],{},[5093,13512,13513],{},"第 1 道",[5093,13515,6271],{},[5093,13517,13518],{},"蒙眼摆盘 —— 刺身后面上甜品，全程油炸，一片混乱",[5078,13520,13521,13524,13526],{},[5093,13522,13523],{},"第 100 道",[5093,13525,6282],{},[5093,13527,13528],{},"学会了「怀石一般先清淡，收尾浓郁」",[5078,13530,13531,13534,13536],{},[5093,13532,13533],{},"第 500 道",[5093,13535,6293],{},[5093,13537,13538],{},"学会了「味噌和白饭总是成对出现」「不能连上三道油炸」",[5078,13540,13541,13544,13546],{},[5093,13542,13543],{},"第 1000 道",[5093,13545,6304],{},[5093,13547,13548],{},"设计出食客真正信任的套餐",[758,13550,13551,13554],{},[778,13552,13553],{},"小明从未记住过一条规则。"," 没有人告诉他「不能连上三道油炸」，也没有人解释「味噌配白饭」。他只是一遍遍地出菜→追溯→调整，直到旋钮自然地稳定在那些反映这些规律的位置上。",[758,13556,13557],{},"这就是「习得直觉」的意思。无师自通，不是因为天才，而是因为重复。",[758,13559,13560],{},[3036,13561],{"alt":13562,"src":6322},"训练循环：出菜→打分→追溯→调整（1,000 轮），旁边是师傅的三种方法：记忆、因材施教、越来越轻",[3590,13564,13565],{},[758,13566,13567],{},[2156,13568,13569],{},"智慧不是从规则里来的，是从重复里来的。",[2163,13571],{},[753,13573,13575],{"id":13574},"第七章出师-小明独自掌厨","第七章：出师 —— 小明独自掌厨",[758,13577,13578],{},"训练结束。旋钮锁定在最终位置。不再是练习 —— 真实的食客已经就坐。",[772,13580,13581,13584,13587,13590,13593],{},[775,13582,13583],{},"服务铃响起 🛎️ —— 「开始一套新的怀石」",[775,13585,13586],{},"小明根据当前旋钮位置设计第一道菜",[775,13588,13589],{},"第一道菜过一遍流水线，从中推导出第二道应该是什么",[775,13591,13592],{},"第二道推导出第三道，第三道推导出第四道",[775,13594,13595],{},"一直到小明自然地产出收尾铃声 🛎️",[758,13597,13598],{},"每一道菜都是即兴的 —— 没有预先写好的菜单 —— 但都建立在之前所有菜的基础上。就像真正的怀石主厨在读桌 —— 第一道菜塑造了第二道，整顿饭的氛围决定了结尾。",[826,13600,13602],{"id":13601},"冒险旋钮温度-temperature","冒险旋钮（温度 Temperature）",[758,13604,13605,13606,11959],{},"厨房门口有一个旋钮，它不影响厨艺水平，只影响",[2156,13607,13608],{},"风格",[804,13610,13611,13616,13621],{},[775,13612,13613,13615],{},[778,13614,6374],{},"（极度保守）→ 永远选最安全的选项 → 稳，但无聊",[775,13617,13618,13620],{},[778,13619,6380],{},"（略有冒险）→ 基本连贯，偶有惊喜",[775,13622,13623,13625],{},[778,13624,6386],{},"（大胆冒进）→ 可能惊艳，也可能一塌糊涂",[758,13627,13628],{},[3036,13629],{"alt":13630,"src":6393},"冒险旋钮温度表：0.1（极度保守）到 0.7（ChatGPT 标星）到 1.5（大胆冒进），三档各配对应风格的出菜示例",[758,13632,13633],{},"ChatGPT 大约在 0.7。有一点创意，但不会失控。",[826,13635,13637],{"id":13636},"关于发明菜品幻觉-hallucination","关于「发明菜品」（幻觉 Hallucination）",[758,13639,13640],{},"训练结束后，小明生成了「Karia」、「Yeran」、「Liole」—— 这些是 Karpathy 的模型实际输出的名字，听起来像真实的人名，感觉合理，好像某个地方确实有人叫这个名字。但大多数可能根本不存在于任何训练数据中。",[758,13642,13643],{},"他没有在撒谎。他只是在沿着习得的规律延伸 —— 在统计上看起来合理的组合。他满怀把握地呈上来，但它们从未真实存在过。",[758,13645,13646],{},"当 ChatGPT 自信地引用一篇不存在的论文，或者说出一个不存在的日期时，它做的事和小明发明「karia」完全一样。它没有事实核查站，只知道什么「味道对」，不知道什么「是真实的」。",[3590,13648,13649],{},[758,13650,13651],{},[2156,13652,13653],{},"小明可以用一种从未被收获过的食材，设计出一套无懈可击的怀石套餐。这是这间厨房的工作方式决定的代价。",[2163,13655],{},[753,13657,13659],{"id":13658},"第八章从街边小摊到米其林三星","第八章：从街边小摊到米其林三星",[758,13661,13662],{},"小明用 4,192 个判断旋钮和 32,000 套训练记录学会了做菜。ChatGPT 呢？",[5072,13664,13665,13677],{},[5075,13666,13667],{},[5078,13668,13669,13671,13674],{},[5081,13670],{},[5081,13672,13673],{},"小明（街边怀石）",[5081,13675,13676],{},"ChatGPT（米其林三星）",[5088,13678,13679,13690,13701,13712,13723,13734],{},[5078,13680,13681,13684,13687],{},[5093,13682,13683],{},"判断旋钮",[5093,13685,13686],{},"4,192 个",[5093,13688,13689],{},"数千亿个",[5078,13691,13692,13695,13698],{},[5093,13693,13694],{},"训练记录",[5093,13696,13697],{},"32,000 套名字序列",[5093,13699,13700],{},"万亿套序列（整个互联网）",[5078,13702,13703,13706,13709],{},[5093,13704,13705],{},"菜品词汇",[5093,13707,13708],{},"27 个字母 + 一个铃",[5093,13710,13711],{},"约 10 万个词块",[5078,13713,13714,13717,13720],{},[5093,13715,13716],{},"厨房",[5093,13718,13719],{},"一口灶（MacBook）",[5093,13721,13722],{},"数千口灶并行运转（GPU 集群）",[5078,13724,13725,13728,13731],{},[5093,13726,13727],{},"训练时长",[5093,13729,13730],{},"1 分钟",[5093,13732,13733],{},"数个月",[5078,13735,13736,13741,13746],{},[5093,13737,13738],{},[778,13739,13740],{},"烹饪原理",[5093,13742,13743],{},[778,13744,13745],{},"完全相同",[5093,13747,13748],{},[778,13749,13745],{},[758,13751,13752],{},[3036,13753],{"alt":13754,"src":6525},"规模对比：小明的街边小摊（4,192 个旋钮、MacBook、32,000 套记录）vs ChatGPT 的米其林厨房（数千亿旋钮、数千块 GPU、万亿条记录）",[758,13756,13757],{},"同一间厨房，天壤之别的规模。",[826,13759,13760],{"id":13760},"三星厨房的额外步骤",[758,13762,13763],{},"ChatGPT 没有止步于基础训练。两个额外阶段让它从技术上正确变成真正好用。",[758,13765,13766,13769],{},[778,13767,13768],{},"第一阶段 —— 换菜单（有监督微调 SFT）："," 小明先在简单的名字序列上训练，建立起基础的「顺序感」。然后换成多轮对话 —— 复杂的来回交流 —— 继续训练。同样的出菜→追溯→调整算法，不同的训练材料。就像厨师先把鸡蛋做到炉火纯青，再去研习法餐。基本功不变，曲目扩展。",[758,13771,13772,13775],{},[778,13773,13774],{},"第二阶段 —— 请来食评家（基于人类反馈的强化学习 RLHF）："," 小明做两道菜，食评家选出更好的那道。根据食评家的偏好调整旋钮，反复数百万次。",[758,13777,13778],{},"这就是为什么 ChatGPT「有礼貌」、「乐于助人」—— 不是因为有人写了一条规则说「要有礼貌」。而是因为人类食评家，在数百万次比较中，持续选择了那些感觉有帮助、有分寸的回应。那些偏好被烤进了旋钮里。",[758,13780,13781,13782],{},"贯穿始终，核心从未改变：",[778,13783,13784],{},"出菜 → 打分 → 追溯 → 调整。",[3590,13786,13787],{},[758,13788,13789],{},[2156,13790,13791],{},"从 200 行代码到数千亿参数：同样的六种手法，同样的三个步骤。只是更多旋钮，更多菜品，更多食客。",[2163,13793],{},[753,13795,13797],{"id":13796},"这间厨房说白了","这间厨房，说白了",[758,13799,13800],{},"下次有人告诉你 AI 很神秘、很可怕、或者即将取代人类 —— 想起那个坐在吧台前的食客。",[758,13802,13803],{},"没有菜单，菜品一道道端上来，一切都显得理所当然。",[758,13805,13806],{},"这背后：一个学徒，一面墙的判断旋钮，32,000 套记录。没人给他规则，只有出菜→打分→追溯→调整。做足够多次，旋钮自然稳定下来。",[758,13808,13809],{},"这不是智慧，是被打磨得足够精细、以至于与品味无法区分的统计直觉。",[758,13811,13812],{},"但这么简单的机制 —— 六种操作，三个步骤，一个循环 —— 居然能产出一个可以聊天、解释概念、写文章、让你感觉到「有人在」的东西？",[758,13814,13815,13818],{},[2156,13816,13817],{},"这才","是真正令人叹为观止的地方。",[758,13820,13821],{},"200 行代码。六种手法。一间厨房。",[758,13823,13824],{},"这就是 ChatGPT 的全部。",[2163,13826],{},[758,13828,13829],{},[2156,13830,13831,13832,13835],{},"想看那实际的 200 行，Andrej Karpathy 的 ",[2177,13833,5551],{"href":5549,"rel":13834},[2181]," 值得一读。他建了这间厨房，我只是给它起了个名字，带你转了一圈。",[2163,13837],{},[753,13839,13840],{"id":13840},"比喻对照表",[5072,13842,13843,13853],{},[5075,13844,13845],{},[5078,13846,13847,13850],{},[5081,13848,13849],{},"GPT 概念",[5081,13851,13852],{},"厨房比喻",[5088,13854,13855,13863,13871,13879,13887,13895,13902,13910,13918,13926,13933,13941,13948,13956,13963,13971,13978,13986,13993,14000,14008,14016,14024,14032,14040],{},[5078,13856,13857,13860],{},[5093,13858,13859],{},"模型（Model）",[5093,13861,13862],{},"小明，学徒厨师",[5078,13864,13865,13868],{},[5093,13866,13867],{},"参数（Parameters）",[5093,13869,13870],{},"4,192 个判断旋钮 —— 编码小明习得的顺序直觉",[5078,13872,13873,13876],{},[5093,13874,13875],{},"训练数据（Dataset）",[5093,13877,13878],{},"前辈大师的 32,000 套怀石记录（不是食谱，只有顺序）",[5078,13880,13881,13884],{},[5093,13882,13883],{},"Token（词元）",[5093,13885,13886],{},"怀石套餐里的一道菜",[5078,13888,13889,13892],{},[5093,13890,13891],{},"分词器（Tokenizer）",[5093,13893,13894],{},"把一整套套餐拆解成一道道单菜的系统",[5078,13896,13897,13899],{},[5093,13898,6673],{},[5093,13900,13901],{},"服务铃 🛎️ —— 「新套餐开始」/「套餐结束」",[5078,13903,13904,13907],{},[5093,13905,13906],{},"词向量（Embedding）",[5093,13908,13909],{},"风味档案卡 —— 16 个数字编码一道菜的性格",[5078,13911,13912,13915],{},[5093,13913,13914],{},"位置编码（Position Embedding）",[5093,13916,13917],{},"序列标签 —— 「这是第 3 道菜」",[5078,13919,13920,13923],{},[5093,13921,13922],{},"注意力机制（Attention）",[5093,13924,13925],{},"副厨查阅笔记本，参考之前的菜品",[5078,13927,13928,13930],{},[5093,13929,6705],{},[5093,13931,13932],{},"Q = 「我现在需要什么？」K = 「我之前提供过这个」V = 「如果你选我，这是我的食谱」",[5078,13934,13935,13938],{},[5093,13936,13937],{},"多头注意力（Multi-head Attention）",[5093,13939,13940],{},"4 位副厨同时查阅，各自追踪不同维度",[5078,13942,13943,13945],{},[5093,13944,6721],{},[5093,13946,13947],{},"笔记本 —— 不需要重新品尝已经记录过的内容",[5078,13949,13950,13953],{},[5093,13951,13952],{},"MLP（前馈网络）",[5093,13954,13955],{},"后厨 —— 向同事请教之后，独立思考",[5078,13957,13958,13960],{},[5093,13959,6737],{},[5093,13961,13962],{},"品控 —— 差的清零，好的通过",[5078,13964,13965,13968],{},[5093,13966,13967],{},"残差连接（Residual Connection）",[5093,13969,13970],{},"每一步之后混回去的那一勺原汤",[5078,13972,13973,13975],{},[5093,13974,6753],{},[5093,13976,13977],{},"每个检查站之前的漱口",[5078,13979,13980,13983],{},[5093,13981,13982],{},"损失（Loss）",[5093,13984,13985],{},"食客的惩罚分 —— 离完美有多远？",[5078,13987,13988,13990],{},[5093,13989,6768],{},[5093,13991,13992],{},"把原始评分转换成概率排名",[5078,13994,13995,13997],{},[5093,13996,13248],{},[5093,13998,13999],{},"沿着流水线往回追「太咸」追到旋钮 #347",[5078,14001,14002,14005],{},[5093,14003,14004],{},"梯度（Gradient）",[5093,14006,14007],{},"每个旋钮的影响系数 —— 它对最终结果的推动力有多大？",[5078,14009,14010,14013],{},[5093,14011,14012],{},"链式法则（Chain Rule）",[5093,14014,14015],{},"沿着流水线逐步相乘",[5078,14017,14018,14021],{},[5093,14019,14020],{},"Adam 优化器",[5093,14022,14023],{},"师傅 —— 有记忆、因材施教、越来越轻手的调节大师",[5078,14025,14026,14029],{},[5093,14027,14028],{},"温度（Temperature）",[5093,14030,14031],{},"冒险旋钮 —— 低 = 保守，高 = 创意",[5078,14033,14034,14037],{},[5093,14035,14036],{},"幻觉（Hallucination）",[5093,14038,14039],{},"发明「karia」—— 统计上合理，从未真实存在",[5078,14041,14042,14044],{},[5093,14043,6824],{},[5093,14045,14046],{},"街边小摊 → 米其林三星 —— 同样的原理，天壤之别的规模",{"title":839,"searchDepth":708,"depth":708,"links":14048},[14049,14050,14051,14057,14058,14064,14068,14072,14075,14076],{"id":12887,"depth":708,"text":12888},{"id":12940,"depth":708,"text":12941},{"id":13000,"depth":708,"text":13001,"children":14052},[14053,14054,14055,14056],{"id":13032,"depth":110,"text":13033},{"id":13087,"depth":110,"text":13088},{"id":13114,"depth":110,"text":13115},{"id":13138,"depth":110,"text":13138},{"id":13166,"depth":708,"text":13167},{"id":13220,"depth":708,"text":13221,"children":14059},[14060,14061,14062,14063],{"id":13239,"depth":110,"text":13239},{"id":13257,"depth":110,"text":13258},{"id":13283,"depth":110,"text":13283},{"id":13377,"depth":110,"text":13377},{"id":13433,"depth":708,"text":13434,"children":14065},[14066,14067],{"id":13466,"depth":110,"text":13467},{"id":13491,"depth":110,"text":13491},{"id":13574,"depth":708,"text":13575,"children":14069},[14070,14071],{"id":13601,"depth":110,"text":13602},{"id":13636,"depth":110,"text":13637},{"id":13658,"depth":708,"text":13659,"children":14073},[14074],{"id":13760,"depth":110,"text":13760},{"id":13796,"depth":708,"text":13797},{"id":13840,"depth":708,"text":13840},"你听说 ChatGPT 很神奇。其实不是。它是一个叫小明的学徒厨师，一面墙的 4,192 个判断旋钮，以及一个不断重复的循环。没有公式，没有代码，只有一间厨房。",{"date":6860,"image":6861,"alt":14079,"tags":14080,"category":2132,"youtube":6868,"published":859},"学徒厨师小明站在日式怀石料理厨房中，面对一整面墙的 4,192 个判断旋钮",[6864,14081,14082,14083,14084],"AI解析","机器学习","Transformer","深度学习","/blogs/zh/chatgpt-explained-kitchen-metaphor",{"title":12809,"description":14077},"blogs/zh/22.chatgpt-explained-kitchen-metaphor","_K6FTpP3hHvbGiWzTjbFakm6LQLN8MKuclIAnSSB96Q",{"id":14090,"title":14091,"body":14092,"description":14459,"extension":2126,"meta":14460,"navigation":859,"ogImage":7254,"path":14465,"seo":14466,"stem":14467,"__hash__":14468},"zh/blogs/zh/23.i-engineered-the-law-of-attraction-with-ai.md","我用 AI 工程化了吸引力法则",{"type":750,"value":14093,"toc":14451},[14094,14097,14100,14103,14106,14108,14111,14114,14117,14120,14123,14126,14131,14134,14136,14140,14143,14169,14172,14175,14180,14183,14189,14235,14238,14243,14246,14248,14252,14255,14258,14261,14267,14273,14279,14285,14290,14293,14319,14330,14337,14340,14345,14347,14350,14353,14356,14371,14374,14379,14385,14388,14390,14393,14396,14402,14405,14408,14410,14414,14417,14420,14431,14434,14444,14446],[758,14095,14096],{},"每天早上，在我打开 WhatsApp 或看手机之前，有一样东西已经在我的收件箱里等着我了。",[758,14098,14099],{},"那是一份摘要。大约 15 到 20 条内容——论文、帖子、讨论串、项目——每一条都附有 AI 生成的摘要和一个相关性评分。高分的是\"必读\"，其余的我可以扫一眼或直接跳过——而那些值得留存的，往往会进入我的写作和博客文章。",[758,14101,14102],{},"这套系统，我叫它 Signal，全天候运行。它监控 Hacker News、Reddit、ArXiv、X、Product Hunt、GitHub Trending 以及其他几个来源。每一条内容都会被 AI 根据我的兴趣、目标和知识盲区打分。每小时重新扫描一次，每天早上准时推送。",[758,14104,14105],{},"我建了 Signal，因为我相信吸引力法则。但我想要一个更好的版本。",[2163,14107],{},[753,14109,14110],{"id":14110},"那个我无法忽视的规律",[758,14112,14113],{},"我从来不是一个会把吸引力法则当成\"心灵鸡汤\"来驳斥的人。我亲身经历了太多次，不可能怀疑它。",[758,14115,14116],{},"大学时，我想创业。没有资金，没有商业计划，没有明确的方向。我有的只是执念——我疯狂地读创业类书籍，不停地谈创业，把自己浸泡在一群正在做事的人里。几个月后，我和朋友在校园里开了一家乐器店。不是我想象中的那个生意，但确实是一门生意。",[758,14118,14119],{},"后来，我想离开中国，在海外重新生活。没有清晰的路径，只是不断地往前走——学习、结识人脉、抓住每一个可以走的小步。慢慢地，一扇扇门打开了。我移民到了加拿大。",[758,14121,14122],{},"再后来，我想进入金融圈。我抓住每一个能靠近这个世界的机会——对话、职位、项目。最终，我成为了一家金融公司的合伙人。",[758,14124,14125],{},"这些没有一件是直线完成的。也没有一件是单靠愿望实现的。但所有这些都遵循着同一个模式：",[758,14127,14128],{},[778,14129,14130],{},"强烈的专注 → 深度沉浸 → 一系列\"巧合\"的叠加 → 抓住这些巧合采取行动。",[758,14132,14133],{},"我不只是想要某些东西。我让自己彻底浸泡在那个世界里。然后，事情自己来找我了。",[2163,14135],{},[753,14137,14139],{"id":14138},"吸引力法则究竟是什么","吸引力法则，究竟是什么",[758,14141,14142],{},"剥去所有神秘主义的包装，机制其实很清晰：",[772,14144,14145,14151,14157,14163],{},[775,14146,14147,14150],{},[778,14148,14149],{},"意图"," — 你定义什么对你重要",[775,14152,14153,14156],{},[778,14154,14155],{},"注意力"," — 你让自己暴露在相关的信息和人脉中",[775,14158,14159,14162],{},[778,14160,14161],{},"模式识别"," — 你的大脑开始过滤世界，寻找相关信号",[775,14164,14165,14168],{},[778,14166,14167],{},"行动"," — 你抓住那些\"并非巧合的巧合\"采取行动",[758,14170,14171],{},"第三步有真实的神经科学支撑。大脑的网状激活系统（RAS）每秒处理约 1100 万比特的信息——但意识层面只能处理约 40 比特。RAS 就是那个过滤器，它决定什么进入意识，什么被忽略。",[758,14173,14174],{},"当你设定一个清晰的意图，你就在重新编程这个过滤器。突然间，你开始注意到那篇一直存在的文章、那个你早就该认识的人、那个就放在眼前的机会。",[758,14176,14177],{},[3036,14178],{"alt":14179,"src":6964},"RAS：意图如何重新编程你大脑的过滤器，从每秒 1100 万比特降到约 40 比特",[758,14181,14182],{},"这不是魔法。这是信号获取。",[758,14184,14185,14186],{},"而让我恍然大悟的是：",[778,14187,14188],{},"这和推荐算法的工作方式完全一样。",[5072,14190,14191,14201],{},[5075,14192,14193],{},[5078,14194,14195,14198],{},[5081,14196,14197],{},"吸引力法则",[5081,14199,14200],{},"推荐系统",[5088,14202,14203,14211,14219,14227],{},[5078,14204,14205,14208],{},[5093,14206,14207],{},"设定意图",[5093,14209,14210],{},"定义用户画像",[5078,14212,14213,14216],{},[5093,14214,14215],{},"让自己暴露在信息中",[5093,14217,14218],{},"摄入内容流",[5078,14220,14221,14224],{},[5093,14222,14223],{},"RAS 过滤相关性",[5093,14225,14226],{},"排序模型打分",[5078,14228,14229,14232],{},[5093,14230,14231],{},"\"巧合\"浮现",[5093,14233,14234],{},"个性化信息流呈现",[758,14236,14237],{},"吸引力法则是一台生物推荐引擎。而和所有推荐引擎一样，它有一个硬性上限——带宽、处理速度、你能用来阅读和连接的有限时间。",[758,14239,14240],{},[3036,14241],{"alt":14242,"src":7028},"吸引力法则 vs 推荐算法——同一个四步结构",[758,14244,14245],{},"AI 没有这些限制。",[2163,14247],{},[753,14249,14251],{"id":14250},"signal-如何运作","Signal 如何运作",[758,14253,14254],{},"2025 年初，我开始认真学习 AI——智能体、大语言模型、整个技术版图。一个主题反复出现：最顶尖的 AI 研究者们正把大量资金投入到能够建立更丰富世界内部表征的系统中，让机器能像人类一样过滤和推理。Yann LeCun 离开 Meta 去创建 AMI Labs，Fei-Fei Li 以 50 亿美元估值为 World Labs 融资。方向已经很清楚：AI 的未来不只是生成文字，而是建模什么是重要的。",[758,14256,14257],{},"我不是在试图构建 AGI，我只是想为自己建一个更好的过滤器。Signal 就这样诞生了。",[758,14259,14260],{},"架构很简单：",[758,14262,14263,14266],{},[778,14264,14265],{},"信息源"," → Hacker News、Reddit、ArXiv、X/Twitter、Product Hunt、GitHub Trending、Lobsters",[758,14268,14269,14272],{},[778,14270,14271],{},"打分"," → 每条内容都由 LLM 根据我的个人兴趣画像评估：我当前的关注领域、我正在尝试回答的问题、我想补足的知识盲区",[758,14274,14275,14278],{},[778,14276,14277],{},"输出"," → 高分内容是\"必读\"，进入我的晨间摘要；其余内容可查阅，但优先级较低",[758,14280,14281,14284],{},[778,14282,14283],{},"节奏"," → 每小时重新扫描一次，每天早上推送一次",[758,14286,14287],{},[3036,14288],{"alt":14289,"src":7076},"Signal 系统架构——信息源、AI 打分和晨间摘要",[758,14291,14292],{},"这套系统实际上做到了：",[804,14294,14295,14301,14307,14313],{},[775,14296,14297,14300],{},[778,14298,14299],{},"将我的注意力带宽扩大了约 100 倍。"," 我通过 AI 摘要每天\"阅读\"500 多条内容，再决定哪些值得深度阅读。",[775,14302,14303,14306],{},[778,14304,14305],{},"消除了噪音循环。"," 不再无休止地刷手机，不再被为平台留存时长而非为我的成长而设计的算法干扰。",[775,14308,14309,14312],{},[778,14310,14311],{},"制造幸运的巧合。"," 我否则不会发现的论文，不会知道去找的人，不会联系起来的想法。",[775,14314,14315,14318],{},[778,14316,14317],{},"随时间复利增长。"," 我把兴趣画像定义得越清晰，信号就越精准。",[758,14320,14321,14322,14325,14326,14329],{},"这正是技术界所说的从",[778,14323,14324],{},"注意力经济","向",[778,14327,14328],{},"意图经济","的转变。",[758,14331,14332,14333,14336],{},"注意力经济——我们熟悉的社交媒体——优化的是平台参与度，它捕获你的注意力并出售。意图经济则不同：这些系统为",[2156,14334,14335],{},"你的","目标而优化，而不是平台的指标。",[758,14338,14339],{},"剑桥大学的研究人员已经发出警告：AI 很快将实时拦截你正在形成的意图，并将其卖给广告商。这场军备竞赛无论如何都要来——问题只是你是主动建立自己的系统，还是让别人帮你建。",[758,14341,14342],{},[3036,14343],{"alt":14344,"src":7131},"注意力经济 vs 意图经济——算法在为谁服务？",[2163,14346],{},[753,14348,14349],{"id":14349},"飞轮效应",[758,14351,14352],{},"传统的吸引力法则是被动的：设定意图，等待宇宙送来。这个框架一直让我不满意，它把太多东西交给了偶然性。",[758,14354,14355],{},"AI 增强版是一个主动飞轮：",[758,14357,14358,14360,14361,14364,14365,14367,14368,14370],{},[778,14359,7148],{}," → 捕获你领域中重要的信息\n",[778,14362,14363],{},"知识"," → 随着每天处理精选内容而不断加深\n",[778,14366,14167],{}," → 更有质量——更好的判断、更敏锐的写作、更准确的对话\n",[778,14369,7148],{}," → 随着行动产生新的连接、新的信息源、新的反馈而持续改进",[758,14372,14373],{},"一个具体的例子：Signal 在 2025 年底开始密集地推送世界模型相关研究——LeCun 的论文、Jim Fan 的机器人工作、AMI Labs 的早期报道。我深入研究了这个方向，开始以当时还不普遍的方式谈论世界模型的应用场景。这让我进入了一些否则不会参与的对话，而这些对话又带来了新的信息源，反哺回 Signal。",[758,14375,14376],{},[3036,14377],{"alt":14378,"src":7168},"Signal 飞轮——Signal → 知识 → 行动 → Signal",[758,14380,13245,14381,14384],{},[778,14382,14383],{},"工程化的意外","——刻意设计让有价值的\"偶然\"更频繁发生的条件。",[758,14386,14387],{},"不是宇宙在派送。而是你建了一个让送达更可能、更频繁、更精准的系统。",[2163,14389],{},[753,14391,14392],{"id":14392},"为什么现在很重要",[758,14394,14395],{},"我们生活在信息爆炸、注意力稀缺的时代。普通人每天接触数千条内容，却几乎没有留存——因为决定他们看什么的算法，优化的是平台参与度，不是个人成长。",[758,14397,14398,14399,14401],{},"在这个环境里胜出的人，不会是那些拥有最好 AI 工具的人。现在每个人都能用 GPT。真正有差异的是：你是否建立了一个以",[2156,14400,14335],{},"具体意图为服务对象的系统——持续地、自动地、在你睡觉时。",[758,14403,14404],{},"Signal 没有让我变得更聪明，它改变了我接触到的是什么。而随着时间的推移，你接触到的东西，塑造了你知道什么、认识谁、能做什么。",[758,14406,14407],{},"这不是显化。这是架构。",[2163,14409],{},[753,14411,14413],{"id":14412},"建立你自己的-signal","建立你自己的 Signal",[758,14415,14416],{},"我没有停止相信吸引力法则。我只是决定把它工程化。",[758,14418,14419],{},"我会问任何思考这件事的人：",[804,14421,14422,14425,14428],{},[775,14423,14424],{},"你现在想把什么吸引到你的生命里？",[775,14426,14427],{},"你每天需要持续看到什么样的信息才能实现它？",[775,14429,14430],{},"如果你不再把这件事交给一个不了解你的算法，而是建立一个真正了解你的系统，会发生什么？",[758,14432,14433],{},"你不需要相信吸引力法则才能从这里找到价值。你只需要相信更好的信号处理。",[758,14435,14436,14437,14440,14441,14443],{},"关于 Signal 的具体构建方式——架构、打分提示词、画像设计——我会在后续文章中分享。如果这是你想持续关注的内容，欢迎关注我。我在 ",[2177,14438,7232],{"href":7230,"rel":14439},[2181]," 写关于构建真正有效的 AI 原生系统的内容，也在 ",[778,14442,3956],{}," 通讯中持续更新。",[2163,14445],{},[758,14447,14448],{},[2156,14449,14450],{},"关注我，获取更多关于构建服务你意图——而非算法意图——的 AI 系统的内容。",{"title":839,"searchDepth":708,"depth":708,"links":14452},[14453,14454,14455,14456,14457,14458],{"id":14110,"depth":708,"text":14110},{"id":14138,"depth":708,"text":14139},{"id":14250,"depth":708,"text":14251},{"id":14349,"depth":708,"text":14349},{"id":14392,"depth":708,"text":14392},{"id":14412,"depth":708,"text":14413},"吸引力法则一直对我有效——所以我用 AI 把它变成了一套自动化系统。Signal 全天候监控 Hacker News、Reddit、ArXiv 等平台，根据我的个人兴趣画像为每条内容打分。这就是工程化的意外。",{"date":7253,"image":7254,"alt":14461,"tags":14462,"category":3016,"youtube":7260,"published":859},"一个人被发光的信号流包围，代表有意识的信息过滤系统",[2130,14463,14464,14328],"Signal系统","个人系统","/blogs/zh/i-engineered-the-law-of-attraction-with-ai",{"title":14091,"description":14459},"blogs/zh/23.i-engineered-the-law-of-attraction-with-ai","RZbTIatb0Ge2Ov-PyMXc4Uw6Xr8meH9RjjUjRZTGXnk",{"id":14470,"title":14471,"body":14472,"description":14558,"extension":2126,"meta":14559,"navigation":859,"ogImage":7363,"path":14565,"seo":14566,"stem":14567,"__hash__":14568},"zh/blogs/zh/24.in-the-age-of-ai-look-up-at-the-stars.md","越是 AI 时代，越要仰望星空",{"type":750,"value":14473,"toc":14551},[14474,14477,14480,14484,14487,14490,14494,14497,14500,14503,14508,14512,14515,14518,14521,14526,14530,14533,14536,14539,14542,14545,14548],[758,14475,14476],{},"最近在工作里，我看到一个很稳定的现象：同样的模型，同样的任务，不同人用出来的结果完全不同。有人反复改 prompt，最后还是泛泛而谈；有人一两轮就能把问题打穿，直接进入可用状态。",[758,14478,14479],{},"这件事表面看是 prompt 技巧的差距，实质上是认知资产的差距。AI 降低了执行成本，也抬高了判断力的价值。越是强大的工具，越会把人背后的业务理解、客户感、产品判断和战略视野放大出来。",[753,14481,14483],{"id":14482},"prompt-是业务判断的接口","Prompt 是业务判断的接口",[758,14485,14486],{},"好的 prompt 不是咒语，本质上是一个人对问题的拆解能力。它要求你先知道目标是什么，约束是什么，哪些 context 必须给，哪些细节会干扰，以及一个真正可用的答案应该长什么样。",[758,14488,14489],{},"这背后不是模板，而是行业经验、客户理解、产品感、审美和决策训练。高手不是会写更长的 prompt，而是脑子里有更清楚的模型。他在输入框里写下的几句话，其实是把自己的业务判断压缩成了一个接口。",[758,14491,14492],{},[3036,14493],{"alt":14483,"src":7291},[753,14495,14496],{"id":14496},"销售也是同一件事",[758,14498,14499],{},"销售里也有同样的现象。有的人见一次客户，机会基本就稳了；另一些人材料做得很全，会议开得很多，却始终推进不了。差别不在于谁更努力，而在于谁更快识别决策链、利益点、风险点和真正阻力。",[758,14501,14502],{},"表面看，这是一场 meeting 的差距。实际上，这是商业判断的差距。AI 只是把这种差距搬到了输入框里：同样一句问题，有的人问出来只是问题，有的人问出来已经带着方向、取舍和判断标准。",[758,14504,14505],{},[3036,14506],{"alt":14507,"src":7307},"销售判断与客户理解",[753,14509,14511],{"id":14510},"ai-会重新定价人的能力","AI 会重新定价人的能力",[758,14513,14514],{},"过去，很多执行能力本身就很值钱。写文案、做方案、写代码、做分析、整理资料，这些都需要时间，也需要专门训练。AI 之后，执行会越来越便宜；越便宜的东西，越不构成核心壁垒。",[758,14516,14517],{},"真正变贵的是另一层能力：问题定义、判断标准、业务 context、取舍能力、品味，以及对人的理解。AI 不是简单的平均器，而是杠杆。杠杆不会自动让所有人变得一样强，杠杆会放大操作者。",[758,14519,14520],{},"有判断的人，会更快验证、更快迭代、更快学习。判断弱的人，也会产出更多东西，但很多只是更多“像那么回事”的东西。这就是我认为 AI 时代最关键的变化之一：产出会越来越多，稀缺的是判断。",[758,14522,14523],{},[3036,14524],{"alt":14525,"src":7326},"AI 时代的认知半径与战略高度",[753,14527,14529],{"id":14528},"仰望星空不是浪漫是战略","仰望星空，不是浪漫，是战略",[758,14531,14532],{},"所以我说“越是 AI 时代，越要仰望星空”，不是文艺表达。这里的“星空”，指的是认知半径和战略高度。执行速度越快，方向感越值钱；内容生成越便宜，观点越值钱；信息越容易获得，判断越值钱。",[758,14534,14535],{},"对企业家来说，真正的问题不只是“怎么把 AI 用得更熟”。这个当然要做，但更底层的问题是：我有没有足够高的视野，判断什么值得做？有没有足够深的行业理解，判断机会在哪里？有没有足够强的客户感，判断什么是真需求，什么只是噪音？",[758,14537,14538],{},"AI 可以放大执行，但它不能替代方向。方向来自人，来自长期积累的认知资产，也来自对市场、客户和机会的持续判断。",[753,14540,14541],{"id":14541},"我的结论",[758,14543,14544],{},"接下来我对 AI 的投入，会分成两层。第一层是工具和 workflow，继续追求效率；第二层是更底层的认知资产，包括行业理解、商业洞察、客户感、产品判断、审美和历史感。",[758,14546,14547],{},"前者决定我能跑多快，后者决定我往哪里跑，以及 AI 到底能把我放大到什么程度。AI 时代最危险的不是不会用工具，而是把工具带来的流畅感，误认为自己的能力。",[758,14549,14550],{},"同样的模型，同样的输入框，最后拼的不是谁更会问，而是谁背后有更大的世界。",{"title":839,"searchDepth":708,"depth":708,"links":14552},[14553,14554,14555,14556,14557],{"id":14482,"depth":708,"text":14483},{"id":14496,"depth":708,"text":14496},{"id":14510,"depth":708,"text":14511},{"id":14528,"depth":708,"text":14529},{"id":14541,"depth":708,"text":14541},"越是 AI 时代，仰望星空越不是浪漫表达，而是企业家需要的认知半径、战略高度和判断力。",{"date":7362,"image":7363,"alt":14560,"tags":14561,"category":3417,"youtube":7369,"published":859},"复古商业插画：指南针、市场地图和 AI 输入框共同指向战略远方",[2130,14562,14563,14564],"创业","判断力","战略","/blogs/zh/in-the-age-of-ai-look-up-at-the-stars",{"title":14471,"description":14558},"blogs/zh/24.in-the-age-of-ai-look-up-at-the-stars","DmyatGIKDofJrx7qA8i8AFJ-ogNZVnyAMCiA6dlSPFE",{"id":14570,"title":14571,"body":14572,"description":14885,"extension":2126,"meta":14886,"navigation":859,"ogImage":7715,"path":14892,"seo":14893,"stem":14894,"__hash__":14895},"zh/blogs/zh/25.fable-5-managing-ai-autonomy.md","Fable 5 改变了 AI 工作的最小单位",{"type":750,"value":14573,"toc":14876},[14574,14577,14580,14583,14586,14599,14602,14605,14609,14612,14615,14618,14621,14624,14629,14633,14641,14644,14647,14650,14653,14658,14662,14668,14671,14674,14677,14680,14683,14686,14691,14695,14698,14701,14704,14707,14710,14713,14716,14719,14723,14726,14739,14742,14745,14748,14768,14771,14774,14785,14788,14793,14797,14800,14803,14806,14811,14816,14821,14826,14831,14836,14841,14846,14849,14852,14855,14858,14861,14864,14867,14870,14873],[758,14575,14576],{},"我给 Claude Fable 5 一个需要跑几个小时的任务。第一次，我感觉自己不是在 prompt 一个模型，而是在启动一次工作运行。",[758,14578,14579],{},"这个差别只有真正体验过才明显。普通 chat 模式里，我还在工作内部。我问一个问题，看结果，纠正它，补 context，再问一轮，几分钟就要 steer 一次。但 Fable 的感觉不一样。我给它一个目标，它开始规划、分发、研究、综合，最后交付了一组很漂亮的页面。表面结果是页面，更重要的结果是那次 run。",[758,14581,14582],{},"我觉得这才是关键。",[758,14584,14585],{},"Fable 5 不只是产出了更好的 output。它把 AI 工作的最小单位，从 response 变成了 run。",[758,14587,14588,14589,14593,14594,14598],{},"Anthropic 在 ",[2177,14590,14592],{"href":7398,"rel":14591},[2181],"2026 年 6 月 9 日发布 Fable 5 和 Mythos 5","，强调它们比之前的 Claude 更适合长时间自主工作。三天后，Anthropic ",[2177,14595,14597],{"href":7404,"rel":14596},[2181],"发布声明","，说美国政府发出 export-control directive，要求暂停 foreign nationals 对 Fable 5 和 Mythos 5 的访问。为了合规，实际结果是 Anthropic 暂时关闭了所有客户访问，并表示正在努力恢复。",[758,14600,14601],{},"发布和暂停访问，其实从两个方向说明了同一件事。发布说明：agent 可以跑得更久、更可靠地分发任务、更大范围地使用工具。暂停访问说明：一旦 agent 足够强，control、governance 和边界就不再是后置问题。",[758,14603,14604],{},"这篇文章真正想讲的，不是 Fable 很强。真正的启发是：当一个 agent 可以连续跑几个小时，稀缺能力就不再是写一个聪明 prompt，而是设计一个 operating contract，让这次 run 有边界、可检查、可回滚。",[753,14606,14608],{"id":14607},"最小单位从-response-变成-run","最小单位从 response 变成 run",[758,14610,14611],{},"Response 很小。它是一段文字，可能加一次 tool call，可能改一个文件。如果错了，我纠正它。如果不完整，我再问一次。错误成本通常是一轮对话。",[758,14613,14614],{},"Run 不一样。Run 有持续时间，有状态，会消耗 token，会调用工具，会读文件、写 artifacts、分发 subagents、形成中间假设，并且可能走很远之后我才检查结果。一个坏假设的成本，不再是一段坏答案，而可能是一整条错误工作分支。",[758,14616,14617],{},"所以更长的 autonomy 改变了注意力的经济学。在 prompt 模式里，我的注意力就是 control loop。在 run 模式里，我的注意力前移和后移。运行前，我要定义目标、context、约束、权限、checkpoint 和 stop condition。运行后，我要检查 evidence，再判断结果能不能用。",[758,14619,14620],{},"模型更强了，但工作系统反而更不能随便。如果目标模糊，autonomy 会放大模糊。如果约束不清楚，autonomy 会探索错误空间。如果 agent 能无边界地行动，autonomy 会把能力变成 operational risk。",[758,14622,14623],{},"所以 \"管理 agent\" 这个说法还不够准确。真正的工作是 run design。",[758,14625,14626],{},[3036,14627],{"alt":14628,"src":7438},"Response 和 run 是两种不同的 AI 工作单位",[753,14630,14632],{"id":14631},"fable-让-control-surface-变得可见","Fable 让 control surface 变得可见",[758,14634,14635,14636,14640],{},"Anthropic 自己的 ",[2177,14637,14639],{"href":7448,"rel":14638},[2181],"Fable prompting 文档"," 很有意思，因为它不只是讲 prompt。它讲 long runs、effort levels、progress claims、explicit boundaries、parallel subagents、memory systems 和 scaffolding changes。这已经是另一类指导。",[758,14642,14643],{},"如果一个模型可以跑几个小时，界面就不能只是一个 text box。界面需要 timeout behavior、progress indicators、asynchronous check-ins、基于 evidence 的状态报告，以及定义什么时候应该 pause 的机制。如果它可以分发 subagents，harness 就要决定什么时候 delegation 有用，subagents 如何沟通，它们的工作如何被 review。如果它可以维护 memory，系统就需要一个地方记录 lessons，同时避免坏假设污染后续工作。",[758,14645,14646],{},"换句话说，模型能力提升，会倒逼它周围的产品和 workflow 重构。",[758,14648,14649],{},"这也是我在 Fable 那次 run 里的真实感受。我不只是在看 \"模型答了什么\"。我其实在问另一组问题：它决定研究什么？它分发了什么？它带着哪些假设继续往前？它把 effort 花在哪里？什么 evidence 支撑最终结果？如果 run 跑偏，我应该在哪里介入？",[758,14651,14652],{},"这些不是 prompt 问题。这是 operating-system 问题。",[758,14654,14655],{},[3036,14656],{"alt":14657,"src":7469},"围绕 agent run 设计 operating contract",[753,14659,14661],{"id":14660},"superpowers-是一种日常演练","Superpowers 是一种日常演练",[758,14663,14664,14665,11959],{},"这也是为什么 Fable 的体验立刻让我想到 ",[2177,14666,7481],{"href":7479,"rel":14667},[2181],[758,14669,14670],{},"Superpowers 不是另一个模型。它是一个给 coding agents 用的开源 skills framework 和 software development methodology。它的价值不是让模型突然变聪明，而是在模型周围加上一套更有纪律的工作方式。",[758,14672,14673],{},"这个模式很简单：澄清目标，整理 spec，把 spec 变成 implementation plan，通过 subagents 或结构化阶段执行，review 工作，并在声明完成前验证。Skills 变成了 agent 的 reusable operating contract。",[758,14675,14676],{},"这会改变我日常使用 Opus 4.8 和 Codex 5.5 的方式。我可以给 agent 一个有意义的目标，让它工作一到两个小时，因为它不是在自由发挥。它有流程。它会读代码、写计划、按计划执行、检查自己的工作，再把 artifacts 交给我 review。",[758,14678,14679],{},"模型是 engine。Skills 是围绕 engine 的 operating discipline。",[758,14681,14682],{},"这个区别很重要，因为它防止我们得出错误结论。结论不是 \"Fable 是未来，所以等更强模型就行。\" 结论是：模型越强，周围的工作系统越重要。Superpowers 的价值在于，它让我们在每个模型都能跑半天之前，先练习这种习惯。它让 autonomy 变得可读、可控、可检查。",[758,14684,14685],{},"所以 \"skill\" 这个词比表面上更重要。一个好的 skill 不是 prompt template，而是一块可复用的 operating judgment：什么时候澄清，如何计划，什么叫 verification，什么时候用 subagents，什么时候要求 review，什么 evidence 足够支持 \"done\"。",[758,14687,14688],{},[3036,14689],{"alt":14690,"src":7505},"Superpowers 给模型加上一套 operating discipline",[753,14692,14694],{"id":14693},"这不只是-project-management","这不只是 project management",[758,14696,14697],{},"一个自然的反驳是：这听起来不就是 project management 吗？",[758,14699,14700],{},"某种程度上，是的。这正是重点。AI agents 越强，工作就越不像聊天，越像委托。委托一直都需要目标、context、review 和 accountability。",[758,14702,14703],{},"但 agent delegation 有三个差别，让旧习惯不够用。",[758,14705,14706],{},"第一，agents 以机器速度行动。一个人误解任务，通常很快会产生 friction：提问、延迟、明显偏差。Agent 可以安静地把大量 compute 花在错误理解上，然后交付一个看起来很完整、但方向错了的 artifact。",[758,14708,14709],{},"第二，agents 通过工具行动。它们不只是思考，还会读、写、调用 API、修改文件、发送消息，未来甚至可能花钱或发布内容。风险从 \"bad answer\" 变成了 \"bad action\"。",[758,14711,14712],{},"第三，agents 是没有原生 accountability 的概率性 worker。人可以在组织和社会关系里承担决定。Agent 可以产出工作，但责任仍然留在设计这次 run 的人或团队身上。",[758,14714,14715],{},"所以这不只是普通管理。它是面对一种新 worker 的管理：快、不累、有用、不稳定、能使用工具，但并不真正负责。",[758,14717,14718],{},"Operating contract 就是让这种 worker 变得可用的东西。",[753,14720,14722],{"id":14721},"governance-是同一个问题的放大版","Governance 是同一个问题的放大版",[758,14724,14725],{},"Fable 暂停访问这件事，把这个问题放到了地缘政治层面。但同一个模式在每个尺度都存在。",[758,14727,14728,14729,14733,14734,14738],{},"Anthropic 的声明讨论了 safeguards、red-teaming、jailbreak resistance、monitoring、customer data retention 和 defense in depth。Databricks 介绍 Fable 5 时，也把它放在 ",[2177,14730,14732],{"href":7546,"rel":14731},[2181],"governance、audit logs、tool-call policies 和 cost controls"," 里讲。Gartner 在 ",[2177,14735,14737],{"href":7552,"rel":14736},[2181],"2026 年 5 月"," 也提醒，不能对所有 AI agents 使用同一种 governance，因为组织必须区分 agent 的 autonomy level 和它被授予的 access scope。",[758,14740,14741],{},"这个区分非常关键。危险的组合不是 intelligence 本身，而是 autonomy 加 access，却没有相应的 control system。",[758,14743,14744],{},"一个只能 observe 的 agent，是一种风险。一个能 draft recommendation 的 agent，是另一种风险。一个可以在 approval 后行动的 agent，又是另一种风险。一个可以跨系统 autonomously act 的 agent，已经是完全不同的类别。",[758,14746,14747],{},"这同样适用于 individual builders。每次我让 agent 跑之前，都要知道自己授予了哪种 autonomy：",[804,14749,14750,14753,14756,14759,14762,14765],{},[775,14751,14752],{},"它只能读、总结和提建议吗？",[775,14754,14755],{},"它能改文件吗？",[775,14757,14758],{},"它能安装 dependencies 吗？",[775,14760,14761],{},"它能调用外部服务吗？",[775,14763,14764],{},"它能创建 branch、push code 或 publish content 吗？",[775,14766,14767],{},"它不确定时可以继续，还是必须在某些边界停下来？",[758,14769,14770],{},"这些问题不是 bureaucracy。它们就是工作的 control surface。",[758,14772,14773],{},"对每一次有意义的 agent run，我现在会问三个问题：",[772,14775,14776,14779,14782],{},[775,14777,14778],{},"它能做什么？",[775,14780,14781],{},"我怎么知道它做了什么？",[775,14783,14784],{},"如果它做错了，我怎么停止、回滚或修复？",[758,14786,14787],{},"如果答不上来，我就还没有设计这次 run。我只是启动了它。",[758,14789,14790],{},[3036,14791],{"alt":14792,"src":7610},"Agent governance 是 autonomy 和 access 的匹配问题",[753,14794,14796],{"id":14795},"真正的新能力是设计-operating-contracts","真正的新能力是设计 operating contracts",[758,14798,14799],{},"我觉得 AI-native work 正在走向这里。",[758,14801,14802],{},"Prompt 仍然重要。好的 prompt 是 judgment 的接口。但当工作最小单位变成 run，prompt 就不够了。Run 需要 operating contract。",[758,14804,14805],{},"这个 contract 不一定复杂。我现在会从七个部分想：",[758,14807,14808,14810],{},[778,14809,7628],{}," 什么结果重要？什么算成功？",[758,14812,14813,14815],{},[778,14814,7634],{}," 哪些文件、事实、docs、rules 或 examples 是 authoritative？",[758,14817,14818,14820],{},[778,14819,7640],{}," Agent 能读什么、写什么、调用什么、花费什么、发布什么？",[758,14822,14823,14825],{},[778,14824,7646],{}," 它应该在哪里 pause、report 或 ask for approval？",[758,14827,14828,14830],{},[778,14829,7652],{}," 什么 proof 足够支撑 progress claims 和 final claims？",[758,14832,14833,14835],{},[778,14834,7658],{}," 允许花多少时间、成本和探索空间？",[758,14837,14838,14840],{},[778,14839,7664],{}," 如果 run 产出错误结果，怎么处理？",[758,14842,14843],{},[3036,14844],{"alt":14845,"src":7671},"Agent run 的 operating contract stack",[758,14847,14848],{},"这比 \"prompt engineering\" 更有用。它更像设计一条小型工厂线。有些工作可以交出去，在后台运行。有些工作必须我深度参与。Leverage 来自知道这两者的区别。",[758,14850,14851],{},"这也是我最近更大的工作感受。AI 让我觉得自己更 powerful，因为更多项目可以同时动起来。我可以让一个 agent 起草，让另一个 agent research，让另一个 agent coding，让另一个 agent 把 plan 变成 distribution assets。但我的 focus 并没有变成无限。恰恰相反，focus 的价值更高了。",[758,14853,14854],{},"瓶颈从 \"亲自做每件事\"，变成了 \"正确路由工作\"。",[758,14856,14857],{},"这个环境里最强的 operator，不是盯着每一步的人，而是能把工作设计好的人：让该自动推进的部分自动推进，让重要判断仍然回到人这里。",[753,14859,14860],{"id":14860},"真正留下来的东西",[758,14862,14863],{},"Fable 5 可能会回来。它可能会变化。也可能很快被另一个模型替代。具体产品周期不是最重要的。",[758,14865,14866],{},"真正重要的是：AI work 正在从 prompt-response loop，走向更长时间、更可委托、更可检查的 runs。一旦这个变化发生，优势就会从 access 转向 orchestration。",[758,14868,14869],{},"最强模型当然仍然重要。但最强模型放在一个弱 operating system 里，可能只会产出昂贵的混乱。稍弱一点的模型，放在一个有纪律的 workflow 里，反而可能产出可交付的工作。",[758,14871,14872],{},"这就是我想继续练习的能力：不是只学会要 output，而是学会设计 agent 做真实工作的条件。",[758,14874,14875],{},"未来不属于最会 prompt 的人。它属于那些能让 execution 跑在前面，同时不让 judgment 掉队的 builders。",{"title":839,"searchDepth":708,"depth":708,"links":14877},[14878,14879,14880,14881,14882,14883,14884],{"id":14607,"depth":708,"text":14608},{"id":14631,"depth":708,"text":14632},{"id":14660,"depth":708,"text":14661},{"id":14693,"depth":708,"text":14694},{"id":14721,"depth":708,"text":14722},{"id":14795,"depth":708,"text":14796},{"id":14860,"depth":708,"text":14860},"Fable 5 让我看到，AI 工作正在从一轮 response 变成一次长时间 run；未来稀缺的能力，是为 agent run 设计边界、检查点和 operating contract。",{"date":7714,"image":7715,"alt":14887,"tags":14888,"category":2132,"youtube":7721,"published":859},"一个操作者在控制台前监督多条 AI 工作流",[2130,14889,14890,14891],"Agent","工作流","AI 治理","/blogs/zh/fable-5-managing-ai-autonomy",{"title":14571,"description":14885},"blogs/zh/25.fable-5-managing-ai-autonomy","KPbA461XaIC8kSk8LKlmMPhT6fLdILNZHMELKQp2QZs",{"id":14897,"title":14898,"body":14899,"description":15105,"extension":2126,"meta":15106,"navigation":859,"ogImage":7962,"path":15109,"seo":15110,"stem":15111,"__hash__":15112},"zh/blogs/zh/7.build-blog-in-one-day.md","快速开发我的博客",{"type":750,"value":14900,"toc":15097},[14901,14904,14910,14913,14915,14919,14922,14940,14943,14957,14963,14965,14969,14982,14998,15004,15006,15010,15013,15027,15030,15035,15038,15040,15044,15047,15054,15060,15062,15064,15067,15073,15079,15081,15086],[753,14902,14903],{"id":14903},"想法",[758,14905,14906,14907,11959],{},"最近，我终于完成了一个一直以来想做但又不断搁置的事情：",[778,14908,14909],{},"搭建自己的博客",[758,14911,14912],{},"我一直希望有一个属于自己的空间，可以分享人生的经历、记录工作的思考、沉淀学习的成果。朋友圈太快，社交平台太碎片，而博客就像一个数字笔记本，沉稳、自主、自由。写博客对我来说，不是为了流量，而是为了整理思维、留存价值，也给未来的自己一个可以回望的轨迹。",[2163,14914],{},[753,14916,14918],{"id":14917},"我尝试过-ai-工具","我尝试过 AI 工具",[758,14920,14921],{},"在动手之前，我也和很多人一样，对 AI 工具抱有期待。我先后试用了几个生成网站的产品，包括：",[804,14923,14924,14932],{},[775,14925,14926,14931],{},[2177,14927,14929],{"href":7759,"rel":14928},[2181],[778,14930,7763],{},"：UI 风格讨喜，上手门槛低，几分钟就能生成一个视觉不错的页面；",[775,14933,14934,14939],{},[2177,14935,14937],{"href":7769,"rel":14936},[2181],[778,14938,7773],{},"：非常适合快速生成结构化组件代码，特别是结合 shadcn/ui 时，交互非常自然。",[758,14941,14942],{},"但实际使用下来，我遇到了两个痛点：",[772,14944,14945,14951],{},[775,14946,14947,14950],{},[778,14948,14949],{},"细节难以控制","：页面初稿生成得很快，但每次我想做细节调整（比如布局微调、字体规范化、多语言支持），都要绕很多圈，还不如直接写；",[775,14952,14953,14956],{},[778,14954,14955],{},"免费计划限制大","：我用的是免费版本，导致代码每次生成后都不能持久保存，更新也有限制，让我很难持续开发。",[758,14958,14959,14960],{},"所以我最终决定：",[778,14961,14962],{},"AI 可以辅助我写代码，但网站的主架构还是要我自己来掌控。",[2163,14964],{},[753,14966,14968],{"id":14967},"我的最终方案nuxt-content-模块","我的最终方案：Nuxt + Content 模块",[758,14970,14971,14972,14977,14978,14981],{},"经过几轮比较，我选择了用 ",[2177,14973,14975],{"href":7809,"rel":14974},[2181],[778,14976,7813],{}," 搭配官方的 ",[2177,14979,7819],{"href":7817,"rel":14980},[2181]," 模块来自建博客。为什么？",[804,14983,14984,14987,14992,14995],{},[775,14985,14986],{},"Nuxt 是我熟悉的生态，拥有 Vue 的灵活性和 SSR 的强大能力；",[775,14988,14989,14991],{},[841,14990,7819],{}," 让我可以用 Markdown 写文章、结构化处理 frontmatter、自动生成目录、SEO 友好；",[775,14993,14994],{},"支持静态部署，非常适合发布到 Vercel 或 Netlify；",[775,14996,14997],{},"未来如果我想集成搜索、推荐系统、多语言切换、甚至嵌入 AI 辅助阅读，都可以自然扩展。",[758,14999,15000,15001,11959],{},"更重要的是，这套方案让我可以",[778,15002,15003],{},"完全掌控代码与样式结构，不受限于平台或模板框架",[2163,15005],{},[753,15007,15009],{"id":15008},"我用-ai-的方式vibe-coding-augment-code","我用 AI 的方式：vibe coding + Augment Code",[758,15011,15012],{},"虽然我没继续用 AI 一键建站工具，但我还是把 AI 融入了我的开发流程中。",[758,15014,15015,15016,15021,15022,15024,15025,11959],{},"我选择的是 ",[2177,15017,15019],{"href":7858,"rel":15018},[2181],[778,15020,7862],{},"，它不是替我生成整个页面，而是在我写代码的过程中提供“语义级协助”。",[7869,15023],{},"\n我把这个过程称为：",[778,15026,7874],{},[758,15028,15029],{},"它的体验更像是：",[3590,15031,15032],{},[758,15033,15034],{},"“我知道我要改这个卡片组件的排版，但不想一行一行写 CSS，你能帮我试几个版本看哪个最对味吗？”",[758,15036,15037],{},"在这种语境下，AI 不再是替我编码的机器人，而是一个“对设计有 sense 的搭档”。这大大提高了我对样式的控制精度，也让我更享受构建过程。",[2163,15039],{},[753,15041,15043],{"id":15042},"技术只是载体表达才是核心","技术只是载体，表达才是核心",[758,15045,15046],{},"搭博客其实不难，难的是持续写下去。",[758,15048,15049,15050,15053],{},"我在配置博客过程中最大的感受是：",[778,15051,15052],{},"不要追求“完美上线”，而是追求“能写”","。早期不必强求设计有多炫酷、结构有多复杂，先写起来，才是最重要的。",[758,15055,15056,15057,15059],{},"这个博客对我来说，不仅是一个内容平台，更像是一个长期的自我构建系统：",[7869,15058],{},"\n它连接了我生活中的思考、学习中的疑问、工作中的总结，也许未来还会是我构建新产品想法的“起点”。",[2163,15061],{},[753,15063,10381],{"id":10381},[758,15065,15066],{},"从今天起，这个博客不仅是我对过去的沉淀，也是我对未来的承诺。",[758,15068,15069,15070,15072],{},"我会在这里持续记录我的思考与成长：",[7869,15071],{},"\n关于工作，关于家庭，关于技术，关于人生。",[758,15074,15075,15076,15078],{},"不求华丽，只求真实。",[7869,15077],{},"\n愿它成为我持续表达、持续进化的一部分。",[2163,15080],{},[758,15082,7936,15083],{},[778,15084,15085],{},"未来计划预告：",[804,15087,15088,15091,15094],{},[775,15089,15090],{},"下一篇我将分享如何用 Nuxt + Tailwind 实现博客首页的渐进式增强",[775,15092,15093],{},"计划每月总结一次写作心得 + 博客优化迭代情况",[775,15095,15096],{},"正在探索将 GPT 与博客联动，实现“读者问答摘要 + 多语智能推荐”",{"title":839,"searchDepth":708,"depth":708,"links":15098},[15099,15100,15101,15102,15103,15104],{"id":14903,"depth":708,"text":14903},{"id":14917,"depth":708,"text":14918},{"id":14967,"depth":708,"text":14968},{"id":15008,"depth":708,"text":15009},{"id":15042,"depth":708,"text":15043},{"id":10381,"depth":708,"text":10381},"我是如何用开源框架快速搭建个人博客，并计划将它作为记录人生与成长的平台",{"date":7961,"image":7962,"alt":14898,"tags":15107,"category":2463,"topics":15108,"published":859},[2130,2131],[2135,3182],"/blogs/zh/build-blog-in-one-day",{"title":14898,"description":15105},"blogs/zh/7.build-blog-in-one-day","p4G7cOokTD28UrnJVHNYaiXFAJvaXNwXY5SfCj2BnzA",{"id":15114,"title":15115,"body":15116,"description":15127,"extension":2126,"meta":15128,"navigation":859,"ogImage":8227,"path":15131,"seo":15132,"stem":15133,"__hash__":15134},"zh/blogs/zh/8.learn-graph-db-neo4j.md","学习图数据库 - Neo4j",{"type":750,"value":15117,"toc":15124},[15118,15121],[753,15119,15120],{"id":15120},"简介",[758,15122,15123],{},"最近我看到一些关于Graph RAG的讨论，这似乎是构建知识系统的一种强大方式。与此同时，在KubeConf 2025上，我遇到了Neo4j展台的工作人员。了解如何使用图数据库来解决我的问题非常有趣。",{"title":839,"searchDepth":708,"depth":708,"links":15125},[15126],{"id":15120,"depth":708,"text":15120},"最近读了一些关于Graph RAG的内容，开始对图数据库产生兴趣，学习如何在工作中使用它。",{"date":8226,"image":8227,"alt":15115,"tags":15129,"category":2463,"topics":15130,"published":859,"featured":859},[2130,2131],[2134,2135],"/blogs/zh/learn-graph-db-neo4j",{"title":15115,"description":15127},"blogs/zh/8.learn-graph-db-neo4j","OVcEcfPPiEWvgyKhUju35Ss-xHRarsyDJYFo8n1XkyQ",{"id":15136,"title":15137,"body":15138,"description":15272,"extension":2126,"meta":15273,"navigation":859,"ogImage":8227,"path":15276,"seo":15277,"stem":15278,"__hash__":15279},"zh/blogs/zh/9.a-new-chapter.md","新篇章",{"type":750,"value":15139,"toc":15268},[15140,15143,15146,15149,15152,15155,15158,15161,15164,15175,15179,15186,15219,15226,15230,15233,15250,15253,15259,15262,15265],[758,15141,15142],{},"最近，我思考最多的问题是：如何让自己的人生发生真正的、质的改变。",[758,15144,15145],{},"回顾过去这些年，我几乎把 200% 的精力都投入到了工作中。每天都像是打仗，不断向前冲。但我也清楚地知道，这种「单线程的人生」并不可持续。我忽略了家庭责任，错过了与亲人相处的宝贵时光；我忽略了自己的健康，在高压和过度投入中逐渐失去了活力；更重要的是，我渐渐变得按部就班，甚至有些麻木和死气沉沉。这个状态，不是我理想中的人生。",[758,15147,15148],{},"前阵子，家庭经历了一些危机，也促使我必须正视一个很现实的问题：作为儿子、丈夫、父亲，我到底承担起了多少应有的角色？",[758,15150,15151],{},"处理父母与伴侣之间的关系，并不是一个简单的逻辑问题，它没有标准答案。这是一门深奥的情感艺术，需要理解、耐心、共情和持续的沟通。这段时间，我花了更多的心力去梳理这些复杂的人际动态，也逐渐意识到：工作可能是我最容易掌控的领域，而家庭与亲密关系，却是我真正需要修炼的地方。",[758,15153,15154],{},"事实上，只要智商在线、执行力强，职业发展往往是可以通过努力获得结果的。但人生不是只有职业。真正值得追求的，不是一个人的成功，而是一个系统性的平衡：家庭、健康、事业、成长，都在稳步前行。",[758,15156,15157],{},"这让我想到孩子的成长。我们都知道，教育孩子不能只靠「堆时间」，更不能指望一套方法包打天下。它需要你陪伴、观察、倾听、引导，就像打磨一个优秀的产品一样：知识广度、视野高度、执行深度，一个都不能缺。但方式可以不同。用更聪明的方式投入，才能在有限的时间里，实现复利式的成长。",[758,15159,15160],{},"而我，终于开始意识到，我也需要用这种聪明的方式重新设计自己的人生。",[758,15162,15163],{},"从今天起，我希望能用一个全新的视角去对待人生的每一个维度。这个时代，AI 已不再是遥远的概念，它已经渗透进了我们的生活。比起机械地卷在低效的路径上，我希望能善用 AI 这个杠杆，去重构自己的生活方式：",[804,15165,15166,15169,15172],{},[775,15167,15168],{},"用 AI 辅助我更科学地管理健康",[775,15170,15171],{},"用 AI 帮助我更高效地完成工作中的复杂任务",[775,15173,15174],{},"甚至用 AI 辅助我更好地处理沟通、关系和情绪调节",[753,15176,15178],{"id":15177},"为什么我选择-ai","为什么我选择 AI？",[758,15180,15181,15182,15185],{},"随着 AI 的发展，我意识到自己可以",[778,15183,15184],{},"用更聪明的方式过日子","。我们正处在一个罕见的时代：AI 不再只是程序员的工具，它开始成为我们每个人生活的一部分。",[804,15187,15188,15195,15202,15209],{},[775,15189,15190,15191,15194],{},"麦肯锡的研究显示，生成式 AI 每年有望释放 ",[778,15192,15193],{},"2.6 到 4.4 万亿美元","的劳动效率。尤其是知识型工作者，可以用 AI 替代大量重复性和信息密集型任务。",[775,15196,15197,15198,15201],{},"哈佛大学的一项实验证明，在复杂写作与数据分析中，使用 GPT-4 的员工准确率提升了 ",[778,15199,15200],{},"40%","，完成时间减半，满意度也显著提升。",[775,15203,15204,15205,15208],{},"美国 Jobscan 的调查指出，44% 的用户使用 AI 优化简历、准备面试后，",[778,15206,15207],{},"薪资提升了 14%","，甚至得到了更合适的岗位。",[775,15210,15211,15212,15214,15215,15218],{},"日本的一款名为 ",[2156,15213,8337],{}," 的 AI 情绪教练，在用户中减少了 ",[778,15216,15217],{},"31% 的焦虑感","，尤其适用于高压职业人群。",[758,15220,15221,15222,15225],{},"我相信，",[778,15223,15224],{},"如果我真的用心去设计自己的生活，并借助 AI 的力量去优化每一个细节","，也许我可以做到：不用透支身体，不必放弃家庭，还能继续创造职业上的价值——而且可能，比以往更轻松、更聪明。",[753,15227,15229],{"id":15228},"我的ai赋能人生实验从今天开始","我的“AI赋能人生实验”从今天开始",[758,15231,15232],{},"我决定给自己一次真正的重启机会，从工作、家庭、健康、情绪、认知各个维度，用 AI 辅助我设计一种新的生活方式：",[804,15234,15235,15238,15241,15244,15247],{},[775,15236,15237],{},"用 AI 规划每日饮食、作息和锻炼，让我别再靠“意志力”硬撑",[775,15239,15240],{},"用 AI 辅助我优化会议、写作、知识获取，让我用最少的时间完成最复杂的任务",[775,15242,15243],{},"用 AI 生成亲子活动灵感，让我更轻松陪伴孩子、建立高质量连接",[775,15245,15246],{},"用 AI 记录情绪、做 journaling、生成反思，让我更敏锐地觉察自己",[775,15248,15249],{},"用 AI 帮助我设计更好的系统，更加搞笑的工作",[758,15251,15252],{},"很快就是我的 38 岁生日。我不想只是切个蛋糕、喝点酒。今年，我更想用一场彻底的人生重启，给自己一个真正意义上的礼物。",[758,15254,15255,15256,11959],{},"愿接下来的路，",[778,15257,15258],{},"更清醒、更平衡、更有创造力",[758,15260,15261],{},"这篇博客，便是我启动这个「AI 赋能人生再设计计划」的第一步。我希望在这里记录下改变的过程、遇到的困难、技术与心理的突破，也许还会有一些让人惊喜的转变。",[758,15263,15264],{},"很快就是我 38 岁的生日。比起庆祝，我更期待这一天能成为我人生真正发生质变的起点。不是换一种工作方式，而是换一种活法。",[758,15266,15267],{},"愿这个新的阶段，是我活得更清醒、更平衡、更有创造力的开始。",{"title":839,"searchDepth":708,"depth":708,"links":15269},[15270,15271],{"id":15177,"depth":708,"text":15178},{"id":15228,"depth":708,"text":15229},"如何利用AI技术重新设计生活，在家庭、健康、事业和个人成长之间找到平衡",{"date":8399,"image":8227,"alt":15137,"tags":15274,"category":3016,"topics":15275,"published":859,"featured":859},[2130,3015],[3182],"/blogs/zh/a-new-chapter",{"title":15137,"description":15272},"blogs/zh/9.a-new-chapter","sKMsU56pP34HszP4GpkXI-zkYk4_o4gd85Mptxuz85w",1781532112541]