# Blog **Summary:** Articles by Mark Huang about AI, software engineering, and building technology. - Canonical: https://markhuang.ai/blog - Language: en --- ## [Claude Max 20x Hit 100% on Day Four. Why I Moved to Codex](https://markhuang.ai/blog/why-i-cancelled-claude-max-for-codex) 2026-08-08 After Claude Max 20x hit its weekly limit on day four, I moved my coding work to Codex. Rising correction effort made the decision easier. My current setup uses Sol to plan, Luna to implement, Dense-Mem for context, and other model families for review. --- ## [I Keep Asking Why: The Mindset I Want to Keep in the AI Era](https://markhuang.ai/blog/i-keep-asking-why) 2026-07-14 A conversation with my wife made me examine how I became an engineer and architect: learning from people, asking why, testing assumptions, and owning decisions in the AI era. --- ## [I'm Aggressively Trying to Replace Myself With AI](https://markhuang.ai/blog/aggressively-trying-to-replace-myself-with-ai) 2026-07-13 AI becomes more useful as it learns more context, but every email, document, and story moves the privacy boundary. A personal reflection on automation, AI slop, focus, and builder responsibility. --- ## [GPT-5.6 Sol Scores Higher. Why Does My Weekly Limit Vanish Faster?](https://markhuang.ai/blog/what-ai-benchmarks-actually-measure) 2026-07-12 A mixed review of GPT-5.6 Sol leads to a plain-English guide to context limits, agent swarms, and every current Artificial Analysis LLM index—with concrete examples. --- ## [AI MemMail: Let Dense-Mem Triage and Answer Your Business Inbox](https://markhuang.ai/blog/ai-memmail-dense-mem-email-agent) 2026-07-05 A practical setup guide for ai-memmail, the open-source Rust email agent that watches IMAP inboxes, recalls business context from Dense-Mem, and safely replies, forwards, or no-ops through SMTP. --- ## [Agentic Browser Automation Without Giving AI Your Login](https://markhuang.ai/blog/ai-browser-automation-without-sharing-login) 2026-06-26 A beginner-friendly tutorial for building scheduled or queue-driven agentic browser automation with Selenium, noVNC manual login, and a narrow API wrapper so agents can draft website forms without seeing credentials or controlling your PC. --- ## [I Might Be Wrong About Agentool](https://markhuang.ai/blog/i-might-be-wrong-about-agentool) 2026-06-20 A personal automation postmortem about building agentool to make AI CI workflows lighter, then realizing the real cost may be feature maintenance, orchestration complexity, and chasing SDK behavior that Claude Agent SDK and Codex SDK already own. --- ## [Stop Teaching Every AI From Scratch](https://markhuang.ai/blog/centralized-ai-knowledge-graph-dense-mem-case-study) 2026-06-10 A personal Dense-Mem reflection on the problems that pushed me beyond static skills and stale files toward dynamic shared memory, read-only automation context, import/export, and governed knowledge graphs. --- ## [I Feel Sorry for AI](https://markhuang.ai/blog/i-feel-sorry-for-ai) 2026-06-03 Why both AI hype and anti-AI hostility miss the same point: LLMs behave more like straight-A new graduates than senior experts, and useful agents need onboarding, skills, and maintained memory rather than impossible first-attempt expectations. --- ## [Skills + Dense-Mem: Making AI Workflows Learn From Experience](https://markhuang.ai/blog/skills-plus-dense-mem-ai-workflows-learn) 2026-06-02 A hypothesis for combining AI skills with Dense-Mem: keep workflow, safety rules, and acceptance criteria in skills, while memory stores expectations, examples, corrections, failures, and portable skill-pack knowledge. --- ## [Try Dense-Mem in 5 Minutes With the Hosted Demo](https://markhuang.ai/blog/dense-mem-hosted-demo-test-instance) 2026-05-31 A quick tutorial for using the hosted Dense-Mem test instance, connecting Claude Code and Codex to the same temporary memory, and seeing how shared context helps AI work smarter. --- ## [Dense-Mem Quick Start: Give Claude Code and Codex the Same Memory](https://markhuang.ai/blog/dense-mem-personal-server-claude-code-codex) 2026-05-30 A beginner-friendly tutorial for spinning up a local Dense-Mem server, creating your first memory key, and connecting Claude Code and Codex to one shared AI memory brain. --- ## [Secure Dense-Mem on Vultr with Traefik](https://markhuang.ai/blog/secure-dense-mem-vultr-traefik) 2026-05-30 A nontechnical walkthrough for launching Dense-Mem on a Vultr cloud server with Traefik, HTTPS, private control-portal access, and shared memory for personal, family, or work AI tools. --- ## [System Prompt vs User Prompt: The Layer Under GenAI Features](https://markhuang.ai/blog/system-prompt-user-prompt-genai-features) 2026-05-26 A beginner-friendly explanation of system_prompt and user_prompt using ChatGPT, Claude Projects, Claude Cowork, and Claude Code examples. --- ## [AI Memory Beyond RAG: Vectors, Graphs, and Dense-Mem](https://markhuang.ai/blog/ai-memory-beyond-rag) 2026-05-25 RAG is not magic memory. A practical explanation of chunks, embeddings, vector search, graph-backed memory, and why durable AI memory needs provenance, conflict handling, and retrieval policy. --- ## [Is AI Bad?](https://markhuang.ai/blog/is-ai-bad) 2026-05-22 AI can make you faster, lazier, more capable, or more dependent. The real question is not whether AI is good or bad, but which knowledge you are choosing to outsource and whether the trade is worth it. --- ## [From Software Developer to AI Architect: What Changed in One Year](https://markhuang.ai/blog/from-software-developer-to-ai-architect) 2026-05-17 A personal journey from Claude Code skills to TypeScript state machines, MCP tools, and finally SDK-based AI workflows. Skills help, tools help, but prompts are not enforcement and LLMs should not own the control plane. --- ## ["My Company Doesn't Need AI." Think Again.](https://markhuang.ai/blog/my-company-doesnt-need-ai-think-again) 2026-05-17 AI adoption is not just tool selection. Even companies that think they do not need AI need to understand where AI fits, what should stay deterministic, and who owns customization, safety, and long-term control. --- ## [Three Cobblers, One Zhuge Liang: Making Cheaper Models Work Together](https://markhuang.ai/blog/three-cobblers-one-zhuge-liang-ai-architecture) 2026-04-30 A personal AI architecture lesson from the Chinese saying 三个臭皮匠，顶个诸葛亮: why cheaper models fail on giant prompt blobs, and how focused specialist sessions, orchestration, synthesis, and temperature control can make them useful. --- ## [The Determinism Trade: Why I Archived My Agent Framework for 15 Minutes of N8N](https://markhuang.ai/blog/the-determinism-trade) 2026-04-21 I spent two months building OpenHive — my own OpenClaw — to explore 'agent as feature' for a 1-man company. It reached v4, 90% functional, and still shouted log lines I never asked for. I rebuilt the same monitor in N8N in 15 minutes. Here's what I learned about where LLMs actually belong. --- ## [The 1+1 Hypothesis: Can You Break Coding Problems Small Enough for Any LLM?](https://markhuang.ai/blog/the-1-plus-1-hypothesis) 2026-04-08 Every LLM can do 100×100. Every coding LLM can rename a variable. But where does reliability break — and can harness engineering push that boundary? Exploring residual solution entropy, test-first contracts, layered defense architectures, and why blind consensus fails while verified search works. --- ## [No, Chinese Is Not More Token-Efficient Than English for LLMs](https://markhuang.ai/blog/chinese-token-myth) 2026-04-07 A native Mandarin speaker tests the popular claim that Chinese characters save tokens when interacting with LLMs. Across six tokenizers — including Chinese-first models like Qwen, GLM, and DeepSeek — English uses fewer tokens every time. The data, the BPE mechanics, and why character count has nothing to do with token count. --- ## [Automation Without Intention Is Just Faster Chaos](https://markhuang.ai/blog/automation-without-intention-is-just-faster-chaos) 2026-04-03 Three failed pipeline architectures, a lesson about backpressure, and the UAT gate that finally made multi-AI vibe coding work. An experience-sharing post about what broke, what survived, and why knowing what you want matters more than the tools you use. --- ## [Don't You Think Your AI Is Too Optimistic?](https://markhuang.ai/blog/dont-you-think-your-ai-is-too-optimistic) 2026-03-21 RLHF can reward agreement over accuracy, turning AI into a source of sugar-coated bullets — validation that hides failure modes. How persistent adversarial rules change the default from flattery to honest challenge. --- ## [Agent as Feature: What Happens When AI Replaces Your Backend Logic](https://markhuang.ai/blog/agent-as-feature-what-happens-when-ai-replaces-your-backend-logic) 2026-03-10 Gartner predicts 40% of enterprise apps will embed AI agents by 2026. The 'agent as feature' pattern replaces deterministic controllers with reasoning agents. An exploration of what this means for backend architecture and why the potential is real. --- ## [Why One AI Is Never Enough](https://markhuang.ai/blog/cross-family-multi-ai-why-one-ai-is-never-enough) 2026-03-04 Every high-stakes profession requires independent review — medicine, law, science, finance. AI is one of the few domains where people skip this step. 37% of enterprises already use 5+ models, but most do it ad-hoc. Chapter 1 of Cross-Family Multi-AI. --- ## [The Science of Ensemble Intelligence](https://markhuang.ai/blog/cross-family-multi-ai-science-of-ensemble-intelligence) 2026-03-04 Wisdom of crowds meets AI: diverse LLM ensembles outperform 67% of individual models, F1 scores jump from 0.55 to 0.80+, and 56.9% of best solutions come from the weakest models. The math behind cross-family multi-AI. Chapter 2 of Cross-Family Multi-AI. --- ## [Industry Evidence — Healthcare, Finance, Legal, and Beyond](https://markhuang.ai/blog/cross-family-multi-ai-industry-evidence) 2026-03-04 Multi-model AI is already mainstream in healthcare diagnostics, financial risk management, legal analysis, and content moderation. The evidence from four industries — and what it means for cross-family AI adoption. Chapter 3 of Cross-Family Multi-AI. --- ## [The Monoculture Risk — When Every AI Agrees on the Wrong Answer](https://markhuang.ai/blog/cross-family-multi-ai-monoculture-risk) 2026-03-04 The dangerous risk of single-AI dependency isn't outages. It's correlated wrong answers that nobody catches because nothing pushes back. When every team uses the same model family, the same blind spots propagate silently. Chapter 4 of Cross-Family Multi-AI. --- ## [The Cost Question — When Multi-AI Pays for Itself](https://markhuang.ai/blog/cross-family-multi-ai-cost-analysis) 2026-03-04 Multi-AI costs 3-4x more per token — but organizations lose 40% of AI productivity gains to rework. Execution order, task-appropriate scaling, and the 21x ROI gap between mature and immature AI practices. Chapter 5 of Cross-Family Multi-AI. --- ## [The Road Ahead — Building a Cross-Family AI Practice](https://markhuang.ai/blog/cross-family-multi-ai-road-ahead) 2026-03-04 A 5-level maturity model from single model to self-optimizing, practical next steps for individuals, teams, and enterprises, and an honest look at the evidence gaps that still need filling. Chapter 6 of Cross-Family Multi-AI. --- ## [The Multi-AI Thesis](https://markhuang.ai/blog/the-multi-ai-thesis) 2026-03-01 LLMs confirm their own answers over 90% of the time and have a 64.5% blind spot rate on their own errors. Cross-family multi-AI pipelines — Claude reviewing GPT reviewing Qwen — break the self-review ceiling. The research, the costs, and what actually works. --- ## [Why I'm Using a $1 Alibaba Subscription to Explore AI Models Beyond Claude and GPT](https://markhuang.ai/blog/why-im-using-alibaba-bailian-to-explore-ai-models) 2026-02-28 Alibaba's Bailian Coding Plan bundles eight AI models — Qwen 3.5 Plus, Kimi K2.5, GLM-5, MiniMax M2.5, and more — for ¥7.9/month (~$1 USD). A look at the pricing, setup, trade-offs, and what it's like using these models for code review alongside Claude and Codex. --- ## [Nano Banana 2 Is Fast and Cheap, But Is It Better?](https://markhuang.ai/blog/nano-banana-2-fast-cheap-but-better) 2026-02-26 Google's Nano Banana 2 brings 4K resolution, subject consistency, and Pro-level features at Flash speed — but choosing Gemini 3.1 Flash over Pro raises questions about whether this is a quality upgrade or a cost optimization play. --- ## [Permission Syndrome: Why --dangerously-skip-permissions Is the Vibe Coder's Favorite Footgun](https://markhuang.ai/blog/permission-syndrome-dangerously-skip-permissions-footgun) 2026-02-26 Real incidents of AI wiping home directories and production databases — and how VCP's security gate and Docker image let you keep the speed of --dangerously-skip-permissions without the risk. --- ## [Dev Buddy: Multi-AI Development Pipelines for Claude Code](https://markhuang.ai/blog/dev-buddy-multi-ai-pipelines-with-claude-code) 2026-02-25 A step-by-step guide to Dev Buddy — an open-source Claude Code plugin that orchestrates multiple AI models through structured development pipelines with task-based enforcement, parallel specialist analysis, and automatic fix-and-re-review loops. --- ## [VCP: Enforcing Code Standards in AI-Assisted Development with Claude Code](https://markhuang.ai/blog/vcp-enforce-code-standards-with-claude-code) 2026-02-25 A step-by-step guide to installing and using Vibe Coding Protocol (VCP) — a three-layer enforcement framework that catches security vulnerabilities, enforces architecture standards, and orchestrates multi-AI code reviews directly inside Claude Code.