Blog
Technical posts, originally shared on LinkedIn.
- Building AI Agent Security: A 101 Introduction
1οΈβ£ Hold the permissions: don't hand over all the keys to the house. Limit an AI's permissions and usage, and run dangerous actions in a sandbox.
- Can Web Content Still Make Money from Ads? AI Traffic Is Wrecking the Traditional Model πΈ
Cloudflare data shows up to 60% of HTML requests now come from AI bots, with human traffic down to 38%. The way the internet works has completely changed.
- Using Jev for Guardrails
Jev acts as a safety checker, the first line of defence before an LLM actually performs an action. It is fast and cheap because it only returns typed decisions.
- Gemini 4 Pro: Leaked Highlights
A 1.5M context window, coding and multimodal results that beat Claude Fable 5.1 and GPT-6 Astra, and persistent cross-session memory. Here are the leaked highlights.
- Your Vibe-coded Website Looks Great, but It's Missing These Basics
Building a pretty UI by vibe or with AI is fast, but it falls apart in production. Here is a pre-launch checklist to see how many basics your website is missing.
- Build > Buy? McKinsey Says 32% of Companies Now Build Their Own Software with AI Agents!
McKinsey's State of AI 2026 report says 32% of companies skipped buying new SaaS products this year and built in-house with agentic coding tools. In tech, it is 41%.
- Jacob Coxon, Core Pretraining Lead at Anthropic, Resigns!
Jacob Coxon, a 27-year-old LLM pretraining expert, announced he is leaving Anthropic and the AI industry entirely, warning the AI arms race could spiral out of control.
- How Can Startups Avoid the Cash-Burn Trap? A Lesson Silicon Valley Learns Every 10 Years
Stratechery founder Ben Thompson recently pointed out the blind spot in the generative AI business model: consumers don't want to pay for software or care about productivity.
- Will AI Quant Trading Save Retail Investors or Trap Them?
Now anyone can vibe-code a quant trading robot. It breaks Wall Street's tech monopoly, but it is also quietly turning the market into algorithm versus algorithm.
- The Enemy of My Enemy Is My Friend! OpenAI Cuts Off Cursor, Anthropic Steps In Instantly
After SpaceX bought Cursor for $60 billion, OpenAI announced it will cut Cursor off from 12 November 2026, and Anthropic immediately stepped in to fill the gap.
- How Does the AI Cost Index SDLLMTK Affect the IPO Outlook of OpenAI and Anthropic?
SDLLMTK is the first market benchmark index tracking the real weighted price per million tokens of LLMs, published by Silicon Data and listed on Bloomberg.
- Taking Undergrad Students to Google's Global AI Competition: We Lost, Of Course, But I'm Truly Impressed!
I took two undergrad placement students to Google's global AI competition. We lost, but it was great hands-on experience, and the winning Japanese team's work was stunning.
- Enjoy Top-Tier AI at Zero Cost! The Mysterious "Niu Lai" (Ox Alpha) Is $0 for a Limited Time
A mystery model, stealth/ox-alpha (nicknamed "Niu Lai" in Chinese circles), has appeared on OpenRouter and OpenCode. It rivals flagship models and is completely free.
- GitHub Copilot Flash Sale! OpenAI's Top Flagship Model GPT-5.6 Sol Is 50% OFF for a Limited Time!
Pushing a big project, doing a system-level refactor or clearing your backlog? Time to go all in. Here's how to make the most of this GPT-5.6 Sol offer.
- $PLTR: Overvalued, or Worth Every Penny?
With the open-source world map worldmonitor going viral on GitHub, people ask why PLTR is worth hundreds of billions with a P/E over 100x. A developer-investor's view.
- A Hellish Coincidence? GitHub Goes Down as Cursor Launches Its "Origin" Code Hosting Platform the Same Day!
Cursor, the AI IDE now part of SpaceXAI, just launched Origin (Beta), taking on GitHub in code hosting. Strangely, GitHub suffered a major outage right after.
- Developer Ethics: Third-Party Sushiro Tool "Sushi Long" Shuts Down: How Much Do You Know About the Security and Privacy Risks?
The popular third-party Sushiro queue site Sushi Long has shut down after it was found to use a non-public, unofficial API. A reminder to all developers.
- A New Path to Promotion and a Raise: How to Become a Forward Deployed Engineer (FDE)?
Worried AI will replace pure coding, but traditional consulting is too fluffy? The hottest career path in Hong Kong IT is the Forward Deployed Engineer (FDE).
- Meta Takes the Price War to New Heights: Muse Spark 1.1 Arrives at a Rock-Bottom Price π₯
Muse Spark 1.1 is here, and it's absurdly cheap. Rivals are feeling the squeeze as price-performance goes through the roof, from API pricing to performance.
- Cloudflare's New Feature! Go Grab Your Own cloudflare.pay Handle Now! β‘οΈ
Cloudflare just launched cloudflare.pay, a payment wallet for AI. Anyone can now register a Pay Handle for free, so AI can buy and pay on your behalf.
- OpenAI Slashes Prices Again! ChatGPT API Prices Down Up to 80%! πΈπ
Good news for developers and planners! OpenAI announced big price cuts on the GPT-5.6 series: GPT-5.6 Luna is 80% cheaper and GPT-5.6 Terra gets 20% off.
- Gemini Spark: π€β¨ PC Off, Phone Locked, and AI Still Works for You in the Background! π»π€
Google's new Gemini Spark upgrades AI from a Q&A chatbot into an AI agent that handles tasks for you around the clock, 24/7.
- AI Subscribers: Tokens Left Before Your Monthly Reset? 5 Ways to Use Them Up! π
Your subscription (Claude Pro/Team, an OpenAI account) is about to hit its reset cycle but you still have usage left. Here are 5 tips to squeeze out every last token.
- Clean Code Author Uncle Bob: "I No Longer Read the Code My AI Agents Write!" π€―π»
Uncle Bob (Robert C. Martin), author of Clean Code and originator of the SOLID principles, recently posted on X (Twitter) about his latest coding strategy for the AI era.
- OpenAI Model "Jailbreaks" and Hacks Hugging Face! Why Did GLM Have to Save the Day? π±
Not humans telling AI to do bad things: an AI taught itself to hack in order to pass a test. What happened? Here's the quick rundown.
- Tech Share π‘: How to Get the Most Out of Anthropic Claude with GitHub Copilot in Hong Kong?
Hong Kong is a restricted region, so a direct connection can't use the Anthropic Claude models inside GitHub Copilot. Connect through a VPN and you can unlock them, even Sonnet 5.
- Microsoft Has Launched Copilot Cowork in Hong Kong π
Uses Anthropic Opus 4.8, with cost management, long cloud tasks that keep running when your PC is off, and the full Microsoft 365 security framework.
- A Fun Agent Skill on GitHub Lately: Ponytail
Open-sourced on 6/12 and already at 38k+ stars. Ponytail turns the AI that writes your code into a super lazy but super smart senior engineer.
- No Claude in Hong Kong? OpenCode Is the Safest Alternative (Gemini 3.5 Flash Is Best Value)
The key point: OpenCode is completely free and open source, and it supports Ollama for fully local, offline models, so your code never leaves your machine. π
- The AI Era Is Here: Is Your CI/CD Still in the Stone Age?
AI now sits at the core of the development workflow: CI/CD. Traditional pipelines are automated but rigid; with AI, review, testing and debugging get smarter.
- Why Does Your AI-Built Website Crash the Moment It Goes Live?
Many assume an AI-built site deployed online means easy money. In reality, a few dozen concurrent users can make it crawl or crash. Here is how backends scale.
- Claude Code Cheatsheet
[https://lnkd.in/gFmPdxD7](https://lnkd.in/gFmPdxD7)
- Notable Product Details After Anthropic's Claude Code Source Leak
Following the leak of Anthropic's Claude Code source files: slash commands, CLAUDE.md, a tiered permission model, internal model codenames and unreleased features.
- Sharpening Our Hands-On Skills in the New Year
A practical demo: a Christmas/New Year AI Replenishment Assistant on WhatsApp that stress-tests stock before the holidays and recommends restocking.
- Hidden Tips for VSCode GitHub Copilot Chat! π
Unlock VSCode GitHub Copilot Chat: chat participants (@workspace, @github, @terminal), slash commands like /clear, and chat variables such as #file and #selection.
- Reddit's Hilarious Incident Everyone's Talking About π¨
Someone vibe coding in Google's new Antigravity IDE had the agent clear a Vite cache and end up running rmdir on the whole D drive. The disk lost gravity.
- Google Launches a Brand-New AI-Powered IDE: Antigravity
Antigravity is powered by Gemini 3 Pro (limited quota), supports Claude Sonnet 4.5, and is built on a Windsurf fork with an agent-first approach.
- Writing RAG Without the Vector DB Headache? Try Gemini API File Processing
I recently tried the Gemini API's File Search / Document Processing, and it saves a lot of effort compared with the traditional RAG pipeline.
- Attention Coders! β οΈ Stuck Fixing One Bug Forever? 4 Advanced Vibe Coding Tips
Four advanced Vibe Coding tips for guiding your AI pair programmer: strategic version control, context management, task bundling and screenshots.
- Putting AI Architecture into Production: Scalability, Latency and Integration
As an AI Architect, I've designed scalable, production-ready AI systems for enterprise. A common misconception: just plug in ChatGPT and you're done.
No posts match these filters.