Posts
All the articles I've posted.
-
After OpenClaw Shut Down: Rebuilding a Multi-Agent Automation Setup with the Claude Code CLI
When the third-party AI agent framework OpenClaw shut down, I rebuilt the entire multi-agent automation experience with the Claude Code CLI: Telegram bots, scheduled tasks, session persistence — plus every pitfall I hit along the way.
-
Flash-MoE: Running a 397B-Parameter Model on a MacBook at 4.4 token/s
A developer built Flash-MoE in 24 hours: it runs the 397B-parameter Qwen3.5 model on a 48GB MacBook Pro at 4.4 token/s, using only about 6GB of RAM and no cloud GPUs. We break down how it works: SSD streaming, Metal shader optimization, and MoE sparse activation.
-
Apifox Supply Chain Attack Post-Mortem: Your SSH Keys May Already Be Compromised
In March 2026, the Apifox desktop client was hit by a supply chain attack: a JS file on the official CDN was replaced with a malicious version that stole users' SSH keys, Git credentials, and other sensitive data. A technical breakdown of the attack chain, blast radius, and how to check if you were affected.
-
GitHub Squad: Drop an AI Dev Team Straight into Your Repo
Squad, an open-source project, lets you spin up an AI dev team inside your repo with two commands—an architect, frontend dev, backend dev, and tester, each with their own job, collaborating on top of Copilot. A look at its architecture, what it's like to use, and the multi-agent collaboration patterns behind it.
-
Computer-Use: When AI Agents No Longer Need APIs
AI agents are learning to operate computers the way humans do—reading the screen, clicking the mouse, typing on the keyboard. From Anthropic's Claude Computer Use to Microsoft's CUA to OpenAI's Operator, Computer-Use is redefining what "software integration" means.
-
Google Stitch's Big Update: UI Design in Natural Language — Should Figma Be Worried?
Google Stitch just got a major update, evolving from a simple prompt-to-mockup tool into an AI-native design canvas. Infinite canvas, voice interaction, the DESIGN.md design system format, MCP integration — five big features shipped at once. Free, powered by Gemini, and aimed squarely at Figma Make.
-
Karpathy's AutoResearch: Letting an AI Agent Run 700 ML Experiments on Its Own
A deep dive into Karpathy's open-source AutoResearch project: how a 630-line Python script lets an AI agent run ML experiments autonomously on a single GPU, completing 700 experiments in two days and finding 20 effective optimizations. From architecture to practical applications, here's why every developer should pay attention to the "Karpathy Loop" pattern.
-
Reading the MCP 2026 Roadmap: From Local Tools to Production-Grade Agent Infrastructure
MCP (Model Context Protocol) has published its 2026 roadmap with four priority areas: transport evolution, agent communication, governance maturity, and enterprise readiness. A technical breakdown of the concrete problems and proposed solutions in each area.