While setting up the Claude Science container, I learned about a new way of using NVIDIA GPUs inside containers, so I’m noting it down (most of it was written by AI, used as reference material for my own research).

I had already managed to get claude-science running in a container, and basic functionality worked without major issues. But when I wanted to do things like molecular docking, I couldn’t—the container simply had no GPU. However, when I pulled NVIDIA’s official image to build the container the same way I had before, something bizarre happened: every language kernel in CS—Python, Perl, R—stopped working, and the entire runtime environment was rendered useless.

This was truly beyond my understanding. After consulting Gemini manually with no luck, I decided to let opencode + deepseek-v4-flash handle it on its own—not me remotely typing commands and directing it, but an AI agent living inside the container that read the logs, bisected parameters, decompiled the parser, and edited the entrypoint itself. It was a strange and interesting experience; here’s a record of it.

I recently helped a friend build a small tool that processes sequences from an input file and generates XML in a specific format. My friend isn’t comfortable with the command line, so a GUI was needed. The input data also had confidentiality requirements, meaning all computation had to happen locally. Normally I’d build a desktop app, but considering long-term maintenance, I wanted to try an offline-capable PWA. I checked and found that Pyodide already ships pandas as a wasm package, so there was nothing extra to do on the runtime side — let’s go!

Recently, I spent several days deeply experiencing multiple AI Agent products, from WorkBoddy, OpenClaw, LobeHub, DeerFlow2, LirbeChat, DingTalk Wukong to OpenCode, trying almost every tool I could find. Everyone has different needs, but my main focus is on open source and self‑deployable, feature completeness, and practical usability. Below is my personal comparison from a practical usage perspective.

Looking back, I learned to use FrontPage to create web pages since elementary school, but except for GitHub Pages, I never really set up and operated my own website. However, now it’s possible. With the support of agents and SaaS services, creating and launching a website has become quite simple…

I’ve been trying to use AI Agents / workflows to build efficient agent teams or AI workflows to improve productivity. After two weeks of experimentation, I’ve encountered several pitfalls…

Last year, I explored AI topics purely out of personal interest. This time, it’s under pressure from my boss… Even though I’m full of resentment, I still need to write something down to accumulate experience…

git worktree is a powerful feature provided by Git that allows you to check out multiple branches in the same repository simultaneously. Like cherry-pick, I hadn’t heard of this feature before, but in today’s era of extensive AI Agent applications, this feature will likely become as fundamental and essential as commit.

The Lunar New Year holiday is supposed to be a time for lying back, watching videos, and playing games. However, as a dedicated workhorse programmer, how could I possibly let myself stay idle? So, during the break, I tinkered with opencode + MiniMax/Deepseek and came up with a few little things (definitely not because I got utterly wrecked in Street Fighter on New Year’s Eve and needed a distraction).

Previously, I had tried modifying and adding recipes to conda channels. This time, I want to package my DevSSH and upload it to my own channel. I decided to try building a conda package myself.