I recently got a Mac Mini to use as a server. I initially thought that since macOS is based on Unix, it would be pretty much like Linux — just set it up and go. But once I actually started using it, I ran into quite a few pitfalls. Here’s a record of the main issues.

To do some small-scale model training at work, I eventually recommended buying a Mac Mini. The reasoning wasn’t complicated: I needed it for bioinformatics analysis, running agent deployments, and occasionally training models with modest parameter counts — these scenarios are exactly where M-series chips with unified memory shine. The cost of separate RAM plus a large-VRAM GPU far exceeds the Mac platform; do the math and the choice is clear.

And just like that, I unlocked the achievement of pushing forward multiple devices, multiple platforms, and multiple projects all at once.

Recently I submitted a PR to BananaSlice (#11), adding configurable Gemini-compatible API endpoint support. The whole process made me realize: AI Agents not only help people write code faster, but also change how we leverage open source software.

Following up on the previous post [From Fydetab Duo to HP Chromebook X11], I have been using the HP Chromebook X11 for a while. The overall experience is quite good, but I encountered two annoying bugs. Here I record the solutions.

Recently, many token-selling providers have been raising prices in disguised ways. Just as everyone was getting used to “AI is getting more expensive,” DeepSeek did something absurd — they permanently cut the price of V4-Pro to a quarter of the original.

Recently, while doing differential gene expression analysis, I needed to map the expression values of DEGs onto KEGG pathways. I used Bioconductor’s pathview package, but the resulting nodes were ridiculously large and the colors were wrong. With the help of AI, I fixed the issues and also picked up a new patching technique along the way.

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!

It’s been barely a month since I wrote [The Perplexity of Modern Laptop and Tablet Choices], and here I am again.

That’s right — the same person who confidently declared “I’ll stick with the Fydetab Duo for now” at the end of that article is now sitting in front of an HP Chromebook X11 typing these words.

The essence of human nature is “it smells so good” (after swearing off something). My essence is tinkering.

I previously wrote a blog post about using pixi’s tasks feature to fix missing dependency issues with Bioconductor packages (like GenomeInfoDbData, BSgenome.Hsapiens.UCSC.hg38, etc.) after installation. At the time, I only knew the problem existed but didn’t understand the root cause — I was just providing a less-than-ideal workaround.

But recently, with insights from AI responses, I finally figured out the real reason — it all comes down to the post-link script mechanism in the Conda ecosystem.

I recently switched to another batch of commonly used AI tools — some pleasantly surprising, some left me speechless. Here’s my honest record. Multica: Master of Cyber Livestock ManagementMultica’s...