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.

I wanted to contribute to the Fydetab Duo Wiki, but to preview the blog locally, I needed to add pixi or other configurations to the project. These changes were not suitable for the original repository. So I learned how to selectively submit commits, which is where git cherry-pick comes in.

It all started when I tried to use Devpod to set up a container based on Docker Compose, only to find that Devpod didn’t seem to be able to call Docker Compose correctly to create the container. I didn’t think much of it until I searched online… Wait, what? Has this project really been dropped by loft‑sh?

The new year has brought another AI surprise…

After my previous small contribution to a conda‑forge package, I wanted to take on something more advanced: trying to publish the singler‑py Python package into the conda ecosystem. Little did I know that it would turn out to be… somewhat troublesome.

Finally, aider-chat is available as a conda package, which means it can theoretically be installed globally via pixi global. However, during actual installation, you’ll find that one of its dependencies, tree_sitter_languages, doesn’t have a corresponding aarch64 version, causing the installation to fail. This made me wonder: could I rely on AI to solve this?

Those Bioinfo projects at my new workplace are mostly personalized, so each project inevitably involves reading some new papers. Once again, I’ve been feeling a headache from reading so many english papers. It occurred to me that the last time I created a word cloud was back in 2022. Three years have passed, and it’s time to welcome… well, just an update.

From late 2022 to early 2023, ChatGPT exploded in popularity, and I started using LLMs to assist with script writing. Although it was quite useful, the price was relatively high, and payment was always an issue, so its application was limited to simple coding problems.

From late 2024 to early 2025, DeepSeek became a hit. While its answers often weren’t entirely satisfactory, and the addition of a “thinking mode” made the response time a bit slower, it was really cheap! A year later, DeepSeek remains one of the most affordable models in terms of tokens while still delivering decent output. I paid 50 RMB at the beginning of the last year, and now, almost exactly a year later, I still have 27 RMB left… As a result, I’ve become much bolder in applying LLMs to various other areas.

Preparing a resume may not sound difficult, but in practice, it can be quite time-consuming. Nowadays, many jobs have specific requirements for niche knowledge or projects. With so many job seekers, a resume that doesn’t highlight the key points relevant to the job requirements might be overlooked by employers. Therefore, to achieve better results, it’s best to tailor your resume for each position… This is something AI should be good at, but at least for now, I haven’t found a good free tool for this.

Recently, I came across RenderCV, a tool that generates resumes from YAML configuration files. It enables a “configuration-as-resume” approach. Combining it with an AI coding assistant like Aider essentially creates a rapid resume preparation environment.

I’ve been using my current avatar for over ten years. Ever since I started working in bioinformatics, I’ve wanted to add more complexity to it—half with circuit‑board patterns, the other half with DNA patterns, connected by a smooth transition in the middle to symbolize the transformation from biology to information, which fits my professional field. However, I’m not very skilled at photo editing, so I couldn’t achieve the desired effect. This year, the newly released image‑editing models gave me hope to realize my idea.