Journal 2026-02

February 2026 journal

Monthly retrospective

Wrote a few pieces. Building new software keeps drawing my writing in this direction.

In February, too, I did Anki and closed my Apple Watch rings every single day.

Journal

2026-02-28

Tidied up some notes I’d written around 2021: Distorted user-centered design.

Thought about software assets in the AI era.

2026-02-27

Added full-text search to the wiki.

2026-02-23

Finally started using two $200 Claude Code accounts. I pay for one personally; the company supports the other. My monthly LLM subscription spending is now about $1,200.

2026-02-18

Built a prototype collaboration system in which the user and AI agent each have an independent workspace, using Linux virtualization. Again, something that would have taken a long time without LLMs was done in two days.

Having it embrace the “everything is a file” philosophy and make full use of Linux’s existing capabilities produced a robust system without much code.

2026-02-17

Built a WASM program to convert LaTex into PDF, extracting TeX Live files so only the necessary ones download dynamically. Also implemented Language Server Protocol, with good support for renaming, go to definition, and so on. Without LLMs, I wouldn’t even have considered starting, but it was all done in a day.

2026-02-11

Continuing to think about tasteful software, I wrote down a few more fragments of thoughts I’ve carried for a long time:

2026-02-09

Left the AI software dark factory running overnight and woke up to a code-delete-and-run (it deleted all the code and vanished). Just as a gambler who includes going all-in in their betting strategy is bound to lose everything eventually if they keep playing, without constraints this ending seems almost inevitable. Something like a Ratchet is essential.

2026-02-06

With Opus 4.6 and Codex 5.3 (xhigh), even simple prompts increasingly lead to over 30 minutes of independent work. I like the results more, too. It feels as though a well-designed loop could keep them working all day without trouble.

2026-02-01

Created AI and science and am collecting related news.

Built a Chrome extension to help with building skilled intuition for using LLMs, but hardly use it. The easier making things becomes, the more important it seems to have the taste to decide what not to make. I should think even harder about tasteful software.

Joined Corca.