Journal: January 2026

January 2026 journal

Monthly retrospective

January went by so fast. Articles I wrote:

I kept up with Anki and closed my Apple Watch rings every single day.

Journal

2026-01-31

Before asking “Can humans control AI?” we should ask “Do humans have the will to control AI?” The latter is more likely to cause problems than the former. With the popularity of OpenClaw (formerly MoltBot, Clawdbot), people have begun worrying about whether it will become Skynet. Humans are the problem here. See Accidental Skynet.

2026-01-25

I organized my thoughts on Tasteful software.

Former prime minister Lee Hae-chan has passed away. May he rest in peace.

2026-01-20

It seems increasingly common for one person to build and operate multiple services using agents. Usually that person is a developer, so the resulting services do not have outstanding design. But design matters. How can we solve this problem?

2026-01-17

I organized my thoughts on Design in the AI era.

2026-01-11

I read Finding and Fixing Ghostty’s Largest Memory Leak. It describes fixing a memory leak present since Ghostty 1.0 thanks to Claude Code. The leak wasn’t fixed “using Claude Code”; rather, “thanks to Claude Code,” it began causing serious problems for many users and became reproducible enough to fix.

In some ways, it reminds me of the Internet Explorer memory leak that went unfixed for over ten years, until version 7.0, and was finally fixed thanks to Ajax. That case differs slightly: the cause (circular references) was known, but it was left unfixed until Ajax made the problem serious.

Anyway, Ajax spurred enormous advances in browsers. Perhaps terminals will also advance enormously as terminal-intensive programs like Claude Code become widely used?

2026-01-10

Since ChatGPT (November 2023), users seem to divide software features into two categories.

Creative LLM-based features that are vulnerable to hallucinations.
Logic-based features that carry out precisely what they are told without mistakes.

Users must have a meta-mental model for judging which category a feature belongs to, and getting it wrong has costs: wasting time on suspicion where there is no risk of hallucination, or blindly believing incorrect information because they mistakenly think there is no such risk. These costs can sometimes be enormous—for example, a lawyer citing nonexistent precedents.

People planning AI-infused services should help users form this meta-mental model well.

2026-01-09

I am organizing my thoughts on Software project documentation in the AI era, reflecting my various attempts and missteps so far.

I added a few overdue monthly retrospectives and tidied the journal a little. As I did, I began thinking an annual retrospective would be nice too.

The aftermath of an important decision has left many things to change, reducing my breathing room, and I am still scrambling. Things should settle down somewhat next month.

2026-01-01

It’s 2026. How unfamiliar.