Huon on the internet

2026

Gell-Mann AImnesia

Any time I use an AI tool for something I’m deeply familiar with, I find a continual stream of mistakes and inconsistencies. When I use it on a topic I don’t know… everything sounds plausible and I don’t find mistakes. The difference worries me, and sounds a lot like the Gell-Mann Amnesia effect.
21 Apr 2026
Write broken commits for better review

I spend a lot of time reviewing code, and I think it’d be easier if I saw more tastefully broken commits. Commits construct a story about a change: “first this happened, then that, that another thing” is a good narrative; “first everything happened, the end”… not so much. Sometimes, telling that story is impossible without having a commit that breaks tests or doesn’t compile!
17 Apr 2026
Why don't multi-column indices help queries on the second column?

SQL databases support multi-column indices, where a single index can incorporate data from multiple columns. These are powerful, but the order of the columns matters, with differences observable in practice: there can be orders of magnitude between two similar & simple queries, depending on which uses a prefix of the index columns. Why does this limitation exist? Can we have a mental model for the indices that explains the behaviour?
23 Mar 2026
Staying engaged with AI plans: give inline feedback

When I use an AI coding agent to make a plan, I make a habit of opening that plan in my editor and then leaving inline comments with questions, requests and corrections directly in the file. The agent seems to be happy enough to work with these, and I find it gives better results than reading the plan in the agent and using chat to give feedback.
05 Feb 2026

hit counter