I thought it would be cute to follow last night's post with a dawn … but I slept through it. So, another dusk. I learned the other day about [`sunwait`](https://github.com/risacher/sunwait), a command-line tool for predicting local sunrise and sunset times.  I'm interested this year in reconnecting with naked-eye astronomy, and I thought it might be useful to have a table of sunrise and sunset times. Such tables are things I used in past years to plan bicycle rides, so that I wouldn't be out on the road in twilight. But web interfaces take a lot of clicks (especially if you want data for multiple days), and opening a web browser is an opportunity to forget what I'm doing. This is one of the few cases where I find artificial-intelligence chatbots useful: as natural-language search engines with the ability to filter by semantics, rather than by keyword. I got three recommendations, of which `sunwait` was the most directly useful. `sunwait` is a tool that's been around forever, whose idea is to be able to schedule things on your computer to happen at a particular stage of twilight: ```text # Generate a crummy calendar of rise and set times for January $ ./sunwait list 31 d 01 m 01 36N 86W | cat -n 1 06:56, 16:43 2 06:56, 16:44 3 06:56, 16:45 ⋮ ⋮ ⋮ 30 06:49, 17:11 31 06:48, 17:12 $ ./sunwait poll 36N 86W NIGHT # Pause until local dawn, then exit $ ./sunwait wait sunrise 36N 86W ``` Another recommendation was to fetch some JSON object from [`api.sunrise-sunset.org`](https://api.sunrise-sunset.org). I've used their tracker some of the time, and I didn't know there was a public API. But more interestingly, the chatbot recommended a tool called [`jq`](https://jqlang.org/), which is "like `sed` for JSON data." That seems dead useful as well. The "easy solution" was also an apparently-to-me undocumented query string to [`wttr.in`](https://jqlang.org/). That website sends a `pre`-formatted web page to a browser, but an ascii-box drawing to `curl`, probably with some `User-Agent` matching. A rare AI success for me. But a strategy that might be useful to other people: not asking for a fact, but asking for *a tool* that will help me find the fact *myself*. The way to avoid AI hallucinations is to ask for *sources which address your question*, rather than just asking for the answer to your question directly. A wrong answer pollutes the memory; a wrong citation just wastes a few minutes. It occurs to me that asking "who do you know who could help me with my question?" is a very graduate-degree approach to learning things, perhaps not shared by other folks.