Gradient Descent as a Small Practice
A short note on why search, dark mode, and gradient descent all feel like small adjustments toward a clearer site.
Gradient descent is usually introduced as an optimization algorithm, but I like it as a working metaphor for building a personal site. You do not need the perfect first step. You need a surface, a signal, and a way to take the next step with less confusion.
Today the site gained two small pieces of infrastructure: search and manual dark mode. Search makes the notebook more usable once posts, lab notes, and fragments start to accumulate. Dark mode makes the reading surface less rigid; it lets the same text sit comfortably in different contexts without changing its voice.
Both features are adjustments, not reinventions. Pagefind indexes the static output after Astro builds the site, so the writing stays as Markdown and MDX. The theme toggle stores one small preference and lets the design tokens do the rest. The implementation is intentionally plain because the goal is not to create a settings system. It is to make reading and returning easier.
The gradient descent metaphor matters here: every published page gives feedback. A rough note, a broken link, a search result that feels noisy, or a code block that looks wrong in dark mode all become gradient information. The site should improve by following those signals carefully, one small update at a time.