# Yuhang > Notes on everything consumed and worth keeping — papers, articles, podcasts, reports, talks — linked densely enough to navigate as a graph. Plus canvas sketches, photographs and side projects. Every note carries its source, a maturity marker (`seed`, `growing`, `evergreen`) and topic tags. Notes link to each other with wikilinks; those links are the edges in the graph. Some notes exist in both English and Chinese. They are the same note: one slug, one set of links, one node in the graph. Each translation records a hash of the version it was translated from, so a reader can tell whether it is still `current` or has gone `stale` — see `translations` and `source_lang` in /api/notes.json, and prefer the source when the two disagree. ## Structured data - [/api/notes.json](/api/notes.json): every note with topics, source, summary, inbound and outbound links. - [/api/graph.json](/api/graph.json): the knowledge graph as nodes and edges. - [/rss.xml](/rss.xml): notes as a feed, newest first. - Any note is also available as its original markdown at `/notes/.md`. ## Notes (5) - [How notes work here](/notes/how-notes-work/) (own, evergreen, 2026-10-16): The conventions this knowledge base runs on — frontmatter, wikilinks, topics and maturity. Read this one first. - [Attention is all you need](/notes/attention-is-all-you-need/) (paper, growing, 2026-10-14): The Transformer paper. Drops recurrence entirely, and in doing so makes sequence modelling a parallel problem rather than a sequential one. - [Loss landscapes and terrain](/notes/loss-landscapes-and-terrain/) (own, seed, 2026-10-12): Seed. The slope-reading intuition from backcountry skiing keeps wanting to be a statement about optimisation geometry, and I have not made it rigorous yet. - [Terrain as a first filter](/notes/terrain-as-a-first-filter/) (own, seed, 2026-10-10): Seed. Terrain choice does more risk reduction than snowpack assessment, because it is the variable you actually control. - [Sequence and sight-reading](/notes/sequence-and-sight-reading/) (own, seed, 2026-10-08): Seed. Reading a boulder problem from the ground is planning under a model you know to be wrong; the skill is in how cheaply you revise. ## Topics (5) - [Skiing](/topics/skiing/) (3 notes): Snow, slope angle, and the long argument between what is skiable and what is survivable. - [Neural Networks](/topics/neural-networks/) (2 notes): Architectures, optimisation, and what a loss landscape has in common with a mountain. - [Climbing](/topics/climbing/) (1 note): Rock, sequence, and the specific pleasure of a problem that only yields to the right order. - [Generative Visuals](/topics/generative/) (1 note): Fields, flows and noise — making something worth looking at out of a few lines of maths. - [Snow Science](/topics/snow-science/) (1 note): Layers, bonds, temperature gradients — why a slope holds, and then doesn't. ## Projects (1) - [yhzhan.me](/projects/yhzhan-me/) (active): This site. A knowledge graph over daily reading notes, a lab for canvas sketches, photo albums and a project index — one static build, no database. ## Lab (2) - [Contour field](/lab/contour-field/): A scalar field traced with marching squares, deformed live by the cursor. The terrain actually changes — the lines are re-extracted every frame, not warped as an image. - [Flow lines](/lab/flow-lines/): Particles advected through the curl of a scalar potential, accumulating on a canvas that is never cleared. A long exposure of a field that does not exist. ## Optional - [/about/](/about/): who keeps this and what it is for. - [/colophon/](/colophon/): how the site is built, set and coloured.