Notes on everything consumed and worth keeping: papers, articles, podcasts, reports, talks.https://yhzhan.me/enhttps://yhzhan.me/notes/how-notes-work/https://yhzhan.me/notes/how-notes-work/The conventions this knowledge base runs on — frontmatter, wikilinks, topics and maturity. Read this one first.Fri, 16 Oct 2026 00:00:00 GMTgenerativeskiinghttps://yhzhan.me/notes/attention-is-all-you-need/https://yhzhan.me/notes/attention-is-all-you-need/The Transformer paper. Drops recurrence entirely, and in doing so makes sequence modelling a parallel problem rather than a sequential one.Wed, 14 Oct 2026 00:00:00 GMTpaperneural-networkshttps://yhzhan.me/notes/loss-landscapes-and-terrain/https://yhzhan.me/notes/loss-landscapes-and-terrain/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.Mon, 12 Oct 2026 00:00:00 GMTneural-networksskiinghttps://yhzhan.me/notes/terrain-as-a-first-filter/https://yhzhan.me/notes/terrain-as-a-first-filter/Seed. Terrain choice does more risk reduction than snowpack assessment, because it is the variable you actually control.Sat, 10 Oct 2026 00:00:00 GMTsnow-scienceskiinghttps://yhzhan.me/notes/sequence-and-sight-reading/https://yhzhan.me/notes/sequence-and-sight-reading/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.Thu, 08 Oct 2026 00:00:00 GMTclimbing