The pieces prose is built from. Each one consumes only the semantic tier of the tokens — re-skin the palette and all of them follow.
Five variants; title replaces the variant's default label (Note / Intuition / Warning / Important / Abstract — the same defaults the quote syntax uses, localizable through labels). The kitchen-sink shows the quote-syntax and raw-HTML spellings of the same markup.
Inline badges for status and categories, in the four variants PaperCard tags and HubCard chips share: stable beta research new — or a custom tone for topic palettes: custom tone
One grid, three skins (plain, card, cmp). Cards:
Tokens first. Import order is part of the contract: tokens, then the content layer, then site chrome.
Semantic layer only. Components never touch raw colors — re-skin the palette and they all follow.
And the compare skin, made for before/after panels:
Without a guard — an unclosed ** ships to production as two literal asterisks.
With the guard — the build fails at file:line:column with a caret under the marker.
A markdown list on a vertical rail:
file: dependencies.tokens.css, base.css and (for a wiki) browse.css in your global stylesheet.siteMarkdown() into astro.config.The survey workhorse — alias, full title, venue tags, author meta, links, and two dedicated slots:
Recurrent models process tokens one after another — long-range dependencies fade, and nothing parallelizes. Key idea: attention alone, no recurrence — every token attends to every other in one step, and the whole sequence trains in parallel. The architecture nearly every large language model since has built on.
Under a deep-dive heading, a one-line plain-language gloss (.speak) tells the reader what a module does before the mechanism:
Attention is a soft lookup table: every token asks "who is relevant to me?" and gets a weighted mix of the answers.