Changelog · How Lucid is being built

Here's the thing I noticed: good content dies the moment you share it.

Drop a read into a group chat and the platform grabs a dull default thumbnail. The best content ends up looking like spam — nobody clicks.

So every page on Lucid now grows its own card. Share any read and the link auto-renders a branded card with the title, the core takeaway, and the speaker.

It's generated on the fly, and I hit a CJK font trap along the way (everything rendered as tofu boxes) — fixed with on-demand font subsetting.

Packaging is part of the content.

Today's work is invisible but critical: making Lucid legible — to search engines, and to AI.

Structured data, a bilingual sitemap, hreflang, an llms.txt for crawlers. Now when you ask an AI "what does Karpathy think about AGI," it has a shot at citing Lucid.

However good the content is, if it can't be found or cited, it doesn't exist.

Drew comic-style avatars for every speaker — one consistent illustration style, warm cream background, instantly recognizable without the stiffness.

The reads got a new attitude too: no more translation-ese ("This interview explores..."). Just plain talk — what they actually said and why it matters, in human language.

Understanding beats completeness.

Lucid is live.

It started from a long-standing itch: great English long-form interviews run three to five hours, dense with insight — and I never finish them. Sentence-by-sentence translation isn't understanding. Understanding is someone chewing it down, mapping the thread, and explaining it in your own language.

So Lucid doesn't do subtitles — it does "reads": core takeaways, a timeline, a plain-language explainer. Plus Google sign-in and favorites, so what you've watched and want to keep lives in "My Library."

Five hours of English, five minutes of Chinese.