About

What this is

transcripts.wiki is a knowledge base built from podcasts and talks — philosophy, history, science, politics, economics, health, and technology. An LLM reads each source, extracts the key ideas, and integrates them into a persistent, cross-linked wiki — updating concept pages, noting contradictions, building synthesis across sources. You can browse the resulting pages or query the AI chat, which answers strictly from the wiki content.

The primary reading surface is HTML. Each source is rendered as a clean, styled transcript — synthesised from auto-generated captions into something actually readable. The wiki pages (transcripts, concepts, themes) are the synthesised layer on top of that.

What it is not

Not a summary site. Not a generic AI chat. Not a transcript host in the archival sense. The value is in the cross-linked synthesis and the grounded AI chat — not in reproducing source material. Each transcript page links to the original video or episode.

How to navigate

The wiki is organised into several layers:

• Episodes — one summary page per episode, with key ideas and cross-links.
• Concepts — atomic ideas extracted from the corpus. Each concept page synthesises everything the wiki knows about that idea across all sources.
• Themes — curated synthesis pages that span multiple concepts and episodes, organised around a broader argument or question.
• Speakers — guests and presenters who appear in episodes. Speaker pages link to all their episodes.
• People — notable figures discussed or cited in episodes who have not themselves appeared as speakers.
• Notes — deeper source-grounded reading notes for idea-dense episodes, using the Adler four-questions frame.
• Quotes — portable verbatim quotes extracted from episodes, grouped by speaker.
• Books — books mentioned across the corpus, with notes on what speakers said about them.

The AI, and its limits

Two grounded assistants sit on the home page. Ask answers a question in one shot; Teach me works through a topic Socratically, one idea at a time. Both are grounded: they answer only from passages retrieved from these wiki pages, and they show the sources they drew on. The answers are written by Google's Gemini 2.5 Flash; a separate open embedding model does the retrieval — the finding of relevant passages.

Treat them as a way in, not the whole story. Each answer is built from at most a handful of fragments, so the assistant can miss things the wiki actually covers, and — like any language model — it can be confidently wrong. It runs on a free tier, so it may occasionally be slow or unavailable. When you want the substance, read the pages: they hold the full synthesis the fragments are cut from.

Inspirations

The LLM Wiki pattern

The architecture of this wiki — immutable raw sources, LLM-owned wiki layer, persistent compounding synthesis — comes directly from Andrej Karpathy's LLM Wiki pattern (April 2026). The core insight: instead of re-deriving knowledge from raw documents on every query, have the LLM build and maintain a persistent wiki that compounds with every new source. The wiki is the product; the LLM is the maintainer.

In June 2026 Google Cloud formalised the same pattern as the Open Knowledge Format (OKF): a vendor-neutral specification for knowledge as a directory of typed Markdown files with YAML frontmatter, crediting the same Karpathy pattern. Built on that pattern months earlier, this wiki already matches the shape OKF describes — every page carries a type, and the root holds an index.md catalogue and an append-only log.md.

HTML as a reading surface

The decision to use HTML rather than Markdown as the primary reading format was inspired by Thariq Shihipar's "The Unreasonable Effectiveness of HTML" (May 2026): HTML files produced by AI trade a document you would skim for one you would actually read. Andrej Karpathy agreed : asking an LLM to structure its output as HTML and opening the result in a browser is simply a better reading experience than a wall of Markdown.

The teaching workspace

The wiki's learning layer — the private /learn skill that teaches from the corpus over many sessions, and the public /teach skill that turns a topic into one shareable HTML lesson — grew out of Matt Pocock's teach skill . His insight: treat a directory as a stateful teaching workspace and produce each lesson as a beautiful, self-contained HTML file, tied to why the learner cares. The wiki adapts it by grounding every lesson in the trusted corpus and adding spaced retrieval for durable memory.

Sources

The corpus is growing — see the Episodes index for what is live.