The compile step

Transcribing audio gets you text. It does not get you an answer to anything - the transcript is still organized the way the speaker talked, not the way a question is asked. Ask it something and you are back to reading the whole thing.

The method treats the transcript as raw material, not a deliverable. The deliverable is what a model writes after it reads the transcript: a page organized around a concept, not around the order someone happened to say it.

The compile pass does not summarize a transcript into one blob. It pulls out the concepts the transcript actually contains and gives each one its own page. A podcast about three different topics becomes three pages, not one long recap.

This is the part that looks slow and pays off later. A single page per concept means the next transcript that touches the same concept edits that page instead of writing a new, disconnected summary next to it.

Every page names the other concepts it touches and links to their pages. A page on 'context windows' links to the page on 'attention', which links to the page on the paper that introduced it - because that is how the source material actually talked about it.

A pile of unlinked summaries is a filing cabinet. Links turn the same material into a network you can walk - from one claim to the next related claim, across sources that were never in the same conversation.

Every page opens with a header listing the transcripts it was compiled from - which show, which episode, which source. Nothing on the page floats free of where it came from.

That header is what lets you trust the page later. Six months on, when the page has been edited a dozen times by a dozen new transcripts, you can still walk back to the exact recording a sentence traces to.

The first transcript compiles into a handful of thin pages. The tenth transcript does not start over - it edits the pages that already exist and adds the ones that are missing, and the links get denser with every pass.

That is the actual argument for compiling instead of just archiving: a library of transcripts sitting on disk is worth exactly what you read out of it today. A compiled network is worth more after every transcript than it was before.