Knowledge base
Turning audio into knowledge
Field notes and essays on the method: pull a transcript, compile it into a knowledge network, query it, build from it. Start here.
Method & workflow
The knowledge network
Pull a transcript, compile it into a linked network, then work every question and every post out of that network. The loop we run, and why it compounds.
Read →The compile step
How a model turns a raw transcript into a linked encyclopedia page: one concept, sources in the header, every claim traceable to a line.
Read →Compiled once, not retrieved every time
Why we don't RAG our own knowledge - and the cases where retrieval still wins.
Read →Cross-domain links are the product
The insight lives where a podcast claim connects to an article claim. Only a linked network surfaces it.
Read →From wiki to content
Writing a post the thinking is already done for. The artefact is a view onto the network.
Read →
Foundations
Read, don't listen
Audio is opaque. Text is searchable, quotable, and the only form an LLM can actually read. The case for transcripts, anchored on Karpathy.
Read →Timestamps make a transcript navigable
A transcript without per-line timestamps is half a transcript. Jump to the moment, quote the sentence.
Read →Who said what
Speaker diarization, and why attributing a claim to a speaker is what makes a transcript usable as knowledge.
Read →A transcript you own vs one you rent
Audio behind an app disappears. Plain text you keep is durable, greppable, and portable.
Read →
Sources & ingestion
Lexicap at home
Transcribe a whole back-catalogue overnight - the effort Karpathy spent by hand in 2022, automated on your Mac.
Read →Scraped captions vs real transcription
Why YouTube's auto-captions aren't enough: no punctuation, no speakers, rolling duplicates.
Read →Follow creators, don't chase links
Subscribe once; new episodes and reels arrive already transcribed.
Read →