Methodology

How the Library is built

The TWiST Intelligence Library turns conversation into structured knowledge. Each episode travels through a four-stage pipeline before it becomes a page you can link to.

1
Source

Transcript capture

We start from the authoritative video on the This Week in Startups channel. Each episode is transcribed with speaker diarization, then aligned with the YouTube timestamp so every claim has a traceable source moment.

2
Entity extraction

Founders, companies, VCs, concepts

A structured-output model reads the transcript and extracts five classes of entity: guest founders (with role + company), companies discussed (with description + metrics), venture capital firms and partners (with investment context), tactical concepts (with a canonical formula), and axioms the guest stated as general rules.

Quantitative claims and proof points are preserved as the guest stated them, with a from transcript label. These are verbatim figures that can be checked against the source video.

3
Synthesis

Thesis, axioms, contrarian take, playbook

From the entity set we synthesize the episode’s core thesis (one paragraph), 3–5 axioms (operator rules), 2–4 decision rules (if/then heuristics), a contrarian take, a four-step operator playbook, and a one-line formula. This is the structure Karpathy proposed for an “LLM wiki” — compile the signal, keep the source.

4
Cross-linking

Every mention becomes a hyperlink

When a founder’s thesis references another company, investor, or concept that exists in the library, the mention is auto-linked. This is the difference between a search index and a wiki: you can trace an axiom on one page to a founder on another to a portfolio position on a third.

Scope of this release

What the pilot slice covers

10
Episodes
14
Guests
13
Companies
5
VC Firms
12
Concepts

This is intentionally a narrow slice. Thin mentions and side references do not get standalone pages; they appear only where they belong in context. The goal of the pilot is depth per page, not breadth of index.

Roadmap

What comes next

Expansion will work backward through the TWiST archive one season at a time. Two additional passes are planned after corpus expansion: a citation layer (timestamp anchors on every claim) and a longitudinal view (how a founder’s thesis or a company’s metrics evolve across multiple appearances).

Questions, corrections, or requests for specific episodes to prioritize? The Library is a living project; feedback shapes what gets compiled next.