How This Handbook Works
This Handbook is the current best-understood picture of everything processed into Duke's Tech Knowledge Database. It is organized by TOPIC, not by the date material was received, so you can always find the up-to-date picture on a subject in one place.
It is rebuilt — not just appended to — each time new source material is merged in. That means sections may be reworded, reorganized, or strengthened over time as more evidence accumulates, even if no single new fact was added.
Structure
- General Operating Principles — standing rules that apply regardless of topic. This section starts empty and fills in only as genuine cross-topic lessons emerge.
- Topic Sections — each with three parts: Core Principles (the underlying facts/patterns), Key Facts & Examples (concrete extracted detail — numbers, comparisons, named tools), and Source Articles (traceability back to Notion IDs).
- Change Log — at the very end. Every merge gets a row: Source ID, date, what was added, which section. This is what makes the Handbook auditable rather than a black box.
Traceability
Every claim in this Handbook can be traced back to a specific Notion database row via its DK-### ID. If you want the full original summary behind any point made here, look up that ID in the Duke's Tech Knowledge Database Notion database — the full transcript is attached to that row.
What's new in v15
This rebuild merges sixteen entries (DK-137 through DK-152), filed together on 20 October 2026 after finished drafts had waited up to eighteen days for a reading. Six of the sixteen are long panel podcasts, and that shapes how the batch has been handled: one guest appears in four entries and one panel in five, so a claim repeated across them has been counted as one source, not several. The batch argues more than it reports, and the arguments are the value. Jensen Huang says AI is software and every incident an engineering failure; Geoffrey Hinton, replying, concedes his own risk figure is a signal rather than a measurement (DK-137, DK-143). Anthropic put model welfare into its usage policy, and Mustafa Suleyman's objection is that the speculation itself becomes training data (DK-152, DK-142). Investors argue the AI trade's real weakness is financial — fixed-price contracts on floating-rate debt — while tokens get cheaper and the hardware beneath them dearer (DK-150, DK-138). One new General Operating Principle: liability. Who carries the cost of harm turns out to be the question underneath the safety argument in four separate topics, from sources that do not share a stage. Two existing principles were sharpened: test a figure against its own episode before testing it against the world, since six figures in this batch contradict themselves; and count independent sources in people, not videos. Several things were kept out or held back on purpose. Allegations about Bill Gates made on an investor podcast are excluded entirely, because there is nothing to check. A pattern that looks like a principle — AI cracking the fields where answers can be verified first, now seen in mathematics, physics and biology — stays inside AI Model Landscape, because all three accounts come from the same panel. Two Anthropic disclosures that reached the archive only through television anchors are recorded as unverified until traced to what Anthropic actually published.
Version 15, compiled 20 October 2026. 151 source entries (DK-2 – DK-152). Claims are recorded as their sources framed them; figures described as claimed or reported are not independently verified.