# Notes

## User preferences

- **Generalist, not builder.** No code, no math. Wants concepts, vocabulary, and the
  judgment to read news/pitches critically. Success = sharp conversation + hype radar.
- Wants provenance. Cite claims to regulators/standard-setters over vendors or blogs.

## Teaching notes

- The core skill of this course is *classification under uncertainty*: take a vague claim
  and sort it. Lessons should end with the user able to place real-world headlines, not
  recite definitions.
- The single most load-bearing distinction: **predictive ML (old, tabular, works)** vs
  **generative AI / LLMs (new, language, hallucinates)**. Most confusion and most hype
  collapses once the learner reliably tells these apart. Taught in Lesson 0001.
- Push every new term into `reference/ai-finance-glossary.html` and hold to those
  definitions — the field abuses words ("AI", "algorithm", "model") as marketing.
- Anti-hype is a first-class goal, not a footnote. "AI-washing" is a real, named risk.

## Session log

- 2026-07-27 — Lesson 0002 (the application tour: five stops — fraud, credit, trading,
  robo, document work — each sorted with the two-AIs lens; new tool = the *fit test*:
  labelled history? clear answer? future like past?). Opens with retrieval of lesson 1
  (spacing). Key frame planted: excitement of a use case ≈ inverse of its fit. Indexes
  updated. Lesson 1 evidence still pending — user hasn't reported the headline-sorting
  drill; probe next session before treating the two-AIs split as established. Next:
  the risk lens (lesson 03, planned card added).
- 2026-07-25 — Workspace created. Mission set (generalist / literate track, see LR-0001).
  Grounded in FSB/GAO/CFTC/OECD. Glossary started. Lesson 0001 (the two AIs: predictive ML
  vs generative AI, + how to classify a headline and name its risk) delivered. Course +
  root index wired. Next candidate: the application tour (fraud/credit/trading/robo) OR
  the risk lens (bias, black-box, systemic correlation) — pick by what the user reacts to.