advantage / artificial-intelligence — courseware track
A short courseware track that reads SLeeLa's AI code through the advantage model: which approach should I draw on, and what must I check before I do?
It follows the house style of the tutorial series
(standalone lessons, prev/next headers) and points at the real code in
/artificial-intelligence. Nothing
here claims a measured result; numbers in exercises are illustrative.
The track
| # | Lesson | You will learn |
|---|---|---|
| 01 | Evidence before trust | Treat a model score as evidence: reward vs baseline, confidence n/(n+prior), "unknown is not zero" |
| 02 | Assumptions of a model | Record held/doubted assumptions for neural nets, character recognition and LLMs |
| 03 | Counsel: choosing an AI approach | Rank approaches with mean × confidence × trust and read the cautions |
| 04 | The high speed trader | A runnable model: 5,000+ evals/s, about 2,000 trades/min (advanced-medium) |
| 05 | Pragmatics and evolution | The 24 pragmatics in three phases; online weight learning (advanced) |
| 06 | Fail closed, and the brilliant | Safe halting, refusal over clamping, and the insight channel (advanced) |
| 07 | The Moral Evaluator | 41 yes/no points, cardinal vetoes, merit and risk (advanced) |
| 08 | Tall Science | The United States as a cataloged Item; nouns and names read respectively (advanced) |
| 09 | The medium and the bridge | A user program loads the medium, then AI modules; messages flow both ways (advanced) |
AI modules are Advantage™ components under lib/advantage/modules; a user program is in user-program.
Prerequisites
- advantage README (math core, store, counsel)
- SUPPORTED.MODELS.md §1 (the foundation)
- Optional: neural-nets, character-recognition, large-language-models
Honest limits
- Illustrative data only; replace with your own evaluation outcomes.
- Fetched/remote knowledge is untrusted and never scored (see
enrich). - Lessons 01-03 add no code; lessons 04-09 teach the Advantage™ modules in
lib/advantage/modules.