mirror of
https://github.com/harvard-edge/cs249r_book.git
synced 2026-07-20 21:20:19 -05:00
Vijay Janapa Reddi
38a6693724
Pedagogical alignment pass across all 35 Beamer decks (Vol I + Vol II), auditing each deck against its source textbook chapter and the project's slide/speaker-note standards. Content alignment: - 92 numeric/equation corrections to match the chapter's canonical (mlsysim-computed) values: e.g. Iron Law of Scale restored to 4 terms (T_sync coordination term), 175B checkpoint 1,050->1,750 GB / 4.5->7.6 PB, H100 ridge point made internally consistent (989 FP16), GPT-4 memory FP32->FP16 label, twelve->thirteen invariants in the conclusion. - 119 chapter-grounded additions: missing war stories (Amazon recruiting, Knight Capital, OVH fire, OpenAI gang-scheduling, Google Photos, etc.), thinner-than-chapter Fallacies & Pitfalls slides, and named equations / lighthouse archetypes folded back in. - Course-overview rebuilds: Bitter Lesson, Software 2.0, silent-vs-loud failure (Vol I); Three Walls and Iron Law of Scale through-line (Vol II). Lecture/render quality: - Added [N min] timing tags to 958 speaker notes that lacked them. - Reconciled learning objectives in 17 decks to the chapters' current LOs. - Fixed 3 compile failures (corrupt image regen, \mlsysfocus un-nesting, literal # escaping) and drove every overfull hbox/vbox to zero across all 35 decks, preferring slide-splits over content cuts (no content removed). - New thirteen-invariants figure for the conclusion deck. All 35 decks compile with 0 errors and 0 overfull boxes.