Skip to content

lectures.md

Latest commit

 

History

History
101 lines (87 loc) · 5.73 KB

lectures.md

File metadata and controls

101 lines (87 loc) · 5.73 KB
Raw

lectures.md — Course Parameter Spec (single source of truth)

This is the only file you edit to add, remove, or reorder lectures. The three run prompts (run1_spine.md, run2_body.md, run3_assemble.md) read it; they never ask you to edit anything else.

How it works

  • slug is a stable id and the folder name under lectures/. Never rename a slug once a lecture has an approved spine — it's the key everything joins on.
  • order sets sequence in the assembled document. Use gaps of 10 so you can insert a lecture between two others (e.g. 15) without renumbering anything. Reordering changes only assembly (Run 3); it never invalidates a spine or body.
  • papers are arXiv ids. Include figure-only / referenced papers too (e.g. a scaling-laws paper you only cite a figure from); mark them role: referenced.
  • figures is the explicit, curated placement list — Claude does not subselect or add figures. This is a placement list, not a license to structure the notes around figures (Run 2 stays prose-first regardless). Spec the subset you actually want shown.
  • page_budget_total (in defaults) is the approximate target length of the whole compiled document, in pages. Run 1 apportions it into a per-lecture page_allowance proportional to load. It is a target biased toward concision, not a hard cap — substance wins over the number, and overflow is flagged in a lecture's README rather than padded or amputated.
  • Changing a lecture's papers/slides/video/figures changes its param hash, which marks its spine stale (Run 1 will offer to regenerate). Changing only order or title does not.

Figure entry notation (per entry, one figure)

"<id> Figure N" paper figure · "<id> Table N" table · "slide N: <desc>" lecturer-original slide figure · "<url> | title: <t> | source: <page>" external web image — fetch and attribute to the named source/page · qualifiers (left), (right), (top half, Pythia) crop to that panel · (refer back) same image as an earlier entry in the same lecture — reference once, never re-insert · "refer back to previous lecture's figure, do not include" the image was introduced in an earlier lecture — do not re-include it; the back-reference is added at assembly by Run 3, not baked into the frozen body · (slide-only: not in <paper>) use the slide version · "A (left) + B (left)" present as one figure / tight pair · source: slide-original | paper-adapted | paper-clean provenance hint (else inferred by the ladder).

course:
  name: "CS329A: Self-Improving AI Agents"
  institution: "Stanford University"

defaults:
  page_budget_total: 20          # whole compiled document, approximate target (pages)
  density: concise               # prefer the shortest version that loses no substance
  audience: "CS300-series graduate students"

lectures:

  - slug: course-overview
    order: 10
    title: "Course Overview"
    papers:
      - "Note: for this particular lecture the goal for this introductory lecture is not summarizing papers, but only relying on the lecture transcript, while including the figures specified below with referencing the papers and sources they are coming from."

    slides: "/Users/laszlo/Library/CloudStorage/OneDrive-Personal/Stanford/CS329A/Course_Materials/course_files_export/Lectures/Fall 2025 - Lecture 1 CS329A.pdf"
    video:  "/Users/laszlo/Library/CloudStorage/OneDrive-Personal/Stanford/CS329A/Course_Materials/CS329A on 9_22_2025 (Mon)_default_b0f8e752.mp4"

    figures:
      - "2001.08361 Figure 1" # LLM scaling law
      - "https://huggingface.co/blog/assets/33_large_language_models/01_model_size.jpg | source & context: https://huggingface.co/blog/large-language-models)"
      - "slide 19 pretraining-finetuning-inference pipeline | source: slide-original"
      - "2407.21787 Figure 1" # Large Language Monkeys
      - "https://www.anthropic.com/_next/image?url=https%3A%2F%2Fwww-cdn.anthropic.com%2Fimages%2F4zrzovbb%2Fwebsite%2F58d9f10c985c4eb5d53798dea315f7bb5ab6249e-2401x1000.png&w=3840&q=75 | title: The evaluator-optimizer workflow | source: https://www.anthropic.com/engineering/building-effective-agents"
      - "2405.15793 Figure 1"
      - "2408.06292 Figure 1"

  - slug: test-time-compute
    order: 20
    title: "Test-Time Compute Scaling"
    papers:
      - { id: 2407.21787, role: core }      # Large Language Monkeys
      - { id: 2502.17578, role: core }      # How Do LLM Monkeys Get Their Power Laws
      - { id: 2408.03314, role: core }      # Scaling LLM Test-Time Compute Optimally
      - { id: 2409.15254, role: core }      # Archon
      - { id: 2001.08361, role: referenced } # Kaplan scaling laws (figure + analogy only)
    slides: "/Users/laszlo/Library/CloudStorage/OneDrive-Personal/Stanford/CS329A/Course_Materials/course_files_export/Lectures/Fall 2025 - Lecture 2 CS329A.pdf"
    video:  "/Users/laszlo/Library/CloudStorage/OneDrive-Personal/Stanford/CS329A/Course_Materials/CS329A on 9_26_2025 (Fri)_default_ceead911.mp4"
    figures:
      - "slide 5: pretraining-finetuning-inference pipeline | do not include, just refer back to previous lecture's figure!"
      - "2407.21787 Figure 1 | just refer back to figure in previous lecture, do not include" # Large Language Monkeys
      - "2407.21787 Figure 2" # Large Language Monkeys
      - "2001.08361 Figure 1 | just refer back to figure in previous lecture, do not include" # LLM scaling law

      - "2407.21787 Figure 5"
      - "2502.17578 Figure 2"
      - "2407.21787 Figure 7"
      - "2407.21787 Figure 8"

      - "2408.03314 Figure 5"
      - "2408.03314 Figure 2"
      - "2408.03314 Figure 1"

      - "2409.15254 Figure 2"
      - "2409.15254 Table 3"
      - "2409.15254 Figure 6 (left) + Figure 7 (left)"