presentation

Progressive Cramming — 3-minute video deck

A self-contained reveal.js slide deck for the short video presentation of the paper. Speech text and plan live in video_speech.md; the spec is at .omc/specs/deep-interview-revealjs-video-slides.md.

Present / view

It's fully offline — just open the file in a browser:

xdg-open presentation/index.html      # or: open … (macOS)

Or serve it (some browsers restrict autoplay/notes from file://):

cd presentation && python -m http.server 8000   # -> http://localhost:8000

Controls

• → / Space advance · ← back · Esc / o slide overview
• S speaker view (the full speech is in each slide's speaker notes)
• ↓ from the final slide → hidden Q&A backup slides (capacity scaling, attention-mass-vs-causal, solution diversity, PCA curves)
• F fullscreen · B blackout

Export to PDF / PPTX

reveal.js exports via the browser's print-to-PDF, but you must use the special ?print-pdf URL — a plain Ctrl+P prints only the current slide and drops the layout.

PDF (recommended):

1. Open the deck in Chrome/Chromium with ?print-pdf appended — ideally over a local server so every asset loads:

   cd presentation && python -m http.server 8000
   # then open:  http://localhost:8000/index.html?print-pdf

   Hard-refresh once (Cmd/Ctrl+Shift+R) so the latest CSS loads.
2. Print (Cmd/Ctrl+P) → Destination Save as PDF, Layout Landscape, Margins None, and enable "Background graphics" (required for the colors and the figure cards). Save.

Notes:

• Use Chrome/Chromium — Firefox/Safari handle the print stylesheet differently.
• A PDF is static, so the trajectory video shows its poster frame, not playback.
• The @media print block in css/custom.css re-flows the auto-fit (.fit/.grow) media for print; without it the absolutely-centered images collapse and go missing.
• Speaker-notes pages: append &showNotes=separate-page to the ?print-pdf URL.

PPTX (no clean reveal→PPTX path):

• Make the PDF above, then soffice --convert-to pptx slides.pdf (LibreOffice) — each slide becomes a full-page image.
• For an editable deck with the video playing, build one PNG per slide and embed the mp4 on the trajectory slide via python-pptx (ask for a script).

Structure

8 main slides mapped to the 7 speech beats (the trajectory beat is a vertical pair: PC1–PC2 spatial path → optimizer-steps/effort path, where the dwell-and-leap basins pop). Light interactivity: fragment builds, in-slide video that auto-plays on entry, and an interactive attention-knockout slider on slide 7.

Layout

index.html            the deck
css/custom.css         theme (paper accent palette, dark)
js/deck.js             video autoplay + knockout slider
reveal/                vendored reveal.js 5.2.1 (offline)
assets/                videos + figures (built by scripts/build_assets.sh)
scripts/build_assets.sh

Rebuilding assets

assets/ is generated from existing repo artifacts (no GPU, no retraining):

bash presentation/scripts/build_assets.sh      # trajectory videos + paper PDF figures -> PNG
bash presentation/scripts/render_artifacts.sh  # paper TABLES + TikZ schematics + repo QR -> PNG

• build_assets.sh copies the rendered trajectory videos from artifacts/experiments_progressive/sl_4096_Meta-Llama-3.1-8B_lr_0.1/ and converts the relevant paper/figures/*.pdf to PNG via pdftoppm.
• render_artifacts.sh compiles the real paper tables and TikZ schematics to tightly cropped PNGs (standalone + pdflatex + pdftoppm) so the deck reuses the actual paper artifacts, and regenerates the repo QR code via segno.

Re-vendoring reveal.js (only if reveal/ is missing)

One-time, needs network:

tmp=$(mktemp -d) && (cd "$tmp" && npm init -y >/dev/null && npm i reveal.js@5 >/dev/null)
cp -r "$tmp/node_modules/reveal.js/dist"   presentation/reveal/dist
cp -r "$tmp/node_modules/reveal.js/plugin" presentation/reveal/plugin
rm -rf "$tmp"

Note: the trajectory GIFs are large (trajectory_pca.gif ~17 MB). The deck uses the MP4s; the GIFs are only <video> fallbacks. If repo size matters, you can delete the assets/*.gif files — the MP4s still play.