Dense Slides: Teaching Gemini to Speak YouTube: Adapting LLMs for Video Recommendations to 2B+DAU - Devansh Tandon

Source Video

Teaching Gemini to Speak YouTube: Adapting LLMs for Video Recommendations to 2B+DAU - Devansh Tandon

Method

This deck is slide-only. The existing captured video frame set supplies candidate frames, then local OpenCV rejects sponsor/title/speaker-only frames, crops visible slide surfaces, deduplicates, and saves the cropped slide images.

Cropped Visible Slides

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.97
Text source: advanced OCR rapidocr-live/bright-screen/contrast.
OCR decision: ready — Product/UI screenshot slide with multiple embedded mobile screens and small labels; OCR should handle the screenshot text better than manual transcription.

Slide text:

AIE HIGHLIGHTS EXTENOEO CHESSTIPS COUNTERPLAY CREATING

Cahs (henTey 0444 Others

100

seE Dems Karubs TIPS CHESS

Home WatchNext Shorts Search

oogle

Microsoft smol ai

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.99
Text source: agent_vision.

Slide text:

Personalized Recommendation Problem

f(user, context) ≈ recs

User Context Recs

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.98
Text source: advanced OCR rapidocr-live/border-trim/opencv-adaptive.
OCR decision: ready — Diagram slide with multiple small labels, captions, and surface names; OCR is appropriate for the small text and layout labels.

Slide text:

Home

AIE Gemini Vidoo Rotrlovsl WatchNext Search Shorts

Music

YouTube LRM Ranking Video Experiments

Base Gemini checkpoint, adapted for YouTube Itions Aligned to tasks Served across surfaces Google Draulut

Microsoft smop

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.95
Text source: none.
OCR decision: ready — Content slide with multiple small example cards, prompt/output boxes, and short IDs; OCR is likely better than manual reading for the dense embedded text.

Recreated text/layout view: open HTML recreation
AI slide classifier: content_slide confidence 0.96
Text source: advanced OCR rapidocr-live/full.
OCR decision: ready — Dense prompt-and-example slide with small text boxes and thumbnail captions; OCR is the right first pass.

Slide text:

Us.Android Userdemo 24yr old,Female OLYNPIC

AIE Watch history Context video MENS400M

CHALI 200 Champiors Menls 4x100n Final IParis

WatchNext:men's trackraces,nowomen

LRMprompt

User:region Us|24 years female devlce

tle WHAT ACOMEBACK!IMen's 4OOm Paris2024 highlights SemanticID PARIS2O24 7.0 7.00

Watch history: SID1Taylor Swift 100 260.00 Kria Hni 100320.00 401260.008 LRM:findsrelated women'sraces UNCATCHEEnwordnecod40m hudes|Pais Olymgics N8C toots Sydey McLaoghin-Leronewas U5woendosingincedle in40lylo fhOlyict Paris Olymgics NBC Ipots

Google Yolibe

Microsoft smol?

Classification audit: raw/sources/slide-ai-classification/dense/LxQsQ3vZDqo/audit.json