You already know: image encoders and token counts per image (Computer vision); ASR with timestamps (Audio & speech); chunking, embeddings and retrieval (D7); cost per request (D12).
Why this matters for Atlas. Video is images × time × audio, and that multiplication is what makes it hard. A one-hour lecture at 30 frames per second is 108,000 frames; even at a few hundred tokens per frame that's tens of millions of tokens, far beyond any context window and any budget. Every practical video system is therefore a story about reduction: sample frames sensibly, detect scene changes, lean on the audio transcript, compress each segment, index it, and only look closely at the parts a question needs. This lesson builds frame sampling and shot detection, reviews how temporal models work, and designs a lecture-video question-answering pipeline with timestamps.
1. The size problem and frame sampling
The question. How many frames does a model actually need to see?
Start from nothing. Adjacent frames are nearly identical: video is massively redundant. Strategies, from cheapest to smartest: