Reimers: a 10h 4K movie has 7T pixels — single-vector video embeddings can't work

Nils_Reimers · x · 2026-11-17

Nils Reimers breaks down why enterprise video understanding is hard: a 10-hour 4K movie has 7 trillion pixels and the relevant info can be anywhere, so compressing it all into a single 1024-dim embedding vector fails. Information can live on-screen, in audio, or spread temporally — you may need to combine something visible at minute 5 with something said at minute 45 and an event at minute 90. That's why vendors cap video length (AWS Nova at 30s, Gemini embeddings at 120s per request), while most benchmarks only test short silent clips; his goal is ingesting videos of any length.

Original post →

More from Multimodal

Multimodal channel →