What's this about?
DeepMind is bringing agentic video understanding to its latest Gemini models: better accuracy analysing video while using up to 88% fewer tokens.
Why is video so expensive in tokens?
Because a model sees video as frames, and every frame becomes a large block of visual tokens. A minute at even a few frames per second is an enormous input, which is why video understanding has been priced out of most workflows. "Agentic" here means the model decides what to look at — sampling, seeking to relevant moments, examining closely only where it needs to — rather than ingesting everything uniformly. That is the same shift as agentic search versus stuffing the whole corpus in context.
Why is it worth noticing?
1,106 likes
- An 88% token reduction is the difference between video analysis being a demo and being affordable.
- It pairs with the Locate Anything in Videos paper from yesterday's digest — grounding and efficient understanding are the same problem approached from two sides.
- Watch whether accuracy holds on long videos, where selective attention has the most to miss.