AI Vision Data Squeezed to 2% — Answers Stay Just as Smart

Modern AI that "watches" video works in two parts: a visual encoder that sees the footage, and a language model that thinks about it and answers questions. These parts often live in different places — your phone or a local device, and a distant cloud server. Between them flows a massive stream of data, called "visual-token embeddings." Researchers call this burden "AI traffic," and it can be enormous.

A team from Finland tested a clever solution: use a standard compression system designed just for neural networks (called NNC, essentially a digital squeeze for AI data) to shrink that traffic. They compressed the visual data flowing inside Qwen3-VL-8B, a large video-understanding model, while leaving its knowledge untouched. The result shocked them: the data size dropped to just 2% of the original, yet the AI's answers to video questions stayed almost as accurate as when it received full-quality data.

Why does this work? The compressed data looks messy on the inside — heavily rounded-off numbers with big errors. But the AI doesn't need precise pixel-perfect information. It cares about rough shapes, layout, and overall structure — the "big picture" rather than the fine details. That's why they recommend optimizing for task performance, not for making the data look clean to human eyes.

What does this mean for you? Video AI tools — like assistants that summarize meetings, analyze security footage, or caption clips — could get dramatically faster and cheaper to run. Carriers would carry far less data, and users on slow networks could still get smart answers. The catch: 98% is the cliff edge. Squeeze harder, and accuracy collapses. And this was tested on one model and one set of benchmarks, so real-world rollouts will need more testing.