The phrase covers three different technical products: completed clips generated faster than playback, streams that emit frames while generating, and action-conditioned worlds that react continuously.
Why the definition changed
Traditional AI video returned a file after a wait. Once generation time falls below consumption time, applications can prepare the next scene while the current scene is still playing. This changes the product from a file generator into a live system.
What does not qualify
A fast interface, a progress preview, real-time video analysis, and ordinary pre-rendered branching video are useful products, but they do not prove that new video is being generated within the live loop.
How to compare systems
For completed clips, use RTR. For streams, add time to first frame, output cadence, and jitter. For interactive worlds, measure action-to-effect latency, state persistence, and consistency over time.