fal launched fal.live, "a new platform for infinite, interactive AI livestreams. Pick a channel, prompt what happens next, and watch it generate in real time. You aren't just watching the show. You're directing it." Channels like Chaos and Sitcom run as continuous broadcasts; viewers vote on what happens next, and the winning prompt is on screen in seconds. The launch capped a five-day sprint: H3 Max on August 27, H3 Max Live and the Director model on August 29, and fal.live as the consumer version two days later — built in part because the first experimental streams were kicked off Twitch and YouTube. The engine is H3 Max Director, autoregressive and natively continuous with a context window of up to two minutes: where text-to-video and image-to-video stitch independent clips, Director generates each moment conditioned on the footage before it, so shots tie together natively, characters and sets persist, and the stream never has to cut. It reads like the opening of a new genre: once generation runs faster than real time, a video model stops being a renderer and becomes a broadcast, with the audience as the writers' room. Whether fal.live itself endures matters less than the demonstration — interactive, live, natively continuous generative video now exists in public.