fal.live lets viewers vote on an endless AI-generated show

A crowd of strangers is directing a television show that has no writers, no actors, and no fixed script, just a stream of AI-generated video that reshapes itself every few seconds based on what the audience votes for next.

That's the premise behind fal.live, a browser-based platform from fal, the infrastructure company better known for its API layer serving image and video generation models to developers. The service was pulled offline shortly after its initial launch so the team could refine output quality, then relaunched as an open platform anyone can use for free.

Under the hood, fal.live runs on H3 Max Director, an autoregressive variant of MiniMax's H3 Max model. The key detail is memory: the system retains up to two minutes of prior context, letting scenes flow into one another instead of resetting with each new clip, which is the norm for most text-to-video generators today.

Viewers submit prompt ideas, vote on which one plays out, and watch the result render in near real-time, with new scenes appearing within seconds of a vote closing. Every submitted prompt passes through layered automated moderation before it can influence the stream.

The platform currently runs several parallel channels spanning different formats, including sitcom, anime, cartoon, hand-drawn styles, and ongoing serialized storylines, each generated live rather than pre-rendered.

From a phone experiment to a public platform

The idea traces back to indie developer Pieter Levels, who built a prototype called "Infinite Slop," a chat-directed infinite video stream he first tested informally on his phone. A fal engineer later livestreamed that concept on Twitch, and the response appears to have pushed fal to turn it into a standing product.

fal.live turns collective prompting into a persistent, evolving video feed, rather than a library of isolated clips.

Nothing in the current release requires installation or an account beyond browser access, which lowers the barrier for casual viewers to jump in and start voting rather than just watching. For fal, the platform also doubles as a live showcase of its underlying generation infrastructure, running continuously in front of an audience rather than sitting behind a developer API.

Whether continuous, crowd-steered video generation holds attention beyond novelty use is still an open question, but the two-minute memory window marks a concrete technical step past the disconnected, short-burst clips that have defined most consumer text-to-video tools so far.