The feed never resets: each new stretch is built from what the model has already generated, and it can hold as much as two minutes of context.
MiniMax released H3 last month. Fal’s version is tuned for cost and quality, runs on the company’s own inference engine rather than the endpoint MiniMax operates, and is 35 times faster than it by Fal’s own account.
Speed is what had kept generative video in short clips. Even drastic shortcuts fell well short: consistency models compressed a 30-second generation into a single second and still delivered only about one frame per second, far below anything a person will sit and watch.
Once a generator turns out frames faster than a viewer gets through them, it can simply be left running, and Fal employees left it on in public as an always-on channel. Twitch and YouTube both pulled it at once, so Fal stood up its own service at fal.live — a Twitch Plays Pokémon for generated video, steered by whoever is watching.
The feed runs on H3 Max Director, a continuous version of MiniMax’s H3 Max. The steering runs through prompts written by an LLM that viewers vote up.
Outside the company, Pieter Levels runs Infinite Slop, a round-the-clock channel on Fal’s tuned model that viewers steer the same way, with a slot called the Slop News Network turning breaking X and Hacker News stories into segments. Levels puts the cost of keeping it on air, with no sponsor, at about $4,000 a day.
Latent Space, watching a few seconds of it, called the result pure slop — a fever dream with no plot that nobody would actually sit through. Its conclusion runs the other way too: this is the worst the technology will ever be, and the feed is proof that video both faster than real time and good enough to use can be made.