fal's MiniMax H3 Max renders AI video faster than it plays

AI video crossed a strange threshold this week: it now generates faster than it can be watched.

The speed step

fal post-trained MiniMax H3 Max for throughput, and the numbers change what the model is for. A five-second 768p clip with synchronized audio renders in under three seconds, roughly 35 times the throughput of the official MiniMax endpoint. Below real time, a video model is a clip tool: you prompt, wait, download. Above real time, it becomes a broadcast signal that never needs to stop.

The first channel built on it

Pieter Levels turned the capability into a format within days. Infinite Slop is an endless interactive AI livestream: anything viewers type in chat gets generated next, and the system tries to stitch each new clip to the previous one so there is an ongoing storyline. It is chaotic and unapologetically low-grade, the name is doing honest work, but it is also a genuinely new format: television whose script is written live by its audience.

A parallel experiment in the same vein, a Rick and Morty style interdimensional cable stream, keeps getting taken down, an early sign of how quickly the copyright questions arrive once generation is continuous.

What faster-than-playback changes

Continuous generation moves AI video into territory that used to belong to games and live broadcast: media that responds while you watch. The economics are equally new, since a channel that renders itself has no production backlog and no fixed library. What to watch is whether platforms treat these streams as user content, as automated media, or as something needing new rules, because the takedowns suggest the current categories do not quite fit.