[AINews] Fal’s H3 Max Live breaks the infinite videogen barrier

Fal took MiniMax’s H3 release from last month, post-trained it for cost and quality, then tuned it for its own inference engine. The company says that brought a 35x speedup over the official endpoint. That mattered because generative video has always been trapped by time: even when generation got faster, it still lagged far behind the pace of human attention.

The breakout moment was a live stream that kept generating continuously. Ethan Mollick noticed it first, and then fal employees turned the idea into an “infinite” Twitch stream. Twitch and YouTube quickly kicked Fal off the platform, so the company moved the experience to its own live video service, which it pitched as a kind of “twitch plays pokemon” for video generation.

The stream itself sounds more like a stress test than a product demo. Fal’s own framing is that viewers can upvote LLM-generated prompts, and the system is powered by H3 Max Director, an autoregressive continuous version of H3 Max with up to two minutes of context. In parallel, Fal also launched Reference-to-Video for MiniMax H3 Max, with early preview claims of real-time factor 1 at 768p.

What makes this interesting isn’t that the output looks good. The source is blunt that it doesn’t. It’s that live, audience-steered video generation now exists at all, even if the result is chaotic, low-quality, and barely watchable. That is still a line crossed, and the industry tends to notice those lines only after they’ve already been crossed.