[AINews] Fal’s H3 Max Live breaks the infinite videogen barrier
Fal showed video generation running faster than playback, enough to support a live, chat-directed stream.
The piece says Fal post-trained MiniMax’s H3 release and optimized it on its own inference stack, claiming a 35x speedup over the official endpoint. That crossed the practical threshold where generated clips can appear in seconds and keep up with viewing time. Fal employees and others turned it into infinite livestream experiments, including chat-driven video sequences. The author calls the current output low-quality “slop,” but treats the faster-than-real-time result as the important existence proof. Latent Space's note
The piece says Fal post-trained MiniMax’s H3 release and optimized it on its own inference stack, claiming a 35x speedup over the official endpoint. That crossed the practical threshold where generated clips can appear in seconds and keep up with viewing time. Fal employees and others turned it into infinite livestream experiments, including chat-driven video sequences. The author calls the current output low-quality “slop,” but treats the faster-than-real-time result as the important existence proof. Latent Space's note
score 6