Sep 17, 2026 · 40 min
Real-time video pushes generative AI toward interactive worlds
The Next Frontier of AI Video Is Control
The episode shows how cheaper, faster video generation is moving AI video from short clips toward persistent experiences and professional production systems.
- 1Post-training and systems optimization can sharply reduce video-generation latency and cost while preserving quality.
- 2Real-time streaming opens interactive formats built around memory, continuous scenes, and audience-directed events.
- 3As speed improves, studios increasingly need precise controls, custom workflows, legal safeguards, and data-residency options.
Don't miss
The guests recount how spontaneous Twitch experiments turned continuous H3 Max generations into audience-directed live experiences.
The brief
Jennifer Lee speaks with FAL co-founders Gorkem Yurtseven and Batuhan Taskaya about H3 Max, a post-trained version of MiniMax’s open-weight video model.
The guests argue that speed comes from a stack rather than one breakthrough: post-training, fewer diffusion steps, better GPU use, hardware choices, and streaming infrastructure.
H3 Max’s low latency quickly enabled continuous experiments on Twitch, exposing the limits of last-frame continuity and prompting work on memory, scene transitions, and interactive worlds.
The conversation then shifts from speed to control, covering references, LoRA fine-tunes, camera movement, lip sync, Blender workflows, and specialized professional pipelines.
Hollywood’s demand is pushing FAL toward focused tools, custom intellectual property, legal and data-residency requirements, and production systems rather than one universal director model.
MiniMax H3
TwitchListen to the full episode and explore every guest, topic, and moment on PodLume.