How Real-Time AI Video Is Changing How Creators Work
Gorkem Yurtseven and Batuhan Taskaya of Fal explain how post-training optimization of the open-source Minimax H3 model achieved 35x speed improvements and 2x cost reductions in AI video generation.
The gist
-
Post-training tweaks like reducing diffusion steps and raising hardware utilization from 30-40% to 70-80% can compound into order-of-magnitude speed gains.
-
A new variant called H3 Max Director holds up to 2 minutes of memory, enabling real-time continuous video that users can steer with voice prompts.
-
Hollywood studios are less interested in generating full videos from scratch and more in targeted tools like camera control via JSON, lip-sync, and motion transfer.
-
Adding more GPUs beyond 8 per node hits diminishing returns, so the real efficiency wins come from algorithmic and kernel-level work instead.
Also mentioned
1:38
Minimax H3
· "The MiniMax H3 model is the first truly open-source, very capable, latest generation video model out there...."
5:28
Ideogram
· "We have been experimenting with how we can build post-training infrastructure to take an existing model,..."
5:30
Flux
· "We did one version with Ideogram, we did one version with Flux. So we have been experimenting with how we..."
15:10
Slack
· "I think we broke a record on Slack that day, how many messages were sent in the company."
20:23
H3 Max Director
· "That team just released that model under H3 Max Director which is public for people to use and I think it's..."
◎
Related episodes
Episodes covering similar topics
AMD Buys Fei-Fei Li's World Labs for $8.2B | Meta Poaches MongoDB's CEO | Bessemer Raises $5.75B
open-source models
Master AI Video Marketing w/ Magnific CEO Joaquín Cuenca Abela | AI Basics
ai video generation
More from a16z
How to Build the Next SpaceX or Tesla
Oct 29, 2026