AI Summary
→ WHAT IT COVERS A16z's Jennifer Lee speaks with Fal co-founders Gorka Mirdsevan and Batuan Tashkaya about H3 Max, a post-trained open-source video model achieving 35x speed improvements over base Minimax H3, enabling real-time generation, two-minute scene memory, and professional camera and lighting controls for Hollywood workflows. → KEY INSIGHTS - **Post-training efficiency stack:** Combining RL-based step reduction (50 steps down to 20), kernel optimization, and hardware utilization improvements (40% to 80% MFU) compounds into order-of-magnitude gains. H3 Max Turbo generates five-second video in 1.5 seconds at half the cost, making further speed optimization less valuable than pursuing quality and controllability improvements. - **Real-time video memory architecture:** H3 Max Director maintains scene coherence by attending to compressed representations of the prior two minutes of generated video, then layering an evolving system prompt for context beyond that window. This enables continuous streams up to 60 minutes where characters, environments, and camera positions remain consistent across scenes. - **Blender-plus-AI workflow for 100% controllability:** VFX professionals render low-resolution scene layouts in Blender, then pass that video as a reference input to H3 Max. This combination gives creators near-complete control over composition and motion, and GPT-integrated Blender scene generation further automates the pipeline for studio-level production work. - **Structured camera control via JSON input:** Fal's post-training infrastructure conditions H3 Max to accept precise camera trajectory descriptions as structured JSON, specifying position and angle at each timestamp. The model treats this as the sole source of truth, eliminating camera hallucination and enabling 3D scene reconstruction from a single input reference image. - **Hollywood is Fal's fastest-growing segment:** Studio adoption was near zero one year ago and now dominates Fal's generative media conference attendance. Studios need targeted point solutions — video extension, camera adjustment, lighting changes — not full generation from scratch. Fal addresses legal barriers by offering US-hosted models including Stable Diffusion and supporting custom IP unlocking for studio-owned content. → NOTABLE MOMENT After H3 Max launched on a Saturday, a Fal engineer began streaming continuous AI-generated video from his personal laptop on Twitch using prompt tricks to maintain narrative coherence — an unplanned demonstration that the model had crossed the real-time generation threshold without any internal coordination or preparation. 💼 SPONSORS None detected 🏷️ Generative Video, AI Inference Optimization, Real-Time AI, Hollywood AI Workflows, Post-Training
