Skip to content
a16za16z

How Real-Time AI Video Is Changing How Creators Work

a16z General Partner Jennifer Li sits down with fal co-founder Gorkem Yurtseven and Head of Engineering Batuhan Taskaya to discuss what changes when generative video becomes fast enough to run in real time. They unpack the technical work behind H3 Max, fal’s post-trained version of MiniMax’s open-weight video model, and how combining model post-training with systems and hardware optimization significantly reduced generation time while maintaining quality. That speed has enabled experiments with continuous video, including streams that can remember previous scenes and respond to new directions while they’re running. They also discuss why the next challenge may be less about speed and more about control, from camera movement and lighting to characters, motion, and lip sync. And they explore what those capabilities could mean for professional creative workflows, where artists and studios need predictable tools rather than simply generating a video from a prompt. Timestamps: 00:00 - Intro 00:47 - Meet fal & the H3 Max Launch 02:45 - From Inference Platform to Post-Training: Why fal Made This Bet 06:33 - The Secret Sauce: System Work, Architecture & Cost/Latency Wins 10:44 - The Twitch Moment: Real-Time Video the Day After Launch 13:11 - How the Model Actually Remembers What Happened in a Scene 15:47 - The Economics: Serving Costs, Chip Footprint & Streaming Experiences 22:02 - Beyond Consumer: Unlocking Hollywood-Grade Controllability 28:37 - Prompting a Director: Camera Angles & Scene Understanding 35:43 - Where Video Models Go Next & the Creator Economy Shift Resources: Follow Gorkem Yurtseven on X: https://x.com/gorkem Follow Batuhan Taskaya on X: https://x.com/isidentical Learn more about fal: https://fal.ai Follow Jennifer Li on X: https://x.com/JenniferHli Stay Updated: If you enjoyed this episode, be sure to like, subscribe, and share with your friends! Find a16z on X: https://twitter.com/a16z Find a16z on LinkedIn: https://www.linkedin.com/company/a16z Listen to the a16z Show on Spotify: https://open.spotify.com/show/5bC65RDvs3oxnLyqqvkUYX Listen to the a16z Show on Apple Podcasts: https://podcasts.apple.com/us/podcast/a16z-podcast/id842818711 Follow our host: https://x.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see http://a16z.com/disclosures.

Gorkem YurtsevenguestJennifer Lihost
Sep 17, 202638mWatch on YouTube ↗

CHAPTERS

  1. 0:00 – 0:40

    Why video generation is hitting “token market fit” (and why H3 Max matters)

    The conversation opens on the idea that generative video is becoming a high-volume “token market fit” category—where individuals can productively spend massive amounts of compute. The hosts frame H3 Max as a model that could enable a long-awaited consumer breakout moment by making video fast and cheap enough to be truly interactive.

    • Generative media (like coding agents) is positioned as a high-token-demand market
    • Creators can spend thousands of dollars/month worth of tokens on video workflows
    • H3 Max is framed as enabling the next big consumer moment via real-time capability
  2. 0:40 – 1:38

    H3 Max launch context: benchmark separation and viral attention

    Jennifer and Gorkem discuss the launch moment and the striking benchmark charts where H3 Max sits far apart from other models—even on a log scale. They set up the core question: how fal achieved both major speed gains and strong quality.

    • H3 Max appears in a “league of its own” on speed/quality graphs
    • Benchmark separation is so large it’s shown on log scale
    • Internet reaction: viral tweets and rapid experimentation
    • Prompt to explain how post-training delivered speed without quality collapse
  3. 1:38 – 3:01

    Why fal bet on post-training an open-weights frontier video model

    Gorkem explains that MiniMax H3 was the first truly capable, next-gen open model that fal could legally and technically build on. Open weights created the opportunity to combine fal’s serving expertise with deeper model work to unlock big product gains.

    • MiniMax H3 is described as the first truly capable open-source modern video model
    • Open weights enabled fal to add capabilities they couldn’t on closed models
    • Model supports references and uses familiar modern video architecture
    • fal chose to “go all in” because the timing and access finally aligned
  4. 3:01 – 6:26

    From inference platform to system/model co-design: breaking the efficiency “roofline”

    The team describes how pure systems optimization typically yields only 2–3x gains before hitting a utilization ceiling. Post-training plus system/model co-design is presented as the lever that pushes performance beyond that roofline—toward order-of-magnitude improvements.

    • Systems-only optimization can raise speed ~2–3x but hits a roofline
    • Goal: maximize video pixels per chip via deep inference engineering
    • fal had already experimented post-training infra on image models (Ideogram/Flux)
    • Combining frontier open video model + learned infra unlocked ~order-of-magnitude speed
  5. 6:26 – 10:07

    The “secret sauce” stack: fewer steps, higher MFU, and an optimized multi-stage pipeline

    Batuhan breaks down the compounding optimization approach: first improve quality via post-training/RL, then reduce diffusion steps without losing quality, and optimize every stage of the pipeline. They emphasize that video inference is not one model call, but multiple components that each need tuning.

    • Key lever: reduce diffusion steps (e.g., 50→20) while preserving quality via post-training
    • Improve quality first, then apply speed optimizations to land at equal/higher quality
    • Raise hardware utilization (MFU) from ~30–40% toward ~70–80%
    • Video pipeline includes prompt expansion LLM, latent generation, VAE decode, and optional upscaling
    • Single-shot prompting (low batch, no caching) requires different LLM optimizations
  6. 10:07 – 10:57

    Hardware realities: 8-GPU node serving, Hopper→Blackwell gains, and “beyond real time” goals

    They explain that most video models run on an 8-GPU single node because scaling beyond that suffers communication overhead. Newer GPUs (e.g., Blackwell) improve wall-clock speed, which becomes critical when trying to generate multiple seconds of video in under a couple seconds.

    • Serving configuration: typically one node with 8 GPUs; scaling beyond reduces efficiency
    • Blackwell-class hardware can provide ~2–3x wall-clock improvements
    • Cost may not drop proportionally because newer hardware is priced similarly
    • Fast hardware matters most for generating video faster than real time
  7. 10:57 – 12:26

    H3 Max Turbo: trading a small quality drop for extreme real-time speed and lower cost

    fal ships an even faster variant, H3 Max Turbo, targeting near-identical quality at the high percentile while doubling speed and cutting cost further. The takeaway: at these speeds, the next frontier shifts from speed to quality and controllability.

    • Turbo can generate ~5 seconds of video in ~1.5 seconds
    • ~2x faster and ~2x cheaper than H3 Max (with a small quality loss)
    • At current speed/cost, the priority shifts toward quality and controllability
    • Speed improvements unlock entirely new interactive product experiences
  8. 12:26 – 13:04

    Controllability roadmap: references, LoRAs, lip-sync, and camera/style modules

    Gorkem outlines how adding references dramatically increases control and becomes the default way users work. They describe building an ecosystem of LoRAs and specialized fine-tunes—lip sync, camera angles, styles—made easier by open-source foundations.

    • References (beyond text/image-to-video) become the default for controllability
    • LoRA ecosystem planned/active: lip-sync, camera-angle, style variants
    • Open source accelerates community and internal experimentation
    • Focus is moving from base speed to layered control features
  9. 13:04 – 14:18

    Launch validation and the “Twitch moment” right after release

    They recount delaying messaging until external eval platforms confirmed results, since internal numbers seemed “too good to be true.” Immediately after launch, spontaneous real-time streaming experiments appeared—demonstrating a new category of interactive video use.

    • fal spent heavily on internal evals, then waited for external validation
    • The results were unexpectedly strong, prompting caution before launch
    • Real-time streaming emerged almost instantly post-launch (Twitch/online demos)
    • The product story quickly evolved from model release to new experiences
  10. 14:18 – 16:12

    Internal creativity burst: three parallel viral projects and fal Live via WebRTC

    Gorkem describes a company-wide sprint where multiple teams independently built applications and optimizations around H3 Max. These included WebRTC-based live experiences and several prototypes that all gained traction on social media.

    • Model release triggered an “all-hands creativity explosion” internally
    • Multiple parallel projects: apps, LoRAs, and live streaming experiences
    • WebRTC live experience developed rapidly
    • Distributed team iterated around the clock over several days
  11. 16:12 – 25:26

    From stitched clips to true continuity: two-minute memory and H3 Max Director

    They explain the limitation of naive continuity (feeding the last frame into the next clip) and how it lacks real memory. The Director variant extends attention and continuity so the model can maintain scene coherence and react to new instructions in-stream.

    • Stitched-clip approach: last-frame carryover, but no real memory across clips
    • Director approach: extend chunk length and retain attention to prior video
    • Memory can reach ~2 minutes while maintaining real-time generation
    • H3 Max Director enables long continuous, action-controlled video sessions
  12. 25:26 – 26:45

    How the model “remembers”: compressed raw-video attention + evolving system prompt

    Batuhan details the memory mechanism: the model retains a compressed representation of recent raw video, plus an evolving higher-level prompt-like state for longer horizons. This combination supports coherence across multiple scenes and enables “infinite” streaming behaviors.

    • Remembers recent past via compressed raw-video representations (not full frames)
    • Two-minute window retains detailed scene continuity (last 4–8 scenes)
    • Beyond that: an evolving system prompt tracks higher-level world state
    • Enables coherent, long-running interactive narratives and social experiences
  13. 26:45 – 28:31

    Economics and infrastructure: serving costs, chip footprint, and splitting workloads across hardware

    The discussion turns to how complex video pipelines can benefit from heterogeneous deployment—running different stages on different hardware for efficiency. They also discuss partial on-device execution paired with cloud control planes for creators and streamers.

    • Video inference pipelines have distinct components with different hardware sweet spots
    • Future efficiency: run pipeline stages on different hardware configurations
    • Possibility of adapting optimizations for consumer/edge hardware
    • Hybrid model: local streaming + cloud-based director/control plane
  14. 28:31 – 34:11

    Hollywood-grade workflows: Blender references, camera JSON control, and closing the studio gap

    They describe a pro workflow where Blender renders provide reference video for near-100% controllability, and H3 Max’s speed makes iteration practical. They then dive into structured camera control (JSON trajectories), plus a broader toolkit of post-training capabilities aimed at professional reliability.

    • Blender + AI reference video yields near-100% controllability for VFX workflows
    • LLMs can generate Blender scenes; H3 Max enables rapid iteration over variants
    • Camera control via structured JSON (time-indexed position/angle) with strong adherence
    • Post-training infra enables adding capabilities (camera, lighting, motion, lip-sync) to models
    • Goal shifts to 99.9% reliability for professional production use
  15. 34:11 – 38:56

    Hollywood adoption accelerates: point solutions, legal/data residency, and what comes next

    Gorkem argues studios want targeted tools (extend clips, adjust camera/lighting) rather than end-to-end “generate a movie” systems, and fal is positioned to build those point solutions. They also highlight legal/data residency progress and rising studio participation, signaling rapid growth in professional AI usage.

    • Hollywood is fal’s fastest-growing segment; usage was near-zero a year ago
    • Studios want point solutions (extend video, camera, lighting) more than full generation
    • fal aims to bridge the gap between research lab outputs and real studio needs
    • Progress on IP workflows and data residency; US-hosted options remove blockers
    • Shift in industry events: conferences now dominated by studios planning deeper integration

Get more out of YouTube videos.

High quality summaries for YouTube videos. Accurate transcripts to search & find moments. Powered by ChatGPT & Claude AI.