a16zHow Real-Time AI Video Is Changing How Creators Work
a16z
Introducing H3 Max 0:00
The conversation opens with Fal introducing H3 Max, a post-trained version of the open-weight Minimax H3 video model. The hosts describe it as being in a league of its own on benchmarks, so far ahead that the results had to be plotted on a log scale just to fit on the graph. A turbo version of the model can generate a 5-second video in about 1.5 seconds, at roughly half the cost of the original release.
Why post-train an open model 2:31
Fal explains why they chose to post-train Minimax's H3 model rather than just serve it. H3 was the first truly next-generation video model that was fully open source, capable, and built on a familiar architecture, which made it a rare opportunity. Generative media, alongside coding agents, is described as reaching what they call token market fit, meaning a single person can productively spend around 10,000 dollars a month on tokens. Since around April, the whole industry has been constrained by compute, so any efficiency gain frees up capacity for more generation elsewhere.
Stacking optimizations for speed 5:00
The speed gains come from layering multiple techniques rather than one trick. First, the model is post-trained so it can run in fewer diffusion steps, going from around 50 steps down to about 20, while a quality-focused pass ensures none of that improvement is lost. On top of that, systems and kernel engineering push hardware utilization from a typical 30 to 40 percent up toward the practical ceiling of 70 to 80 percent. Every stage of the pipeline, from expanding the prompt with a large language model, to generating the video in latent space, to decoding it back into pixels and upscaling, is separately optimized, and the combined effect compounds into roughly a 35x speed-up over the original Minimax model without sacrificing quality as measured by ELO score.
Hardware and serving setup 10:33
Moving from older Hopper chips to newer Blackwell chips brings roughly a 2 to 3x improvement, mainly cutting wall-clock time rather than cost, since the newer hardware also costs more. Most video models today, including this one, run across eight GPUs in a single node, because scaling beyond that introduces communication overhead that reduces efficiency. A week after releasing H3 Max, the team found they could run it twice as fast at about 97 percent of the original quality, and released that as H3 Max Turbo, which can produce a 5-second video in about 1.5 seconds at half the cost.
An internal burst of creativity 14:31
The model's release triggered a company-wide surge of spontaneous projects. One engineer, Rahan, started livestreaming continuous H3 Max generations from his own laptop on Twitch, prompting the model to keep a coherent story going. Around the same time, Twitter figure Levels.io built his own website for infinite streaming, and an internal team was separately building a more seamless continuous version with about two minutes of memory, where characters entering a scene are reflected coherently. All three efforts went viral independently within days, without any coordinated launch plan.
H3 Max Director and its memory 20:30
That continuous-memory work became H3 Max Director, a public model able to generate up to 60 minutes of continuous, action-controlled video, where prompts like a person entering a room are reflected almost immediately while earlier elements of the scene stay consistent. The team built demo channels, including one full of chaos and one styled like 1980s cartoons, and let viewers vote on the next action in the story. Since release, H3 Max has become the platform's most popular video model by more than double the volume of any other. The model remembers roughly the last two minutes of raw video in detail, and beyond that keeps an evolving system-style summary of scene structure covering the last four to eight scenes, allowing it to hold onto the overall coherence of a world over much longer stretches.
Splitting the Pipeline Across Hardware 27:01
Video models are as complex as large language models because their pipelines have many components, and different parts often need different hardware to run efficiently. To push costs down further, the plan is to split the pipeline across different hardware types, including letting parts of it run on consumer machines at home, with a director and control layer still living in the cloud.
Blender Plus AI for Full Control 29:00
A popular workflow among professional visual effects artists is rendering a rough, low resolution scene in Blender and feeding it into an AI video model as a reference, which gives close to full control over the output. After the H3 Max model launched, people began using GPT Astra to generate scenes inside Blender and then passing them to H3 Max, which works especially well because it is fast enough to try many variations in parallel, opening a new pipeline for Hollywood and professional creators.
Pushing Toward Full Controllability 31:00
With a strong base model in place, the next month or two will focus on controllability rather than raw quality, including a lip synchronization model that takes audio and a video or image reference, and motion controls that let a dance move be applied precisely to an AI generated character. The goal is 99.9 percent reliability instead of the 80 to 90 percent that basic prompting gives. A recent camera control launch lets users supply a structured, JSON style description of exact camera position and angle at each point in time, which the model treats as the sole source of truth, and this required significant tuning of the base model itself.
Building Infrastructure for Any Model 33:02
Rather than doing one-off training runs, the team is building infrastructure that can add capabilities like camera control, lighting control, voice assignment, and character reference to any video model, open source or closed source. This addresses what is described as the biggest gap in the market, the lack of controllability, which has evolved from text to video, to image to video, to reference to video, with each new capability compounding on the last.
Hollywood Becomes the Fastest Growing Segment 34:00
Hollywood usage was nearly nonexistent a year ago and is now the fastest growing segment, with studios like Amazon MGM releasing tools backed by this infrastructure. Studios want small, targeted solutions such as extending a video slightly, adjusting camera angles, or changing lighting, rather than generating entire scenes from scratch, and building these post training solutions is seen as closing the gap between what research labs build and what professionals actually need.
Solving Legal and Licensing Barriers 37:00
Beyond model capability, a major barrier has been legal and data residency issues, and progress has been made there too, including a system allowing studios to bring their own intellectual property into the models. The Seance model is now hosted in the United States, removing what was described as the last missing piece Hollywood studios wanted, and usage is expected to grow ten to a hundred times in coming months.
A Changed Conference Audience 37:31
At the upcoming Generative Media Conference, the second one held, the audience has shifted from mostly curious consumer AI attendees and a few Hollywood executives last year to now being dominated by studios, including new AI-focused offshoots of major studios and the biggest Hollywood studios themselves, reflecting their plans to integrate AI into existing workflows.
AI-generated summary. It can be wrong or incomplete - check anything that matters against the original.
