a16z Podcast

Gorkem Yurtseven and Batuhan Taskaya on generating video faster than it plays

Gorkem Yurtseven & Batuhan Taskaya· Co-founder and Head of Engineering at fal
·~39 min·English·a16z
MultimodalInferenceGPUAI Infrastructure
TL;DR

fal post-trained MiniMax's open-source H3 video model into H3 Max, combining model-side post-training with kernel and systems work to generate a five-second clip in about 1.5 seconds. Generating video faster than it plays turns it from a batch render you wait on into a live medium you can direct.

01Core Mental Model

Token Market Fit

Generative media, like coding agents, has 'token market fit': a single person can productively spend around ten thousand dollars a month on tokens, so demand runs far ahead of compute and every efficiency gain is the whole game.

generative media is is I would say along with the coding agent market uh what we call is is token market fit and the way we define it is as can a single person productively spend a lot of tokens and and the the amount is like 10k a month something like that.

Gorkem Yurtseven & Batuhan Taskaya, a16z Podcast
Key Insight
Framing video as 'token market fit' reframes fal's business: growth has been compute-constrained since April, so how many tokens a chip can emit per second sets the pace, which is why a serving company went into post-training at all.

02The Result

Five Seconds in 1.5 Seconds

fal's H3 Max matches or beats the base model's quality while running far faster; its Turbo variant trades a small, noticeable quality drop to generate a five-second clip in about 1.5 seconds.

We have a version called HDMax Turbo that's public that can generate like a 5second video in like 1.5 seconds.

Gorkem Yurtseven & Batuhan Taskaya, a16z Podcast
Key Insight
H3 Max beats the base model on both speed and quality, so the main win did not come from trading quality for speed: post-training raised the ceiling first. The separate Turbo variant is where fal then trades a little quality back for roughly another 2x.

03How: Two Levers

Two Levers Past the Roofline

The speedup compounds two independent levers: post-training the model to run in about 20 diffusion steps instead of 50, and kernel work that lifts chip utilization from 30-40% up toward 70-80%.

This is just essentially how do you go from like running 50 steps to running something like 20 steps, right?

Gorkem Yurtseven & Batuhan Taskaya, a16z Podcast
Key Insight
The two levers are qualitatively different: fewer steps lowers how much compute a video costs, while higher utilization gets closer to the compute the chip can already deliver. Only the second is capped by hardware, which is why post-training carried most of the gain.

04Systems

Not an Agent Decode

A video model is not one call but a pipeline (prompt-expansion LLM, diffusion, VAE decode, upscale), each stage optimized separately; even its LLM stage runs single-shot at low batch with no cache reuse, the opposite of an LLM agent decode.

This is not like something like an agent decoding LM workload where you have very high cache rates where you have higher sessions. It's a single shot. You take you give a prompt you get a prompt back and there's no caching. you're operating at low batch sizes.

Gorkem Yurtseven & Batuhan Taskaya, a16z Podcast
Key Insight
Because the prompt LLM runs single-shot, at small batch with no cache to reuse, it needs a different set of optimizations than a cached, batched agent decode. And the diffusion model and the VAE each need their own, so fal tunes every stage of the pipeline separately rather than reusing one recipe.

05Memory

Two Minutes of Memory, an Hour of Coherence

H3 Max Director keeps a compressed memory of the last two minutes of generated video, then carries an evolving system prompt from two to sixty minutes to hold a scene together while a viewer steers it live.

remembers like the last few scenes. If you think a scene is like 15 30 seconds, then it remembers like the last four to eight scenes and then on top of that there's like a continuously evolving gradually evolving system prompt

Gorkem Yurtseven & Batuhan Taskaya, a16z Podcast
Key Insight
The two-tier memory bets that recent frames still need pixels but distant context only needs a summary, the same instinct behind caching: keep what is recent, compress the rest. It is what lets an action-controlled stream run for an hour without paying the exploding cost of attending to every past frame.

06How the Field Moves

The Burst, Not the Line

Unlike an LLM's steady line of gains, generative media stays quiet for months and then bursts all at once when the base model, latency, and controllability finally line up.

This happens at fal once in every couple of months where like the whole company gets gets hold of something and and the creativity just explodes and everyone is just working on a new little app or or a different optimization Laura whatever it might

Gorkem Yurtseven & Batuhan Taskaya, a16z Podcast
Key Insight
The burst is not planned: the guests stress nobody designed these launches. But fal keeps separate research, inference, and ML teams, so when a capability lands several of them can grab it at once and turn one jump into many products before the quiet returns.

07What's Next

From Prompt to Director's Chair

The next frontier is control, not speed: references, camera angles supplied as JSON, lighting, and LoRA fine-tunes, all aimed at making the model do exactly what a director asks.

you just essentially like underneath you give a JSON of like I want camera at like 0000 at T0. I want camera at like 90 degrees angle at T1.

Gorkem Yurtseven & Batuhan Taskaya, a16z Podcast
Key Insight
Supplying camera position as structured JSON, not a prompt, signals where video tooling is heading: away from coaxing a model with words and toward controls a professional can rely on, with fal targeting 99.9% reliability so the output can be trusted to hit the mark.

08The Market

Hollywood Wants Point Solutions

Hollywood is fal's fastest-growing segment, and what it wants is small point solutions (extend a clip, change the camera, fix the lighting) rather than generating whole scenes from scratch.

these like small point solutions rather than generating everything from scratch. They want to be able to extend the video a little bit.

Gorkem Yurtseven & Batuhan Taskaya, a16z Podcast
Key Insight
The point-solution demand is why fal invests in reusable post-training infrastructure: one toolkit it can apply to many models to add exactly the control a studio asks for. It starts from open models it can grab the weights of, and says the same toolkit will extend to frontier closed-source models too.