Sep 8, 2025 · 1h 4m · latent-space
A Technical History of Generative Media
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Fal leadership Gorkem Yurtseven and Batuhan Taskaya discuss the evolution of their generative media platform into a $100M+ ARR business with podcast hosts Alessio and Swix. They explore low-level GPU kernel optimizations, the progression from Stable Diffusion to frontier video and world models, and enterprise commercialization across digital advertising.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. How this is scored →
The hosts as informed peer 5.1 Guest teaching 4.8 Guest disagreement 2.0 The hosts pushing back 2.7
speaking balance: gold is the hosts, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 25:01 Batuhan dismisses custom ASICs as uneconomical
Batuhan strongly rejects the premise that specialized ASICs make sense for diffusion models, arguing that GPU overhead is minimal and matrix multiplication hardware remains superior.
Hardest push from the hosts ▶ 1:02:40 Alessio challenges Fal's standard hiring job boardsAlessio pushes back against Fal's standard careers page, advocating that they follow TinyGrad's approach by publishing hard kernel-writing bounties to filter candidates directly.
Biggest teaching moment ▶ 15:06 Batuhan reframes universal 10x inference speed claimsBatuhan educates the hosts on hardware realities, explaining that open-source PyTorch optimizes rapidly on established chips, so the true technical challenge is extracting flops on unoptimized newer architectures like Blackwell.
The host holds their own ▶ 33:37 Alessio cites empirical counterexample against world model physicsAlessio demonstrates deep familiarity with research literature by pointing out that while video models predict planetary orbits visually, they fail completely when tasked with modeling underlying gravitational force vectors.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The hosts as informed peer | Guest teaching | Guest disagreement | The hosts pushing back | Why |
|---|---|---|---|---|---|---|
| Introductions and Origins of the Fal Platform | 4 | 3 | 1 | 2 | The hosts set up the conversation with shared historical context around Fal's early pivot from Python runtimes. The guests explain their developer scale, catalog size, and evaluation methodology in a friendly introductory dynamic. | |
| Chronological Milestones in Generative Media Evolution | 4 | 5 | 1 | 1 | Alessio prompts a timeline of generative media model spikes, prompting Batuhan to deliver a comprehensive monologue tracing Stable Diffusion 1.5 through SDXL, Flux, and Google DeepMind's Veo 3. | |
| Strategic Pivot: Specializing in Diffusion over LLMs | 5 | 5 | 1 | 1 | Swix explores Fal's strategic decision to avoid competing directly with LLM inference providers against Google and OpenAI. Gorkem and Batuhan outline how poor baseline GPU convolution performance opened a greenfield opportunity in diffusion. | |
| Custom Kernels, Inference Architecture, and Latency Moats | 6 | 6 | 3 | 4 | Alessio presses on kernel reuse and asks whether Fal universally offers a 10x speedup. Batuhan clarifies that open-source frameworks catch up quickly, framing their true moat as rapid optimization across new chip architectures like Blackwell. | |
| Proprietary Lab Partnerships and Serverless GPU Infrastructure | 5 | 5 | 2 | 3 | Swix asks whether proprietary model labs send unreleased weights and questions serverless GPU mechanics. Batuhan and Gorkem describe forward-deployed kernel engineering partnerships and their custom multi-cloud orchestration stack. | |
| Next-Gen Hardware: Blackwell, ASICs, and Architecture Trends | 5 | 6 | 4 | 2 | Swix poses the prospect of custom ASICs for diffusion workloads. Batuhan firmly dismisses ASICs due to memory bandwidth realities, high GPU utilization for matrix multiplications, and rapid architectural churn among AI researchers. | |
| Real-Time Generation, Consistency Models, and Frontier Video Models | 5 | 4 | 2 | 2 | Swix asks why consistency models and fast sketch-to-image workflows faded in popularity. Gorkem and Batuhan attribute this to base model quality preferences and explain how generation velocity impacts creator workflows in video. | |
| World Models, Robotics Simulation, and the Video Revenue Surge | 6 | 5 | 2 | 4 | Alessio challenges the assumption that world models truly grasp physical laws by citing planet orbit simulation anomalies. Batuhan responds with a scaling hypothesis, and the group analyzes video revenue share and Chinese open-source lab releases. | |
| Enterprise Commercialization: Advertising, LoRAs, and Workflow Pipelines | 5 | 5 | 2 | 3 | The conversation covers advertising demand, LoRA fine-tuning latency, and ComfyUI workflow pipelines. Alessio questions whether single frontier models replace multi-step pipelines, while Batuhan highlights enterprise demand for pixel-accurate brand consistency. | |
| Requests for Startups, RL on Media, and Engineering Culture | 6 | 4 | 2 | 5 | Swix and Alessio discuss startup opportunities, RL for media, and recruitment. Alessio directly challenges Fal's conventional job listings, urging them to adopt George Hotz-style kernel bounties to screen out vibe-coding applicants. |
Statements from this episode (21)
Yurtseven: Fal.ai has around two million developers on its platform
“Yeah, we have around two million developers on the platform.”
Yurtseven confirms Fal.ai has surpassed $100M in revenue
“And you guys are over a hundred million in revenue, right? Just, this is not, you know, just developers kind of kicking the tires. That's correct. Yeah.”
Taskaya: SDXL generated Fal's first $1 million in revenue
“And then SDXL came, which was like the first major model. That brought, like, our first million in revenue if we consider that”
Taskaya: Flux was the first enterprise-ready generative image model
“The team at Stability left to start Black Forest Labs, which released Flux models. And that was the first model to, you know, breach the barrier of commercially usable, you know, enterprise-ready grade models”
Taskaya: Flux grew Fal's revenue from $2M to $20M in two months
“In the first month of flux models, we reached from, like, two million to ten million in revenue. That was, like, a big jump. Next month, we were at 20, like, it just started going from there”
Gorkem: Hosting LLMs is a bad business due to Google search competition
“Language models, hosting language models is not a good business. At the time we thought, okay, we are going to be competing against OpenAI and Anthropic and all these labs. Turned, turned out that it was even worse because the killer application of language mo…”
Yurtseven: Long-term inference differentiation relies on constant hardware and architecture churn
“I think it's very hard to create this differentiation over long term if there is no new architectures, if there's no new chips, but luckily there is all the time.”
Yurtseven: Customer A/B testing proved higher image latency directly reduces engagement
“One of our customers actually did a very extensive A-B test of, like, they purposely slowed down latency on file to see how it impacts, you know, their metrics, and it had a huge part in it, and it's almost like page load time.”
Taskaya: Fal optimizes inference for four major AI video companies
“We have like four different companies four, four major video companies that we are doing this with and one image company that I don't think we disclosed.”
Taskaya: Custom ASICs do not make sense given low Nvidia GEMM overhead
“What is the overhead of an NVIDIA GAM instruction, right? It's like 16%. So like you're essentially buying a, Matrix multiplication machine. So, like, it doesn't really make sense to specialize it that much.”
Taskaya: Model architectures will keep churning due to researcher novelty bias
“I think the architecture is going to keep changing until this, like, paradigm of, like, you know, researchers changing stuff for the sake of changing, you know, finishes.”
Yurtseven: Real-time generative drawing tools fail to retain users long-term
“I think it makes for a good demo. You know, you could build real-time applications. You could build these drawing applications, things like that. But I don't think people could build applications that like have user retention long-term. Like people couldn't re…”
Fal CEO: Newer Video Models Are Now Much Better Than OpenAI's Sora
“Now we have video models that are much better than Sora.”
Taskaya: 1,000x More Compute and Data Will Yield Accurate Physics Simulators
“It's just a matter of like data scale and like the underlying fundamental architectures. But like, I don't think it's going to change that much. We're just like, we're going to put thousand X more data, thousand X more compute, and we'll get like the best phys…”
Yurtseven: Video models now account for over 50% of Fal's revenue
“That was February, so now, now it's probably over 80. No, 50%. 50? Yeah, okay. It's like over 50. Yeah, yeah, a hundred percent.”
Taskaya: Releasing Video Models Yields Better ROI Than Subpar LLMs
“My guess is, like, training these costs, like, a couple million dollars, which is not that much, especially, you know, like, you know, they're probably backed by some sort of entity... So training these models will bring you a lot of attention and it's more at…”
Taskaya: NSFW content makes up less than 1% of Fal traffic
“Moderation is optional to a level where, like, illegal content is moderated, and we also track, like, the non-illegal content NSFW moderation, and, like, we haven't seen, like, we haven't seen more than one percent.”
Yurtseven: Generative AI will not revolutionize Hollywood due to finite attention
“Like how many movies do you watch a year? Like maybe 20, 25 movies. How many movies in the theater you watch? Three, four at most. So if there are like thousands of movies that a year, like people won't be able to watch all of these movies. Like there's just n…”
Taskaya: 80% of promotional video content will be AI-generated within 12 months
“Like 12 months, I think like 80% of this is going to be generated.”
Yurtseven: Closed-source models cannot foster effective LoRA ecosystems
“Like I've never seen a closed source model that can create a good LoRa ecosystem. It just basically doesn't exist.”
Taskaya: Training a state-of-the-art image model costs under $1M
“Like right now, like if you look, if you want to train a Sota image model, I don't think it's going to cost more than a million dollars. It's extremely cheap. It's like a matter of data engineering effort, cleaning. It's, I think it's a function of data set.”