Training Data The Rise of Generative Media: fal's Bet on Video, Infrastructure, and Speed
45 snips
Dec 10, 2025 
In this conversation, Gorkem Yurtseven, co-founder of Fal, and Batuhan Taskaya, Head of Engineering, share insight into the rapidly evolving world of generative media. They delve into the computational challenges of video models compared to LLMs and discuss the performance enhancements of Fal's tracing compiler and custom kernels. The team highlights the booming demand from AI-native studios and the future of generative video in educational contexts. They also explore how rapid iteration in video model design is shaping the landscape, paving the way for new creative possibilities.
AI Snips
Chapters
Transcript
Episode notes
Generative Media Was Underestimated
- Generative image and video were initially overlooked but matured rapidly into large markets.
- Gorkem Yurtseven says better models and use cases flipped perception from "toy" to "massive."
Tracing Compiler Unlocks Kernel-Level Speed
- Fal built a tracing compiler that finds runtime patterns and replaces them with templated specialized kernels.
- Batuhan Taskaya credits this approach for leading performance across benchmarks while keeping output quality.
Video Workloads Are Compute-Bound
- Video diffusion workloads are compute-bound, unlike many LLMs which are memory-bound.
- Batuhan Taskaya explains video denoising and attention saturate GPU compute rather than memory bandwidth.