Ethan He details the process of building frontier image and video systems, from synthetic data generation and VAE compression to the hidden costs of storage and I/O. He emphasizes that iteration speed and fixing small bugs in data pipelines often yield larger gains than new algorithms. A central thesis is that video models derive their intelligence primarily from LLMs, with prompt rewriting and agentic orchestration driving quality improvements. He predicts the next frontier is not a better video model but a video agent—a system that can plan, generate, edit, and iterate across a creative task using language models as the core reasoning engine. He also explores the future of generative UI, where diffusion models could replace traditional frontends, and defines world models as real-time, interactive, long-horizon video systems. Ethan left xAI to focus on LLM research, believing the bottleneck for video models is now the language and agent component.