Ethan details building Grok Imagine from scratch in three months, emphasizing that fast iteration and fixing small data pipeline bugs drove more gains than new algorithms. He explains that video models are bootstrapped from image models and that their intelligence primarily comes from language models acting as prompt rewriters. The high cost of training video models is driven by massive storage and I/O needs, comparable to LLMs. He defines a world model as a real-time, interactive, long-horizon video, and sees video agents—using LLMs to call generative and traditional tools—as the next major trend. Ethan left xAI to focus on LLM research, predicting that language models will soon manage their own context, and that physical AI may be solved by powerful video models controlling robots as tools.