← back
Generative Video at the Speed of Light — Keegan McCallum, uRun
Takeaway
Efficient video models make interactive visual applications feasible, but serving and streaming infrastructure determine how usable they become.
Summary
- uRun highlights video-model progress in efficiency and continuous generation, using Helios, a distillation of Wan 2.1 14B, as an example.
- Real-time generation enables interactive visual transformations, accessible visual interfaces, avatars, and live steering of creative shots.
- Serving these applications requires geographically distributed GPUs, WebRTC connectivity, synchronized controls, and asynchronous pipelines that combine multiple models.
- uRun proposes a drop-in React component backed by a programmable Python runtime, plus CLI and MCP interfaces for agents building interactive media applications.
real-time-videovideo-inferenceurun
Original description
Ten dollars now buys roughly three hours of continuously generated video, and fifty buys fifteen. Keegan McCallum sets that against the room's own habits, since plenty of hands went up for burning that much on coding tokens inside a single hour. His argument is that the interesting axis in generative video stopped being quality a while ago. Put a real time generation next to one that took minutes and the slower clip still has better motion, but it cost on the order of a hundred times more to produce. Helios, the model he serves, is a distillation of a 14 billion parameter open model, and it is one of at least forty released this year carrying real time or long horizon capability. What that unlocks has less to do with better clips than with a different interaction shape. A webcam that shows you the haircut you are considering. A visual medium for people who do not think in text, which is most of what working with AI currently demands. Content creation that stops being a slot machine where you spend ten dollars a minute on a prompt and some keyframes and hope for the shot. Steering a generation in under a second is a different job entirely. What is left is the serving problem: GPUs positioned globally, WebRTC with ICE and TURN, and several models wired into one continuous streaming pipeline that stays synchronized with user controls frame by frame. Speaker info: https://x.com/keeganmccallum3 / keeganmccallum3 https://urun.sh Timestamps: 0:00 - Generative video along the quality axis 1:24 - The other axis: efficiency and long horizons 3:23 - What ten dollars of generation buys now 4:38 - Magic mirrors, accessibility, and steering shots live 6:33 - The hard part is serving it