Keegan McCallum shifts attention from the visual quality of generative video to efficiency and continuous generation. Using narrated Helios examples, he describes models producing video quickly enough for ongoing interaction rather than requiring users to wait for each finished clip. His comparisons illustrate the talk's argument but do not establish a universal quality or cost advantage.
Keegan McCallum proposes real-time webcam transformations, visual interfaces for people who find text-heavy interaction difficult and creative tools that let users steer a shot while it is being generated. He frames these as opportunities enabled by lower latency, including finer camera control and longer-running world-model experiences.
Delivering that interaction requires more than a model endpoint. Keegan McCallum discusses GPU placement, routing users to compute, WebRTC connectivity, ICE and TURN, and pipelines that connect multiple models while keeping each frame synchronized with user controls. Smooth continuous streaming is presented as a central engineering challenge.
Keegan McCallum describes uRun's proposed combination of a React interface component and a programmable Python runtime for asynchronous media pipelines, alongside CLI and MCP access for agents building applications. He concludes that serving and integrating the available models is a major frontier for interactive AI, while leaving application outcomes and wider accessibility benefits to be demonstrated.
Watch on YouTube
