That distinction became particularly relevant during presentations about world models: systems designed to understand or generate environments that can respond to users rather than remaining fixed images or videos.
World Labs researcher Yi Hua presented Atlas, a multimodal world model developed to operate across text, images, video, and 3D data within a shared spatial context. Runway Research Director Jimei Yang discussed Solaris, an early research project that generates an interactive interface frame by frame. Instead of interacting with traditionally authored code, users provide clicks, drags, and text input that condition what appears next.
Dr. Cao of Tripo introduced Project Eden, Tripo’s world model, and its latest milestone, EdenLive, a real-time generative renderer intended for rapidly prototyping interactive experiences. He argued that although agents are becoming more capable of producing code and assets, someone must still recognize an idea, explain it, evaluate the result, and decide whether it is worth developing.
“Everything started as an idea, but it may have been too much time or bother before,” Dr. Cao said. “Now, if we can describe it, we have tools that can turn a passing idea into a real thing.”
The projects presented during Tripothon offered more concrete examples of what these new workflows might look like.
Jennifer Jing received the San Francisco event’s Best Use of World Labs award for Agent Blindbox, a virtual collectible experience built around characters intentionally styled as “AI slop.” Players open a trash can to reveal different characters, each of which inhabits and interacts with its own environment.