World Labs co-founder Justin Johnson and Sam Charrington unpack several meanings of world modelsA world model learns how an environment is structured and how it may change in response to actions., including implicit knowledge, models of state transitions and systems that generate navigable environments.
Johnson explains how observations, actions and state relate in partially observable decision processes, then compares explicit 3D representations such as Gaussian splatsGaussian splatting represents and renders a scene using collections of spatial Gaussian elements with properties such as position, shape, color and opacity. with models that generate frames directlyA video generation model creates or transforms moving images from text, images, video, audio, or structured controls..
The discussion covers World Labs’ Marble and real-time frame research, consistency, data requirements and the open problem of combining planning, simulation and rendering. Johnson presents unified models as a research direction rather than an achieved result.
Watch on YouTube
