What's this about?
An action-conditioned video world model that renders an endless, interactive world in real time on a single desktop GPU, trained on AAA games and simulation data.
What exactly is a "world model"?
A model that predicts what happens next given an action — press "forward" and it generates the next video frames of walking forward. A learned game engine, useful both for play and for training robots in imagined environments.
Why is it worth noting?
Top HF paper of the day (169↑). "Single desktop GPU" is the headline — world models are leaving the datacenter.