After a brief hiatus, the paradigm that once produced superhuman video-game agents now stands poised to transform language models from exceptional retrieval systems into autonomous beings capable of purposeful action.
Yet today's agents remain confined to tasks that occupy only a few human minutes; they falter when confronted with true long-horizon challenges. The liberation of agents in such extended domains requires environments in which they can learn through lived experience. These environments will become an immensely valuable asset in the AI supply chain.
Rather than concentrating on scaling RL training itself, we are focused on scaling the creation of environments. is constructing the foundation model capable of generating worlds for any task.