@Mkrn@jpschroeder Based on their introductory article, the model currently takes in text as input. Probably keeping track of the world state with textual descriptions of key items on the page? Idk, I'm also very curious. Same with the doomo on the typesafe website