Moonlake: Causal World Models should be Multimodal, Interactive, and Efficient — with Chris Manning and Fan-yun Sun

Moonlake AI, co-founded by Stanford's Chris Manning and Fan-Yun Sun, is building causal world models that are multiplayer, interactive, and indefinitely persistent — contrasting sharply with Genie 3's single-player, 60-second limit. Their approach bootstraps from game engines and trains custom agents to achieve efficiency through structure and causality rather than blind scaling. Ian Goodfellow is also involved, arguing current SOTA world models still fall far short on interactivity and efficiency.