endtail

we produce real environments and multimodal data for world models, starting with video games.

we’re building datasets that keep video, audio, player inputs, and changes in world state in sync. we have our own action-labeling models to label player actions in gameplay footage.

endtail focuses on less common interactions, failed attempts, and recovery. our goal is to build environments where we can return to the same starting point, try different actions, and record each outcome. this lets researchers compare what happens when the action changes.

our team has worked with frontier labs and co-authored nomos 1, a mathematics model, with nous research.

contact: i@endtail.com