Pantograph pretrains on internet video to build a goal-conditioned Minecraft agent

A General Goal-Conditioned Minecraft Model

Pantograph released Pan, a 4B-parameter Minecraft model that fights mobs, explores, and builds structures. It pretrains on internet-scale video using hindsight relabeling—later frames become the goal for earlier frames, so no hand-labeled rewards are needed. At inference, you give it a goal image and it acts, even in unseen environments. The post doesn't disclose training data volume, hardware, or wall-clock time.

Why it matters: Clean method: hindsight relabeling from unlabeled video to learn goal-directed behavior, 4B params achieving out-of-distribution goals in Minecraft. But the post doesn't disclose training data scale, success rates, or generalization boundaries — reads more like a research teas...