LeMario: Training a JEPA World Model on Super Mario Bros
SMRTR summary
A developer built LeMario, a world model trained on Super Mario Bros using LeCun's JEPA architecture, to predict future game states from pixels and button inputs. The model learned short-horizon dynamics well, beating baselines by 45.5% over five steps, but failed at actual gameplay because visually similar locations fooled the planner and vertical movement was poorly represented.
SMRTR provides this summary for quick context. The original article belongs to Hacker News.
Read the original article