Show HN: A 14M-param world model of 2D Minecraft you can play on your phone

Category: ai-ml

Tags: world-models, reinforcement-learning, generative-models, benchmark, game-ai

Score: 7.3/10 (Innovation: 7, Technical: 8, Documentation: 9, Utility: 5)

A 14M-parameter world model of Craftax (2D Minecraft) trained from scratch to predict next frames from pixels and actions, playable in-browser via WebGPU. It doubles as an autoresearch benchmark with a frozen evaluator, scoring rules, and a leaderboard, enabling fast iteration on world-modeling techniques under tight compute budgets.

Target audience: AI/ML researchers and engineers interested in world models, generative models, and automated research benchmarks

Repository: https://maharishiva.github.io/tinywmcraft/ · Python · MIT · 3 stars

View on Hacker News