A tiny AI trained for about ten cents imagines a playable Minecraft-like world in your browser
The project doubles as a benchmark: its recipe book lets players spot learned rules going wrong, while a public leaderboard measures more than how convincing the pictures look.
The recipe book says a diamond needs the right pickaxe. The little world on the same page has to remember that from images. If it forgets, the little world keeps generating pictures anyway, leaving the player to discover where its imagined inventory and the real recipe have parted ways.
tinywmcraft puts a Craftax-Classic world model in the browser. Craftax-Classic is the project’s two-dimensional Minecraft-like setting, with gathering, crafting, creatures and survival. During play, the model receives earlier frames and an action, then predicts the next picture, reward and whether the episode ends.
“No game engine and no game state,” the maker writes about what runs on each action. The encoder compresses earlier images, a world model predicts the next latent representation, and a decoder turns it back into the screen you see. Feeding those imagined images back into the model lets you keep walking through its predictions.
That makes it different from a game displaying a machine-generated texture. The simulation you’re pushing against is itself learned. A plausible image can conceal a broken recipe, a forgotten object or an inconsistent inventory. The page puts the real rules nearby, helpfully giving the hallucination something specific to disagree with.
The maker reports 13.9 million total parameters, divided between a tokenizer and the world model, trained from scratch in under ten minutes on one A100; the approximately $0.10 figure uses the page’s own compute pricing estimate for a training run. The project’s development cost did not stop at a dime dropped into a slot.
Its account describes 698 runs in eight days during benchmark development. Those runs used development data and an earlier evaluation, ranked by a visual metric. The current submission score covers appearance, game logic, memory, rewards and deaths. A cheap individual experiment and a cheap completed research project are wonderfully different items on a receipt.
The recipe book gives the comparison some teeth. Wood starts the crafting chain; pickaxe upgrades unlock later materials, with an iron pickaxe needed for diamonds. Those steps make a practical memory check. The model has to maintain a useful relationship between previous actions and what the player can do next, rather than merely drawing another charming patch of ground.
The page says each world begins with an initial frame imagined from noise, rather than replaying a recorded world. It also offers a public benchmark and leaderboard so alternative training attempts can replace the model in the player when they beat the score under the rules.
A small, playable benchmark could make experiments easier to inspect without turning every iteration into an enormous infrastructure project. The unresolved challenge is continuity: a world has to keep its promises after the first attractive frame. Watching what happens after several crafting steps may be more revealing than watching the grass render.
The recipe book is right there.