Playground
small games driven by tiny neural nets
Small games. Everything that moves is pushed around by neural nets with a few thousand parameters each, trained in plain JavaScript (code). No physics engine, no game AI.
World model
Drive the ball with WASD or the arrows. A small net predicts all the physics. “Show ground truth” overlays the real simulation so you can see where the model drifts.
loading…
Arena
Push the crates into the ring. The collisions are learned, not simulated. Autopilot plays it with a policy trained inside the model.
loading…
Colony
The paragraph is alive. Letters hatch, eat other letters, turn into words, fly for a while, then put themselves back.
Every creature on this page is made of text. Each one hatches from a single letter of this paragraph, crawls around eating more letters to grow its body along the beginnings of real words, and dreams in a silk cocoon the moment it finally spells one. Then it flies: a word on wings, looping over the very sentences it was born from. When the flight ends it gives every borrowed letter back, the gaps heal, and an egg or two remains. Nothing here is ever truly lost, except whatever the masked one keeps stealing.
waking the colony…
Chase
The red one picks its moves by simulating futures with the same nets that run the game. The threads on the floor are the futures it considered. Hold shift to see your own.
loading…