01 Dataset
An autopilot bot plays Snake. It heads for the food and avoids trapping itself. Its games are recorded as the training data.
0 transitions
02 World Model Training
Reconstruction loss (L1)
awaiting gradients…
Neuron Corruption
03 Play the Dream
Or drop or paste an image onto the game. The net gets it as a 24×24 image and makes its best guess about what comes next.
Dreamed frames 0 · Longest dream 0 · Epochs 0
DIR
Each tick: (last frame, your key) → CNN → next frame, which gets fed back in. There's no game logic, and that includes food: the network places new food itself, 3 cells to the snake's side, a rule it picked up from the training data. If the snake goes through a wall, that's the network's idea.
Frame predictor: a 5-layer convolutional network (32→64 dilated→64→48→3 filters). Input is a 24×24 RGB frame plus 4 action layers. It's trained with an L1 loss and gives one fixed guess per input, with no memory and no randomness.