You ARE the model.
A roguelike where you pilot a model-ship down the loss landscape to convergence before gradient explosion, catastrophic forgetting, or a hallucination storm takes you out.
Built as a riff on the training process itself: every TRAIN is one epoch, your learning rate is the gas pedal, and your crew are the model’s capabilities: attention, reasoning, recall, calibration. Random events are real ML failure modes, and surviving them means assigning the right capability at the right time.
The glossary doubles as a cheat-sheet for the actual concepts, so dying to reward hacking literally teaches you what reward hacking is. Convergence is the win condition; the abyss is one careless learning-rate bump away.
keys: SPACE to train · [ / ] to tune learning rate
vanilla JScanvasstatic