A playable explainer
Gradient Descent: the loss landscape game
Every AI model you've ever used was trained by one idea: measure how wrong the model is, then nudge it a small step in whichever direction reduces the error — millions of times over. This game puts that idea under your fingers. You are the optimizer. Six short levels, each one a real training problem: local minima, vanishing gradients, learning rates that explode, noisy data.
The game should have opened by itself. If it didn't — or you closed it and want back in:
Runs entirely in your browser (WebGL2, no data leaves your device). Works with mouse, keyboard, or touch. Five to fifteen minutes, English only.
Why a game?
Because "the model rolls downhill on an error landscape" is the single most useful mental model in all of AI — and reading about it doesn't stick the way getting trapped in a local minimum does. Ten minutes in here and phrases like learning rate, divergence, and noisy gradients stop being jargon and start being things that happened to you.
If it clicks and you want the bigger picture, start with the AI Fluency Path or jump straight into the guides.