Mystery

Neural Lab

Build a small neural network and watch it learn, live on your phone. Every number comes from Mystery's own engine.

Pattern to learn Hidden layersneurons per layer, e.g. 8, 8 Activation Optimiser Learning rate Noise

0epochs

—training loss

—training accuracy

—accuracy on unseen data

Filled marks: training data. Hollow marks: unseen test data. The background colour is the network's guess everywhere.

trainingunseen test data

Line thickness shows each weight's size: blue = positive, red = negative.

What is happening, step by step

Forward pass. Each point goes in as two numbers. Every layer multiplies by a matrix of weights, adds a bias and applies the activation.
Loss. The output is compared with the right answer. The loss measures how wrong the network is on average.
Backpropagation. The chain rule gives the slope of the loss for every weight. Mystery checks these slopes against numerical ones.
Update. The optimiser nudges every weight a little downhill. Repeat, and the loss falls.
Unseen data. Test accuracy shows whether it really learned the pattern, or just memorised the training points (overfitting).

Code Doctor

Check one of your apps for real bugs, fix them, and download a fixed zip ready for Cloudflare.

Choose your app's .zip or drop it here. The same zip you drag into Cloudflare Pages.

Your code is checked on this phone. Only when you press "Fix it" are the few lines needed sent to the AI, and secret keys are removed first.