Watch a model memorise the noise
The dots come from a hidden curve plus random wobble. Drag the flexibility slider and watch the model start copying the wobble instead of the pattern.
Try this
- Drag flexibility from 1 to 4. At 1 the line is too stiff and misses the bends (underfitting). Around 3 or 4 it follows the shape. Tick "Show the hidden truth" to confirm.
- Drag flexibility to 15, then tick "Reveal new data". The model's line wiggles through every training dot and the training error is nearly zero. But look at the error on new dots. It jumped. The model memorised the wobble (overfitting).
- Set noise to 0 and keep flexibility at 15. Now the wiggly model is harmless. Then set noise to 1.0: even flexibility 4 struggles. Noise is what makes "trying too hard" dangerous.
This is part of Session 1 of How AI Learns. The full session adds the story, the Python build and a boss challenge.
Start free →