Interactive Lab · Why bigger models aren't smarter
A dozen noisy dots hide a smooth true pattern. Give a model just enough flexibility and it finds the shape. Give it too much and it threads a curve through every single dot, nailing the data it saw and getting everything between them wrong. Drag the complexity knob and watch the moment learning turns into memorizing.
🎬 Watch the 45-second explainer, then turn the knobs yourself below ↓
▶ The playground, drag complexity, watch the fit
Training error · data it fit
Test error · data it never saw
Overfitting is the central failure mode of all machine learning, the reason train/test splits, regularization, dropout, and early stopping exist. The U-shaped test curve you just traced is redrawn on a whiteboard in virtually every ML design review: it's the picture behind 'our offline metrics looked great but production degraded'.
The code that's actually running
Slide degree from 1 up to 14 one notch at a time. Before each notch: predict what happens to training error and to test error. Find the degree where test error bottoms out, that's the sweet spot. Now hit 🎲 Resample a few times at degree 13: the training dots barely move, yet the curve between them flails somewhere completely different every time. If the model can't agree with itself on new noise, what has it actually learned?
The takeaway