Every model lives on a dial of complexity, and both ends are bad.

• Underfitting. The model is too simple to capture the pattern — a straight line through a clearly curved trend. It does poorly on both training and test data. This is high bias.
• Overfitting. The model is too flexible and bends to the noise — great on training data, poor on new data. This is high variance.
• Just right. Complex enough to fit the trend, simple enough to ignore the noise. Good on both.

Too simple, too wild, just right

The dots follow a gentle curve with a little noise. Step through three models: a straight line (underfits — misses the curve), a sensible curve (fits the trend), and a wild high-degree polynomial (overfits — threads every point but lurches between them). Watch the test error tell the truth.

The dial you must tune

Model complexity is the master dial of machine learning, and the goal is always the same: turn it to the sweet spot where test error is lowest. Turn too far either way and the model fails — just for opposite reasons. The precise way to think about this balance is the bias–variance tradeoff.