The tuning parameter ($\lambda$) controls how rough a smoothing spline can be, which also means that it controls its effective degrees of freedom. When $\lambda = 0$, the penalty term has no effect, resulting in a rough and jumpy curve that interpolates the training data. As lambda rightarrow infty, the penalty term dominates, forcing the spline to be completely smooth, which results in a straight line (a linear least squares fit). When $\lambda$ is an intermediate value, the spline balances smoothness and data approximation.