It's a trick. His example for cosine interpolation is misleading in the sense that the sequence of points' heights is oscillating.
The cosine interpolation simply ensures that the interpolated-function derivative is 0 at the control points. If anywhere in the sequence there was a point that is lower / higher than both its predecessors the downsides of this technique would become obvious.
