High-dimensional optimization gets stuck on saddle points, not hills

roon says that in high-dimensional space, hills are rare, so optimization usually ends at saddle points rather than local maxima. The post gives no further detail on the claim or its context.