I understand what numexpr does, I don't understand why I'd use it.

Polars is able to lazy evaluate query plans without any unnecessary intermediate allocations, if I want to do algebra on dataframes, I'd use polars.

The narrow usecase seems to be that you have large matrices such that memory efficiency is a concern, but not so large that they don't fit into memory at all.

My point was that this seems like a very narrow niche to me, where I'd still rather use numba or taichi purely because I don't have to evaluate raw strings and can still rely on linters.