> memorizing fastest algos

I don't think most LC problems require you to do that. Actually most of them I've seen only require basic concepts taught in Introduction to Algorithms like shortest path, dynamic programming, binary search, etc. I think the only reason LC problems stress people out is time limit.

I've never seen a leetcode problem that requires you to know how to hand code an ever so slightly exotic algorithm / data structure like Fibonacci heap or Strassen matrix multiplication. The benefit of these "fastest algos" is too small to be measured by LC's automatic system anyway. Has that changed?

My personal issue with LC is that it has a very narrow view of what "fast" programs look like, like most competitive programming problem sets. In real world fast programs are fast usually because we distribute the workload across machines, across GPU and CPU, have cache-friendly memory alignment or sometimes just design clever UI tricks that make slow parts less noticeable.