upvote
Surely they are testing their optimizations against common benchmarks internally? I bet the "real world task" degradation is larger by some multiple than it appears when measured through a benchmark that is part of the target.
reply