Hacker News Reader

I'm curious what the best performance gain they think they can get by sticking to their current approach (idealistically).

They think (from what I understood) that the limit is "above pypy performance"