The tweet author in the linked thread later writes:

"Weaker memory model makes acquire-release atomics possible to implement much more efficiently, in exchange for not hiding some classes of multithreading bugs"