I'm a bit tired of reading motivated blog posts like this that bring up enough related work that you know they know solutions exist to the problems they're bringing up, but for some reason felt it was not necessary to bring up those solutions. This line in particular is plain false:

> None of the languages have found a way to formally disallow paradoxes like out-of-thin-air values, but all informally disallow them.

The post neglects to mention the Promising Semantics paper of 2017 that resolves the out of thin air problem to (as far as I know) everyone's satisfaction, despite pointing out the previous work that brought up the problem. Similar things are true for ARM's memory model, etc.--this is all stuff that's been mostly resolved within the last few years as proof techniques have caught up with compilers and the hardware. Ironically, the thing that's been hardest to formalize by far in a useful way (outside of C++ consume) is--surprise, surprise--sequentially consistent fences!

It also handwaves away as some unimportant point the reason why compilers provide things like relaxed accesses--it's not just (or even primarily) about the hardware, but about enabling useful compiler optimizations. Even if all hardware switched to sequentially consistent semantics, don't expect languages that aim for top performance to abandon weak memory. And personally, I think it's ironic that at a time when even Intel is struggling to maintain coherent caches and TSO, and modern GPU APIs don't provide sequential consistency at all, people are trying to act like hardware vendors will realize "the error of their ways" and go back to seqcst.