And a group has published an independent working implementation today, nice to see:

https://github.com/tonbistudio/turboquant-pytorch

It has a lot clearer explanation of the method than Google's own post.

Well, yeah. Claude simplified it. That doesn't mean it's a better explanation.

Did it lose important detail?