I don't know a lot about kernel scheduling, memory allocation, I/O scheduling, etc, but from what I understand there are OSes with much cleverer implementations

I don't know a lot about kernel scheduling, memory allocation, I/O scheduling, etc, but from what I understand there are OSes with much cleverer implementations of these things, and OSes which are pretty dumb. I believe Linux is somewhere in the middle. Surely there is lots more scope to optimise these schedulers with regular rule-based approaches, rather than bringing in DeepMind magic? It's always difficult to understand the reasons why machine learning algorithms choose the things they choose, and they often have unintended consequences or awful edge cases. This just seems like it was done for marketing reasons more than technical reasons?