That's a very surprising statement. The number of advances in systems in the 21st century is.. embarrassingly low. It feels like a stale dead field. People keep pumping money into what is basically a dead body at this point.
> consider attending top conferences in your field before committing (e.g. SOSP for OS/Systems, CVPR for vision, CHI for HCI,...). I
I don't think that this is good advice. Certainly I would never suggest it to my own students. It's hard to see what they would gain.
So you go to CVPR or NeurIPS, you see hundreds of posters per session and talk after talk. Maybe you stick around for workshops to interact with people. So what? This gives you zero information about what you will be doing as a PhD student or how useful your PhD will be.
The way to figure out if you want to do a PhD is simple: do research first. Spend a summer in a research lab related to some topic you like and figure out if you like it.
> I was talking to a theoretical Physics PhD at a conference recently, and even he was quite jaded (he switched to CS post PhD).
Speaking of dead fields... I know many postdocs who come from physics into ML because many areas of theoretical physics just don't seem to be going anywhere anymore. Not a good sign!