Comment by sigbottle
There are tons of ML compilers right now, FlashAttention brought back the cache-aware model to parallel programming, Moore's law hit is limit and heterogeneous hardware is taking taking off.
Just some fundamentals I can think of off the top of my head. I'm surprised people saying that the lower level systems/hardware stuff are untransferable. These things are used everywhere. If anything, it's the AI itself that's potentially a bubble, but the fundamental need for understanding performance of systems & design is always there.
Im actually doing a ton of research in the area myself, the caution was against becoming an Nvidia expert narrowly rather than a general low level programmer with Nvidia skills included.