[stealthcat]

The spectrum is too wide, might need to dabble to see which works for you:

- Mainstream vision/text/speech, something Huggingface et al do. The classic advice where you learn SGD, stack net layers and put loss, train, deploy on web, clean and collect data, retrain

- Hardware accelerators, inference and training. Check out startups like tenstorrent, tinygrad, mythic. Also involve compilers, net graph data structure and good software frontends and abstraction to hardware to write models onto hardware.

- Anything else, applying ML for finance, manufacturing, oil and gas, all those boring but important stuff