"Why advertise some super AI chip if it doesn’t do AI stuff? Who knows." => Neural engine is fp16/int8 limited so it wouldnt be good enough for training anyway but its pretty good for inference with CoreML in a lot of cases. Also training anything outside Nvidia ecosystem is very bad cause software was written with CUDA first and everything else second (AMD for example have similar story to Metal backend).
As someone who works on Deep learning for past 6 years I can easily say CUDA wont be replaced anytime soon at least for training, almost every software out there was written for either CUDA or CPU.
And for local model development you don't need your rig to be powerful as training wont be done on your local machine anyway and MacBooks are good enough for that (a lot better than windows for sure).