Agreed, but better-than-middling small model local LLM inference on the M-series has largely been a happy accident that Apple have been silently capitalizing on

Agreed, but better-than-middling small model local LLM inference on the M-series has largely been a happy accident that Apple have been silently capitalizing on without bothering to correct. They were first in mobile, not the best. Generative performance and training are poor to nonexistent, Core ML is incomplete and underdeveloped, memory bandwidth was decreased on the M3 Pros, and eGPUs don't work with Apple Silicon - and what, normalizing 16GB of memory will encourage Apple to stop segmenting when they already have so little to offer? They're paygating /inference/ performance beyond their Pro chips. That's their official stance. Even Intel wasn't prepared for the Cambrian explosion of AGI, and their libraries and APIs have been taught in higher ed for the past decade. The good news is that Apple are not the best or only consumer option and will likely be out-classed by AMD and Intel in the next 18 months. The price of any new Mac for inference today would be better spent renting GPU time.