Yeah this is such a funny thing going around. Try to run or train a small/medium sized model on a MacBook. It doesn’t go very good compared to a dedicated gpu

This is likely the right path in the future but it isn’t there yet today