Read the post: how to train a mixture-of-experts model, knob by knob →