“Future Mixture-of-Experts models may feature experts that vary in computational cost by factors of 100 to 1,000 and have variable-depth processing paths.”