Would this be a reasonable explanation?

> MLPs are universal function approximators, but these models are big enough that it is better to train many small functions rather than a single unified function. MoE is a mechanism to force different parts of the model to learn distinct functions.