An input analyzer that finds out what kinds of tokens the query contains. A bunch of specialized models which handle each type well: image analysis, OCR, math and formal logic, data lookup,sentiment analysis, etc. Then some synthesis steps that produce a coherent answer in the right format.
Eventually there is going to be a model catalog that describes model inputs/outputs in a machine parseable format, all models will use a unified interface (embedding in -> embedding out, with adapters for different latent spaces), and we will have "agent" models designed to be rapidly fine tuned in an online manner that act as glue between all these different models.
This strikes me as the most compute efficient approach.
The model learns how to split them on its own, and usually splits based not on topic or domain, but on grammatical function or category of symbol (e.g., punctuation, counting words, conjunctions, proper nouns, etc.).
I thought half the point of MoE was to make the training tractable by allowing the different experts to be trained independently?