Wonder if you could glue together deep networks such that there's a big reasoning LLM model, and high-speed small network of a different type, and they share a subset of their activations. High bandwidth vector communication. You train them together.

edit to add: Seems this is a well-known idea in ML / robotics already.