I was looking into LLM prompt and context efficacy and found this article. Having one language model snitching on the other might be effective?

I think that the practical uses of LLMs will be in more controlled environments and that on-device NPUs will make good use of domain-specific Small Language Models.