I recently attended a Gartner webinar where they were discussing offloading activities from LLMs. Given the lower energy requirements and the one-shot learning capabilities of functor models I came up with the Global Network Model (GNM). This is preliminary work but it suggests an alternative to scaling for the LLM. Here agents work with functor micro models to discover new data. That data goes to the data processor for cleansing, dedups, canonicalization, learning preparation, etc. Admittedly the data processor and the data science involved are complex and the icon in the diagram below represents one or more large diagrams. Then the learning router submits the learning unit to the appropriate SLM. The LLM can invoke inference on any SLM to gain knowledge, generate code, etc. The Graph Neural Networks are contained in a new type of ensemble that allows messaging between nodes in different models using an orchestrator. 

This can be an alternative to adding transformers and using GPUs in certain cases. Lower energy consuming functor models can support a reduction in overall LLM energy usage and cost. For example AssuredCode from Autonomic AI is a fully deterministic software code generator that has very low latency and low energy consumption. It also uses no external code so faces no legal issues of the types LLMs face. It is also designed as a module for an LLM. This is an excellent candidate for offloading. See this document for further GNM details.

John Harby
CEO
Autonomic AI, LLC
Temecula CA
https://www.autonomicaillc.com
------------------------------