wtf is distributed-cross-moe-interop?

This project explores a way for different AI models to share specialized internal components with each other instead of staying completely separate. Many modern AI models are built as what is called a Mixture of Experts, meaning the model is actually made up of many smaller specialized sub-models, called experts, and only a few of them are used for any given task. Normally each model's experts stay locked inside that one model. This project proposes a system called Cross-Expert Interoperation, where several different AI models, for example one trained on code and another trained on math, can keep their own experts privately but also publish some of them to a shared marketplace so other models can borrow them when useful, while a learning system figures out over time which combinations of borrowed experts actually help and are worth the extra delay. The repository is both a written specification describing this idea formally and a working Python reference implementation. Importantly, the author is upfront that this is mainly a conceptual and research project: the AI models involved are not real trained models, but a simplified toy setup used to test whether the coordination system itself works, including security features like verified requests and protection against replayed messages. You can try it two ways. The simpler option runs everything inside a single Python process as a simulation, comparing how well a marketplace approach performs against just using local experts only. The more involved option spins up several separate networked services using Docker, representing a registry that catalogs available experts, a router that proposes which experts to combine, a learner that tracks what worked, and multiple AI model nodes that request and lend out experts to each other. A typical request flows through checking local options first, asking the router for suggestions, borrowing a remote expert if it looks worthwhile, and reporting back whether it actually helped, so the system learns and improves. The project requires Python 3.10 or newer and includes documentation covering both the underlying idea and how to run it.