What's happening.

Common is online. This site is the first public face of what we are building — a decentralised AI inference network that runs on the devices of its participants. We are early. We are being honest about that. The pages you are reading were built in a few weeks while we were also prototyping the actual network. We will keep adding to both.

The first real piece of the network works. We built a prototype that embeds a user query and seven domain centroids into a shared latent space using sentence transformers, then computes cosine similarity to route the query to the right specialist. It also detects Demand Vector Clouds — regions of the inference space with no good specialist yet. That detection is the foundation for the Soup of Experts architecture. The maths works. Now we need the network to run it on.

We have started talking with researchers at the Australian Institute for Machine Learning at the University of Adelaide. Early conversations around causal reasoning and world models — specifically JEPA (Joint Embedding Predictive Architecture) as an alternative to token-by-token generation for our specialist nodes. Nothing formal yet. We are listening more than talking at this stage.

We moved away from a dedicated router model. Instead, each specialist now has domain awareness built into its system prompt and decides for itself whether a query belongs to it. This removes the router as a single point of failure and cuts one full model call from the latency chain. Testing on STEM and HASS domains. Results are promising. The cold-start time on consumer hardware is still the main problem we have not solved.