https://commonai.com.au/findings/ Working notes from Common AI, a team building decentralised AI in Adelaide, South Australia. Written as we go, including the parts that failed. en-AU Sat, 10 Oct 2026 00:00:00 GMT https://commonai.com.au/findings/2026-10-10-specialists-blended-on-demand/ https://commonai.com.au/findings/2026-10-10-specialists-blended-on-demand/ Sat, 10 Oct 2026 00:00:00 GMT Research Can a specialist be built without training one? A blend of three public LoRA adapters beat the base model by 9 points on 200 questions none of them had seen. The first run of the same experiment stopped. https://commonai.com.au/findings/2026-09-06-first-version-of-our-network/ https://commonai.com.au/findings/2026-09-06-first-version-of-our-network/ Sun, 06 Sep 2026 00:00:00 GMT Milestone An early Alpha version of our network has been released. https://commonai.com.au/findings/2026-08-15-the-graph-is-the-reasoning-layer/ https://commonai.com.au/findings/2026-08-15-the-graph-is-the-reasoning-layer/ Sat, 15 Aug 2026 00:00:00 GMT Research We said CGLA was zero-hallucination and that more was coming. Here are the numbers from two working systems, and what that claim looks like once it is measured. https://commonai.com.au/findings/2026-08-15-what-breaks-at-the-seam/ https://commonai.com.au/findings/2026-08-15-what-breaks-at-the-seam/ Sat, 15 Aug 2026 00:00:00 GMT Research The full results from three experiments on handoffs between small specialist models: what we measured, what the noise floor said, why one fix worked, why the actual MoE bet failed, and why we cannot yet test whether composition works. https://commonai.com.au/findings/2026-07-14-prototype-decentralized-network/ https://commonai.com.au/findings/2026-07-14-prototype-decentralized-network/ Tue, 14 Jul 2026 00:00:00 GMT Milestone We have built the first internal prototype for the decentralized network and it is currently undergoing testing! https://commonai.com.au/findings/2026-06-19-cgla-system-testing/ https://commonai.com.au/findings/2026-06-19-cgla-system-testing/ Fri, 19 Jun 2026 00:00:00 GMT Technical We've been testing a new type of graph based SLM (Small Language Model) making it zero-hallucination, and allowing inference and training on consumer hardware. This architecture is called CGLA (Causal Graph Learning Algorithm), It uses a causal graph to learn relations between cause and effect. We… https://commonai.com.au/findings/2026-06-13-website-live/ https://commonai.com.au/findings/2026-06-13-website-live/ Sat, 13 Jun 2026 00:00:00 GMT Milestone 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. The pages you are reading were built in a few weeks while we were also prototyping the actual network. We will keep adding… https://commonai.com.au/findings/2026-05-20-vector-routing-prototype-complete/ https://commonai.com.au/findings/2026-05-20-vector-routing-prototype-complete/ Wed, 20 May 2026 00:00:00 GMT Technical 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… https://commonai.com.au/findings/2026-04-10-aiml-conversation/ https://commonai.com.au/findings/2026-04-10-aiml-conversation/ Fri, 10 Apr 2026 00:00:00 GMT Research 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… https://commonai.com.au/findings/2026-03-01-self-routing-specialists/ https://commonai.com.au/findings/2026-03-01-self-routing-specialists/ Sun, 01 Mar 2026 00:00:00 GMT Technical We moved away from a dedicated router model. Each specialist now has domain awareness built into its system prompt and decides for itself whether a query belongs to it. That removes the router as a single point of failure and cuts one full model call from the latency chain. We are testing it on…