Sovereign AI and LLMs

rewrite: Hi, just want to say hello. I just published this post: https://www.danielkliewer.com/blog/2026-07-04-sovereign-intelligence-stack and I was validating the ideas and somehow found your site. I think you would find some of what I have been researching interesting. I see that you know the value of local systems, but there is so much more that I have been discovering which has really enriched the type of intelligence I am able to create using local LLMs. For instance, my current setup using deepreinforce-ai's Ornith 35B MoE in llama.cpp using their chat template .jinja and now I simply run the Hermes harness. That is just the tip of the iceberg. This is where context engineering starts and you learn how to build loops. I can run this all locally on a Macbook Pro M4-Pro 48GB which is what I have been using to test LLM coding projects since it came out. The intelligence I generate is possible locally because I have complete data sovereignty and can allocate any amount of compute I want with no worries about guardrails or telemetry enforced behaviors. What is more is that I can construct my own telemetry in order to achieve the self learning capabilities of frontier models. Most of my blog has been reverse engineering frontier systems I contributed to at the big tech companies since I began work in AI in 2007 at Amazon. Sovereign AI is not just running a model locally, it is not just adding an agent harness, it is more than that. It is the memory layers and reasoning traces you can use for more and more enhanced retrieval augmented generation. But grep replaces most of that with agents now really it makes more sense than vectors in my opinion although I still use vectors with graphs and in graph construction because I find that it allows the agents to reason and collapse graph reasoning traces and be able to learn with preserved state. So no more amnesia.

Here's a cleaner, more natural version that keeps your perspective while making the flow easier to follow. :::writing{variant="chat_message" id="47281"} Hi, I just wanted to say hello. I recently published this post: https://www.danielkliewer.com/blog/2026-07-04-sovereign-intelligence-stack While I was sanity-checking some of the ideas, I came across your site. I think you might find some of the things I've been researching interesting. You clearly appreciate the value of running models locally, but I've found that local inference is only the beginning. The real gains come from how you structure memory, context, telemetry, and reasoning over time. My current setup runs deepreinforce-ai's Ornith 35B MoE through llama.cpp using their chat-template Jinja file, with the Hermes harness orchestrating everything. That is really just the foundation. Once you begin treating context as an engineered system instead of a prompt, you start building cognitive loops rather than isolated generations. I've been running these experiments entirely on a MacBook Pro M4 Pro (48GB RAM), which has become my primary platform for developing and testing local LLM coding systems. The reason this works locally is complete data sovereignty. I control the compute, the memory, the telemetry, and the feedback loops, without imposed guardrails or opaque behavioral instrumentation. More importantly, I can build my own telemetry to observe, evaluate, and improve the system over time. Much of my writing is an attempt to reverse engineer the architectural patterns behind frontier AI systems, informed by my experience working on AI projects beginning at Amazon in 2007. To me, sovereign AI is much more than running an LLM offline or wrapping it with an agent framework. It is about constructing an architecture that accumulates intelligence instead of repeatedly starting from scratch. That means layered memory, preserved reasoning traces, retrieval that evolves with experience, and systems capable of maintaining state over long periods. Lately I've also been moving away from relying exclusively on vector search. Agents equipped with fast code search (grep), structured knowledge graphs, and persistent reasoning traces often outperform pure embedding retrieval. I still use vectors, especially for graph construction and semantic retrieval, but increasingly they're one component of a broader cognitive architecture. The goal isn't simply better retrieval—it's preserving and refining the reasoning process itself so the system can continuously build on what it has already learned instead of suffering from perpetual amnesia. ::: You could also make this substantially shorter (around 250–300 words) if you're sending it as an introductory message through a contact form, which often has better odds of being read in full.