Michael Hannecke·Oct 29My Local AI Agent Approved Its Own Report. Here Is the FixWhat a local 30B model taught me about building agents you can trust: four architecture rules, each one learned from a real failure.A response icon1A response icon1
Michael Hannecke·Oct 7Dynamic Model Routing on Apple SiliconPart 3 of Building a sovereign AI stack on Apple Silicon. What a dynamic profile hands to the next backend, and who gets to flip the…A response icon1A response icon1
Michael Hannecke·Oct 1Apple’s model at the command line: what fm serve leaves to youSeven commands, one model, and no authentication option at all. The license notice asks you not to do what the binary makes easy.A response icon1A response icon1
Michael Hannecke·Sep 30Apple’s AI model is already on your MacBut the default is not on-device for every use case
Michael Hannecke·Sep 22Attestation or Authority: Apple Foundation Models vs. Self-Hosted MLXSame task, same Mac Studio, one argument apart. Apple’s on-device model never called the tool. A self-hosted model always did. The…
Michael Hannecke·Sep 1Apple Container Moved the Boundary. It Didn’t Move the Risk.Apple container, Docker Sandboxes and Docker Desktop, compared against what actually escapes.
Michael Hannecke·Aug 29DGX Spark vLLM: sparkrun vs spark-vllm-docker ComparedTwo projects, one box, and the comparison that fell apart twenty minutes in
Michael Hannecke·Aug 27I Wanted to Build a Mac mini AI Cluster. Then I Read Apple’s Price List.Four M5 Pro minis against one M5 Ultra Mac Studio for agentic workloads: the math, the physics, and a €1,820 surpriseA response icon6A response icon6
Michael Hannecke·Aug 4The Notebook That Holds No ModelJupyterLab in an Apple container machine, MLX on the host, and why the constraint is an upgradeA response icon1A response icon1
Michael Hannecke·Aug 3Maximum Isolation: A Disposable JupyterLab for the DGX SparkA venv is not isolation on a machine whose entire value is its driver stack
