PinnedCedric Nanni·Oct 30A Private AI Agent Should Be Able to Call a Smarter FriendHow Pi-Run lets an internal model consult stronger intelligence without automatically sharing its workspace, transcript, or organizational…
PinnedCedric Nanni·Oct 18AI Agents Need Colleagues, Not Just SandboxesWhy I built Guardian as a peer programmer for Pi — and why the human still decides
PinnedCedric Nanni·Oct 12Pi-Run: Beyond Permission Prompts for AI AgentsHow short prediction horizons, explicit authority, portable knowledge, and recovery became one working environment for Pi
PinnedCedric Nanni·Sep 17Sovereign AI: The Model Can Change. Your Knowledge Shouldn’t Have To.Artificial intelligence becomes really interesting for a company when it stops being generic.
PinnedCedric Nanni·Jul 21AI coding in YOLO modeLet the agent execute freely — but never let the development process run without controlA response icon1A response icon1
Cedric Nanni·Sep 2Beyond YOLO Mode: Building Boundaries for Enterprise AI AgentsAI agents need freedom to be useful. They also need boundaries.A response icon1A response icon1
Cedric Nanni·Aug 27What Does a Million Tokens Cost on My €22K AI Server?Using real inference benchmarks to estimate the economics of self-hosting versus the Mistral AI API.A response icon1A response icon1
Cedric Nanni·Aug 27How Close Is My AI Server to a DGX Station? At €22K, It Made Sense. At €50K, Does It Still?After validating the hardware underneath my dual-RTX PRO 6000 server, it was finally time to stop measuring PCIe links and start measuring…
Cedric Nanni·Aug 21Two RTX PRO 6000 Blackwells, One VM: NUMA, PCIe P2P and the Road to Tensor ParallelismIn the first article, From One RTX PRO 6000 to Two: How Close Is My AI Server to an NVIDIA DGX Station?, I started with a simple idea: take…
Cedric Nanni·Aug 21From One RTX PRO 6000 to Two: How Close Is My AI Server to an NVIDIA DGX Station?Last year I built a single-GPU Blackwell server to run large language models on Kubernetes. Since then, I added a second 96 GB GPU. With…