Can I Trust My AI?
I built a sycophancy benchmark for my local AI agent and compared it to frontier models. The biggest surprise wasn't the models — it was how much a simple behavioral layer changed honesty-under-pressure.
Mihai Chiorean
GPU systems, edge AI, and inference infrastructure — from the kernel up.
Latest posts
Everyone keeps rediscovering Claude Code sub-agents
Sub-agents aren't specialists — they're disposable, isolated scratchpads for the noisy, expensive, tool-heavy work, handing back only the part that matters and keeping the main thread clean for decisions.
The hard part of Swift on Jetson was not Swift
A Swift host process driving an NVIDIA DeepStream graph on a Jetson Orin Nano. Swift fit cleanly around DeepStream's C surface — that part was uneventful. What needed fixing was below the language: the container's CDI spec and the Swift cross-compile SDK sysroot.
Closing the Loop: Running the Whole Agent Stack Locally
A 229B model and the full agentic loop — agent, memory, orchestration, inference — running on a $4,700 mini-PC instead of a cloud API. What I'm running, how I got there, and what I would buy if I bought again today.