Parameter counts have been secret since GPT-4. With revised IKP probes we measured six GPT-5.6 variants and three Claude 5 models: sol sits near 2T with fable/opus and kimi-k3; terra and luna are indistinguishable; pro adds no capacity; honest models are hardest to measure.

A node in my home cluster started crashing with no pattern to it — sometimes just a process, sometimes the whole machine. Over the next two weeks, the question I cared about wasn't how to stop the crashes, but where, exactly, it was broken.

A Chinese blog's font optimization turns into a two-week project: self-hosting Source Han Sans, frequency-bucketed slicing, a build-time Heti replacement. The real difficulty, it turns out, isn't the fonts.

LLM addiction runs deeper than flattery — it stems from frictionless reception. Drawing on Iser and Husserl, this essay argues the LLM is history's first omnipotent implied reader, and the real hook is the self that emerges when you write for a reader you believe to be total.

Faced with platform lock-in and technical churn, I designed a personal knowledge-base architecture built to survive decades. This article shares a four-layer blueprint—composability, data ownership, total decoupling—for a truly free and controllable online home.

A five-minute script to fix category typos ballooned into a 2,000-line TypeScript project with layered architecture and production-grade features. This deep dive retraces that overengineering episode and how to turn the process into knowledge assets.