Posts
Running a 744B Model on a Laptop, Then Telling the Author Not to Fix That Bug
colibri claims you can run GLM-5.2, all 744 billion parameters, on about 25GB of RAM. I actually ran it on an M1 Max. Seven seconds per token. Three counterintuitive findings came out of it, and one of them turned my upstream report from "please fix this bug" into "please don't".
The Paper Says ARM Support. I Might Be the First to Actually Compile It
Microsoft released VibeVoice-ASR-BitNet, a speech recognition model that claims real-time transcription on CPUs, "targeting both ARM and x86". I put that to the test on an M1 Max, and debugged my way from a compile error to a beautiful NaN fingerprint, finding three upstream bugs along the way. A story about how to read model cards.
I Held a Funeral for an AI. It Didn't Die.
The day before Fable 5 was set to vanish from my subscription quota, I asked it to distill its judgment into an inheritance its successor could execute. Funeral held, estate settled. Then Anthropic extended it a week.
Nobody performs for a coding agent, so I let one read six months of my logs
6 months, 7580 prompts, 'verify' typed 438 times, the same correction repeated 5 times. A friend sent me a prompt from Twitter; I ran it on my own session logs, got one line that stopped me cold, and open-sourced it as a skill the same day.
CLAUDE.md Is a Score: Running a One-Person Company as an AI Conductor
I don't have employees, and I don't pay for a pile of SaaS. I restructured my work into a system of AI agents: two Macs, one rule file that keeps growing, and four agent roles that each stay in their lane. Here's the architecture and the traps I hit.
I Gave My Blog Both an MCP Server and a CLI, Then Measured the Tokens
'MCP is dead, CLIs are cheaper' takes are everywhere. Why not both? I built both paths for this blog, mostly for fun, and measured the same task over each. The conclusion is more boring than the takes — but it has numbers.
From 3 Hours to 30 Minutes: My Automated Chinese Subtitle Pipeline
A subtitle production line running on Apple Silicon — local ASR, on-screen term extraction, LLM correction, and a fail-safe quality gate. This post covers the architecture, the term-learning loop, and three lessons that took it from "it runs" to "it's reliable."