Writing
Notes from the workbench.
Practical notes on AI adoption and local models, from someone who runs both in production. Grouped by series, in reading order. Posts drafted by the lab's local writer model say so in a byline. New posts land here first; there is also an RSS feed.
For teams rolling out AI
Ownership, sign-off, measurement, and dependency risk. Start here if you run a rollout.
-
The missing 90 percent of AI adoption
73% of surveyed enterprises use AI but only 10% run on it. Missing ownership, sign-off paths, and measured workflows create the gap; the models are fine
-
Routing work between local and cloud
Model rankings churn monthly; a routing policy survives. Classify LLM workloads by risk, privacy, latency, and volume, and keep the decision log.
-
The week the cloud blinked
Fable 5 vanished for 18 days under an export-control order. Frontier model access is now a dependency risk you have to engineer against
-
The open-weight frontier arrived in June
In one June window, open weights reached the coding frontier. What I'd build on, what I'd admire from a distance, and which numbers to trust
-
My timesheet drafts itself, and I still approve every line
An agent I built in monday.com drafts my workday from my calendar, meetings, and email. Nothing reaches the real time log until I approve it, and every uncertain row says why.
Judging and evals
How the lab decides what is good, what is kept, and when to stop.
-
The judge that runs on my desk
I benched local LLMs as judges, watched a bare prompt invert the winner, and swapped a cloud judge for a local one once the rubric was armed
-
I delete more local models than I keep
How a nightly coding benchmark and a strict parity rule turned a season of hyped local models into deleted weights and a stack I trust
-
Self-refinement that knows when to stop
Nous Research's Autoreason finds naive self-refinement makes output worse, then fixes it with a blind tournament where doing nothing can win
Training a local writer
Three parts, in order: distilling a style, a loop that learns from my edits, and a two-retrain night.
-
I trained a local model to write like Claude
I trained a 9B to write like Claude, benched it blind, and the ending flipped: the bigger model with the distilled skill took the seat
-
I built a loop that learns from my editor
Part 2 of the local writer story: a self-improvement loop whose first act was rejecting its own idea, and why that rejection is the feature
-
The night my writer retrained twice
Two retrains in one evening. The second, trained against my style gate's measured failures, moved the 12-brief bench from 88.8 to 97.0.
Memory on a 64GB Mac
The capacity math, the scheduling rules, and the loader that fit a 68 GB model in 42.
-
Local LLMs on a 64GB Mac Studio
The awkward middle tier of local AI in mid-2026, the capacity and bandwidth math that governs it, and a 75 percent verdict on what a 64GB Mac Studio actually delivers
-
Scheduling one big local model on a 64GB Mac Studio
How I schedule large local models on a 64GB Mac Studio with a router and a swap tool instead of juggling memory by hand every day
-
The Mac mini that never swaps
Why the unattended jobs in my local-AI setup run on a dedicated always-on Mac mini instead of my 64GB Mac Studio, serving one small model that never gets evicted.
-
Fitting Qwen3.8 Flash-Next on a 64 GB Mac Studio
How a 68 GB-resident model runs in 42 GB at better-than-reference speed with a row-granular NVMe loader for its 51B n-gram table
The knowledge base and the night shift
What the lab does unattended, what broke, and how the 24/7 operation came together.
-
A wiki my models maintain
My local LLMs maintain a knowledge base: they capture sources, summarize them, and commit the output daily. The RAG engine I built on top was the wrong idea
-
The 4:30am shift
A stack of cron jobs on local LLMs runs my knowledge base overnight. Two failures I watched in the logs changed how I build every local pipeline.
-
I Built a 24/7 Local AI Lab, Then Cut Its Biggest Model by 62%
I paid frontier models to build a 24/7 local AI factory, then cut its active large-model package by 62 percent and moved the recurring work onto hardware I already own.