I built a 24/7 AI lab that works while I sleep, then cut its biggest model footprint by 62%.
It researches, tests, writes, and watches my stack overnight. By morning, it brings me decisions instead of chores.
How Codex orchestration made it autonomous:
https://t.co/7b4iFRiyBo
I retrained the model that drafted this post twice in one night on an 18GB MacBook. Its score on my 12-brief bench went from 88.8 to 97.0 with all briefs passing.
https://t.co/6Xbz8GKIkq
A 64 GB Mac Studio holds one big model at a time and evicts things under memory pressure. Fine when I'm at the desk, wrong for a 5:15 AM job. So the always-on work runs on its own Mac mini: one 9B it never swaps.
The mini serves that one model and never swaps it. It runs the unattended jobs: a fallback summarizer for my wiki, a nightly lint pass at 5:15 AM, and the narration for my morning brief. A weaker model, but always up.
Part 2 of the local writer story. Shout out to @cjzafir: his skill-as-teacher idea sparked this build. The writer now improves itself: it mines my editor's corrections and adopts only rules that win a blind bench. Cycle 1: it rejected its own idea, 6-0.
https://t.co/fCfBLDh9bB
Shout out to @cjzafir: his skill-as-teacher post sparked this build. I distilled my voice into a skill, LoRA-trained a 9B on 267 examples of my writing, and benched it all blind. The twist: the skill itself turned out to be the real asset. Full run:
https://t.co/zP0HhrehcI
I just created my first website within 30 minutes from Ideation to Creation. Without any code. at all. I don't know how to code. This was my gateway into "vibe coding". Pretty addictive! https://t.co/4YpTWaws8P #claudecode
@simontran@greatmattsby I’ve been following @greatmattsby for awhile and knew this would hold. Purchased more at this symmetrical support and has paid off well