The Lab

Opening it up

The tube came first. A CRT can hold a dangerous charge long after it’s unplugged, so I tied a wire to ground and slid a long screwdriver under the anode cap. Then the cables for the floppy, the hard drive, the yoke, the sound and the display came out, gently, since the tube’s neck snaps easily.

Chasing the line

One bright line means the picture’s vertical sweep has stopped. Clean boards and reseated RAM didn’t bring it back, and the display cables tested live. On a friend’s advice I re-soldered the display connector and replaced every capacitor on the three boards, ordering each value from the schematics.

The grey screen

In 2026 it boots from a BlueSCSI until the disk driver loads, then stops at a grey screen. The BlueSCSI’s debug log showed why: the dead internal drive still answers when the driver scans the SCSI bus, and the scan never finishes. Unplugging that drive’s power should let it boot from the SD card; a new drive is the plan.

By the numbers

Machine	Macintosh SE · 1987
CPU	Motorola 68000 · 8 MHz
Paid	$50, on Craigslist
Replaced	150 capacitors
Working	January 28, 2021
Disk	System 6.0.8 · on a BlueSCSI v2

Getting there

It arrived on April 20. With the SSD from my PC inside, it booted Windows 11 and then bootlooped, so it got Linux Mint with XFCE, which has Windows 2000 vibes. Then the i3 window manager went on top, customised a lot.

How I use it

Each of the four workspaces has an icon in the bar along the bottom. It plays my Jellyfin library and reaches the NAS over SSH, and the first note I wrote on it called it a fun laptop for Obsidian writing. That’s the keyboard’s doing.

By the numbers

Machine	MacBook · white, 2010
CPU	Intel Core 2 Duo
System	Linux Mint XFCE · i3
Workspaces	terminal · internet · Obsidian · random
Arrived	April 20, 2026

Who does what

The NAS runs 42 containers: films, music, photos, books and the dashboard. The GPU server does the heavy lifting, local models in Ollama and the face and smart search for the family’s photos, which used to pin the NAS’s Pentium. If the GPU is busy, the photo library quietly falls back to the NAS’s own processor.

Sharing one GPU

Now and then the A2000 goes to a Windows virtual machine. A hook stops everything using the card before handing it over and starts it all again after. The first version restarted the AI mid-handoff, which hung the NVIDIA driver so hard that only the power button could reboot it, so now the handoff only ever stops things.

Upkeep

In August I updated 18 containers across two machines with no failures. The update tool I’d used spoke a Docker API too old for today’s Docker, so a maintained fork replaced it, and the photo library stays pinned, since its database upgrades are one-way. Every app shares one dark theme, injected by a small nginx sidecar where an app has no setting for it.

By the numbers

Containers	42 on the NAS
Models	Gemma 4 · Qwen 2.5 · Qwen 3.5
Monitoring	Prometheus · Grafana · Wazuh
UPS	about 104 minutes of runtime
Shutdown	begins at 15% battery
August	18 containers updated · no failures

Every Sunday

On Sunday night the box’s BIOS switches it on. Ten minutes later a timer on the NAS waits for it to answer, pushes the week’s changes with restic, and switches it off over SSH, about twenty minutes after it woke. If the NAS never does, the box switches itself off after 26 hours.

Why it works this way

Dell’s Auto On reads the clock with no idea of time zones, so the clock is kept on UTC and daylight saving can’t move the wake. Pushing beat pulling: the first scheduled push added 666 MiB in about seven minutes. The 2012 BIOS has no modern interface, so I read its wake schedule over SSH with libsmbios instead of a monitor and keyboard.

Keeping watch

A script on the NAS records each run for Prometheus, touching the box only while it’s awake, and a daily job ages the result, so the dashboard knows how old the last snapshot is while the box sleeps. Under 8 days is green, up to 14 is amber, older is red; I tested red by breaking it on purpose.

By the numbers

Box	Dell OptiPlex 7010 · 2012
Backed up	about 276 GB
First copy	222 GiB · checked clean
Keeps	8 weekly · 6 monthly
Awake	about 20 minutes a week
Fallback	off by itself after 26 hours

Watching

Prometheus scrapes every machine every 15 seconds: each host, each container through cAdvisor, and the GPU’s temperature, load, memory and power. Grafana draws a dashboard per machine, built as JSON and sent in through its API, plus one for the off-site backup.

Security

Wazuh watches logins, sudo and the files that matter on the NAS and the GPU server, and keeps alerts for 90 days. I wrote rules for brute-force logins, root logins and new accounts, then tested them the honest way: failing SSH on purpose, adding a test user, dropping in a cron file.

The dashboard

I rebuilt it in July with a page per machine, in the same navy and neon cyan as the rest of my apps. The map on its home page is one inline SVG of about 7 KB, drawn by a script: a 75-point outline of the US, a dot pattern instead of thousands of circles, and an animated link between home and Michigan.

By the numbers

Scrapes	every 15 seconds
Exporters	node · cAdvisor · NVIDIA GPU
Security	Wazuh · alerts kept 90 days
Map	one SVG · about 7 KB
Built	June 20 · rebuilt July 3

The line

The modem port runs through a DIN-8 to DB-9 cable and an FTDI USB adapter into the GPU server, where a serial login drops straight into tmux. ZTerm 1.0.1 plays the VT100, with its modem init string left blank, since the default AT commands confuse the login.

Making Claude fit

Claude Code draws in Unicode, so a small Python program sits in between and turns its prompt, spinner and rounded boxes into ASCII, and serial logins run with LANG=C. Wide emoji still arrive as two question marks. For the cleanest output there’s a teletype mode that prints one answer at a time.

The disk and the game

A BlueSCSI v2 on the external SCSI port, powered by the Mac itself, serves 2 GB disk images from a 32 GB SD card. The Oregon Trail came down by ZMODEM through a helper that steps tmux aside and back. ZTerm has to decode MacBinary, or the app arrives without the resource fork it needs to launch.

By the numbers

Machine	Macintosh Classic II · 1991
CPU	Motorola 68030 · 16 MHz · no FPU
Screen	512 × 342 · one bit
System	7.1 · on a BlueSCSI v2
Line	19,200 baud · 8-N-1
Terminal	ZTerm 1.0.1 · VT100

From Jellyfin to the card

MacFlim’s encoder reads each video straight from Jellyfin by its URL, so nothing is downloaded first, and dithers every frame for the Classic II’s screen. The two widest are letterboxed; the 16:9 ones are zoomed to fill it, which would crop over a third off a 2.4:1 film.

Getting them onto the Mac

Over the serial line the four would have taken about 35 hours. Instead the SD card went into the NAS, the files came across the network in 2.4 seconds, and two small tools set each file’s type and creator so the player knows them.

Why the GPU didn’t help

The slow work, Floyd–Steinberg dithering and a search for the best encoding of each frame, runs on one CPU core with no GPU path. On a 20-second clip the GPU saved 6%; running several encodes side by side is how to use the other seven cores.

By the numbers

Films	4 music videos · 18.5 minutes
On the card	about 243 MB
Rate	16–21 MB a minute
Encoding	about 12 minutes for all four
Screen	512 × 342 · one bit

What it shows

Two screens, switched with the space bar: the machines with their services and load, and a music screen with a 4 by 3 grid of album covers and the waveform. A Python program on the GPU server gathers the numbers, from service checks, Prometheus and the GPU, and sends them down the serial line as text every 30 seconds.

One bit, eight frames a second

Covers are scaled to 64 by 64 and dithered with Atkinson’s method, MacPaint’s own, after stretching their contrast. The waveform is four nested ribbons drawn from one curve, line weight standing in for colour. Interpolating once per frame instead of once per ribbon cut 384 divides to 48, and only the waveform’s strip, about 5% of the screen, is redrawn.

Receive only

The app never sends a byte: the same port carries my login to Claude, and anything it sent would be typed into that session. To keep the screen from burning in, the whole layout shifts a pixel every minute.

By the numbers

App	about 32 KB of C · Retro68
Memory	a 512 KB partition
Updates	every 30 seconds
Waveform	8 frames a second
Covers	12 · 64 × 64 · Atkinson-dithered
Tests	81 checks · fed 7 bytes at a time

How it hears you

The tweak lives inside iOS 6’s assistant daemon. It copies the microphone’s audio as Siri records, so Apple’s own end-of-speech detection still runs, and sends it to the GPU server, where faster-whisper writes it down. Apple’s connection is stubbed out, so there’s no fake Apple server and no 2012 protocol to imitate.

How it answers

Gemma 4 first picks what kind of request it is from a fixed list, then fills in the details. A handler fetches live data, and the reply goes back as Siri’s own objects, so every card is drawn by Apple’s 2012 code. Routing used to be pattern rules, until “what camera should I film with” brought up a movie called With.

Keeping it honest

An early version said it had set a timer, four times out of four, with no timers at all. Now anything that acts on a device is gated, and the car card only ever says the horn was sent, since no sensor can confirm a honk. The Mach-E and solar cards are web pages for Safari 6: no flexbox, prefixed gradients, fonts cut down to the glyphs they draw.

By the numbers

iPad	4th generation · iOS 6.1.3
Speech	faster-whisper
Model	Gemma 4 12B · on Ollama
A stock quote	about 4 seconds
A search	15–21 seconds
Public build	on my Cydia repo · without the house cards